“Initially when people started working on models for language, the dominant design was LSTMs. Some people remember what those are. They were kind of okay; they were a lot better than what came before that, but they didn’t really scale as well. And then the big thing with transformers was not that transformers were somehow particularly mathematically elegant or anything like that. It’s just that they scaled better. So they were easier to train on very large amounts of data with lots of parameters. The technology proceeds in phases. First you figure out what you can scale, and then you pour on a lot industrial-scale effort and adding lots of data, adding to model size. And that’s when the magic happens.

What I would say about robotics is that it’s not in the GPT-4 to GPT-5 stage, where it’s an industrial-scale effort to make the model bigger and get more capability out of it. It’s in that stage where we’re establishing the fundamental technologies. And because of that, what one should expect to see right now is not necessarily that each month the model gets bigger and more powerful by some predictable kind of scaling curve. It’s that the scaling properties themselves are evolving as we develop the right technologies.

So, where we’re at now is that we’re actually getting all those puzzle pieces in place. And I think it’s actually very close. I think a lot of the puzzle pieces are falling in place, but it’s not yet at that predictable scaling stage. Certainly we’ve seen a lot more happening as we progress, but they were kind of magical because in robotics, basically prior to 2024, the stuff never happened. So the little things that happen, and this was maybe at this point about two years back, we would see things like, okay, we train our policy for folding laundry and it takes individual shirts out of the hamper and tries to fold them. And then one very vivid memory I have in late 2024 is we were watching one of the evals, and it takes out two shirts at the same time. And I’m watching this, I’m like, okay, it’s done for. There’s no way you can possibly do this. And then it puts the two shirts on the table, disentangles them, puts one of them back, and then starts folding the other one. It’s like, wow, okay. In retrospect, you can do some detective work and figure out where it got that from, some piece of training data. It’s exhibiting the common sense you expect people to have.

What would have the biggest impact on the advancement of robotics. Certainly, availability of reliable, low-cost hardware is a big deal. Right now, hardware used for robotics research, a lot of that does come from China. And it’s good, it’s relatively inexpensive, it’s of high quality, and meets the standards that people generally need. But it would be awfully nice to be able to source all that domestically as well. Having an effective positive feedback loop where more deployed robots translates to more model capability, that’s kind of the key. The hard thing in robotics was always generalization. What is very easy to miss in the hype cycle and the demos that people put out, is that when somebody shows a demonstration of their system, the demonstration alone usually doesn’t make it clear what level of generalization is being shown.

The highly acrobatic robot demos, for example, are really exciting to look at. But typically, it’s a show that is obviously rehearsed. It’s not the same as doing a task every time reliably in any home. So you might see the robot doing something fairly mundane and unimpressive, but what’s exciting about it is that it’s doing it with an object that it’s never seen before in an environment that has never been tested before. And that is actually harder than doing an acrobatic backflip that it’s practiced millions of times. One of the steps on the roadmap is to have a robotic system that gets better with autonomous experience in a setting that it wasn’t originally trained for. I get my model, I get my adaptation algorithm, and I put it in a new setting. Maybe it’s a home, maybe it’s a factory, something where it’s doing something real, and it does okay. But then over time, it gets better and better. That would be a major milestone.

Live demos with really long time lapses, I think it’s a great idea. We had our robot assembling boxes at Dandelion Chocolate Factory. So we ran it for several days. But we had this coffee task, which was the robot using an espresso machine to make espresso. We ran it for 13 hours making espresso drinks. It screwed up a few times. It’ll spill all the coffee grounds, and then it needs to go and get a cloth and wipe it down. The way that experiment was done is that there is a high-level prompting. So the prompt is updated roughly every five minutes or so in between semantically coherent tasks. So you tell it: make espresso, clean up the machine, etc. So those steps, the actual clean up the machine, was commanded by a person. In principle, we could automate that. In fact, one of the things we’re spending a lot of effort now on is improving our high-level policy that does those commands. But for that experiment, every five minutes, somebody basically updates what it’s being asked to do, the way we intended. It was like the commands would be like, if you go to an actual coffee shop, you say, “Oh, I want a latte, I want an espresso.” That was supposed to be the prompt. Except then you also have to tell it, “I want you to clean it up before you do the next one.”

One thing that I’ve learned over the last few years is that addressing robotics effectively requires using very broad prior knowledge. For example, I was very skeptical initially about the utility of language. And I think scientifically this is defensible, which is that animals can do some pretty impressive things, but they can’t speak, at least not very eloquently. So maybe our robots should also be able to do stuff, and they don’t necessarily need to understand language. But what’s important is prior knowledge that you can put it in as a scaffold on your learning process. And you can pull in that knowledge in all sorts of ways.”

If we look at the progress in AI—or machine learning, as it used to be called a decade ago—we’re probably in a similar phase now in robotics as we were in digital neural nets one decade ago, around 2016 or so. This was a time when AI had a good understanding of what is in an image, what’s the meaning of a text, and reinforcement learning had just been introduced to give AI superhuman performance. Currently, while robots start being able to do basic tasks, we’re not in a phase yet where they have as good an understanding of the surrounding environment as humans do.

A few months ago, on the Invest with the Best podcast, Levine gave more background on his company and the robotics field:

“Part of the thesis of this company is that we believe doing it at the full level of generality might actually in the long run be easier than trying to special-case very specific narrow application domains. Again, in much the same way that for language models, it turned out to be easier in some ways to solve natural language tasks in their full generality than to narrowly target machine translation or sentiment analysis. In the world of natural language, we saw that there were a lot of efforts to develop domain-specific solutions that tackled specific problems.

For example, somebody would spend a lot of time thinking about how English differs from French and then build a machine translation system. The reason that language models actually took over for all of those different application domains is because they can leverage much broader sources of data. When you can leverage weakly labeled data, like data that you mine from the web in the case of language models, you actually learn more about the world. You establish a foundation of world understanding, and on top of that foundation, it turns out to be much more effective to build out different applications.

To bring this into robotics, the calculus does not look quite the same because in robotics we do not have an internet-sized dataset that we can just draw on. But this notion of understanding the world is actually more important in robotics. If you have many different tasks and many different physical systems, then you can go from training individual dishwashing specialists or laundry-folding specialists to instead training a model that actually understands physical interaction. People can master new skills very rapidly because we understand physical interaction. We can intuitively grasp what is going to happen in a new unfamiliar situation, which lets us bootstrap things really quickly. If we can draw on data from many sources, many applications, and many robots, then we can have a model that has physical understanding, making it much easier to put new applications on top of that platform.

The point of generalization is that it does something relatively mundane that any human could do, but it does it in any situation. We released some demos where we showed our robot cleaning kitchens. If you watch an individual video out of context, it just looks like it is picking up plates, which anybody can do. But the point was that we put the robot into that home just for that demo, and it had never seen training data from that setting. You have to understand what is going on behind the scenes to appreciate why that is pushing the frontier.

In the future, we will have a robotic foundation model that can be adapted to all sorts of applications, running the gamut from bulldozers to humanoids to robotic arms. The fundamentals of how you interact with objects, how things move in the world, and how causality works are conserved across all of these different systems. Machine learning works best when there is a lot of data. If you naively approach a problem like washing dishes by collecting an enormous amount of dishwashing data, it is not cost-effective because you have to repeat that entire process for the next task. Training general-purpose models reduces the data needed per task.

What has changed most recently is the ability to handle unusual scenarios. In edge cases where you lack prior experience, you must rely on knowledge acquired from other sources and ground it in the new situation. Multimodal language models are adept at pulling in broad web knowledge. While they are not inherently grounded in physical situations, they provide a path to import common sense into robotic control. The first deep reinforcement learning systems in the early 2010s were a major milestone because deep RL gives us a way to surpass human-level performance. More recently, the advent of multimodal LLMs adapted for robotic control to supply common sense is a critical advance, and we will likely see several more major breakthroughs over the next few years.

Over the past few years, we started by developing Vision-Language-Action (VLA) models. A VLA model can be understood as an LLM adapted for robotic control. It is pre-trained on text, adapted with extensive web image data to understand visual scenes, and then fine-tuned on diverse robotic data. That serves as the foundation for bringing web-scale knowledge into robot control. On top of that, we explore two primary capabilities: handling unusual situations with common sense, and improving through reinforcement learning. To achieve common sense, we utilize chain-of-thought reasoning. When the robot encounters a scene, instead of immediately executing motor commands, it reasons through the task. If asked to clean a kitchen, it analyzes the visual input, generates an intermediate semantic thought like picking up a specific plate, and then acts. These intermediate inferences benefit from web-scale pre-training.

Reinforcement learning is applied when practicing tasks repeatedly to improve robustness, speed, and precision. In our espresso-making demonstration, the system practiced repeatedly to optimize throughput and reliability. We are continuing to build upon this dual foundation. Nobody knows exactly how much robot data is needed to achieve universal generalization, but we may not need to know in advance. The objective is to deploy systems capable of performing diverse tasks and establishing a data flywheel.

Humans intuitively assume that tasks difficult for us, like advanced calculus, are hard for machines, while tasks easy for us, like grasping a cup, should be simple. In reality, human brains have evolved massive specialized machinery for physical interaction and visual perception. Machine learning shifts this dynamic. Where data collection is straightforward, tasks become tractable even if they are physically intricate. While learning is now widely accepted, there is still debate over end-to-end learning versus modular systems incorporating hand-coded physics. The Bitter Lesson posits that leveraging general methods with data and compute outperforms hand-engineered domain knowledge over time. While incorporating explicit physical models offers short-term structure, end-to-end learning provides the scalability and self-improvement necessary for true open-world generality.

Former Everyday Robots researcher Benji Holson proposed a practical Robot Olympics consisting of everyday manipulation tasks that humans find trivial but robots struggle with—such as opening doors, scrubbing greasy pans, or using a bag to pick up waste. We used this list to test our general onboarding pipeline. Without developing specialized algorithms for individual tasks, our foundation model successfully solved nearly all of them. The only exceptions were turning a dress shirt inside out, which was physically constrained by gripper size, and peeling an orange purely with fingers, which required a small tool due to grip strength limits. This demonstrated the power of a general model to onboard diverse tasks rapidly. Tasks expected to be among the last to be automated by robots involve direct, nuanced physical interaction with humans such as changing a child’s diaper, eldercare, or helping an injured person out of bed. These tasks require precise force modulation, high safety margins, and deep social-physical common sense.”

While it’s an interesting field to follow, it will probably be another 5-plus years before robots really start to scale massively. So, investing in this space is akin to investing in AI in 2015-2016 or so, with the big J-curve in terms of demand only coming 7 years later. We think this is a good framework of how to think about the space as it likely will still take a while before the really big returns come for the winners. Next, we will review a number of the best positioned names in humanoid robotics and their supply chains, and we’ll also overview our semis outlook.

Read more

“We hosted our 3rd annual Silicon Valley AI Field Trip on 8/18-8/19, featuring AI companies, VCs and researchers from Stanford and UC Berkeley/UCSF. Model capabilities are continuing to improve, agents are progressing from assistance to workflow execution and monetization is expanding beyond seats toward consumption, transactions and outcomes. The strongest businesses combine proprietary data, domain expertise, customer context, workflow ownership, distribution, verification and permission to act. Agentic adoption should advance fastest where outputs are verifiable and responsibility for errors is clear. Open models should gain token volume across routine tasks, and frontier models should retain value where reliability justifies higher costs, supporting continued token and compute growth. AI should also expand cybersecurity demand, and physical AI will require distinct architectures and commercialization strategies. For Business & Information Services, AI should increase the value of differentiated information assets and help platforms capture more customer spending through embedded workflows. For Software, select vendors are well positioned to help customers navigate an evolving intelligence curve that segments workflows into frontier vs. non-frontier tokens and enriches models with business-specific context.”

We have a similar view and see multiple winners in this space—frontier models for the most advanced tasks, and open models for routine tasks. Companies with proprietary data will have a strong advantage—the better the quality of your data is, the better the AI’s output will be. This is relevant in both training and inference. We also continue to think that there will be obvious winners in the software space from AI, and that cybersecurity companies and data management systems such as Snowflake and MongoDB are particularly well-positioned.

CLSA tracks a wider database of AI startups and shows that the number of startups hitting key ARR milestones continues to grow rapidly:

We suspect that we’re still only at the very early start of the S-curve when it comes to AI demand. Evercore’s recent channel checks found that:

“Hyperscalers turning away customers because they don’t have enough GPUs.”

And thus AI capex continues to move higher—Nvidia sees 70% revenue growth in calendar 2027 (fiscal 2028) while the semi supply chain continues to be constrained. Nvidia’s CFO estimates that real customer demand is growing 100% year over year:

“We expect to grow revenue by approximately 70% in fiscal 2028. This is a supply-constrained outlook. With cloud industry backlog now greater than $2 trillion, CapEx by the top 5 hyperscalers is expected to reach nearly $800 billion in 2026 and $1.3 trillion in 2027. Incredibly, we are seeing demand acceleration even at our scale. Customers forecasts point to our growth, doubling next year. However, as I mentioned earlier, we expect to grow approximately 70% as we are supply constrained.”

While overall AI capex growth rates are now slowing down—36% expected for ‘27 compared to ‘26 and ‘25 which both grew around 90%—Nvidia’s comments suggest that growth rates for next year would have been almost 1.5x higher if it had not been for constraints in the supply chain.

Over the past two years, we’ve seen Nvidia’s top line growth rates decelerating while AI capex growth rates kept accelerating (chart below). However, in 2027, we see again a reversal—Nvidia will now be accelerating its growth rate again while we see a strong deceleration in AI capex growth:

One possible reason is that AI capex estimates for next year are simply too low. Another reason is that Nvidia is seeing tremendous neocloud demand from around the world—which the data above from Evercore likely isn’t properly capturing. This is Nvidia’s CFO again:

“Hyperscalers will remain a major growth driver, but non-hyperscaler growth, our AICE segment—spanning sovereign, regional NeoClouds, enterprise edge and air-gap data centers—will represent roughly half of our data center business. We don’t own a cloud ourselves, we are a neutral partner to every sovereign and NeoCloud. NeoClouds are emerging everywhere—Firebird in Armenia, Cassava Technologies across Africa, GMI Cloud in Taiwan, Yotta and Neysa in India, Firmus in Australia, YTL AI Cloud in Malaysia—pairing local land, power and operating expertise with our platform.

NeoClouds are seeing strong demand pipelines for many diverse offtakers. Rather than allocating their entire capacity to a single long-term offtake guarantee that lenders typically require to finance a data center independently, we have introduced a revenue-sharing structure. NVIDIA provides a take-or-pay commitment on a portion of the facility’s capacity, a minimum revenue guarantee that gives lenders the confidence to underwrite the project, and in exchange, we share in a portion of the NeoCloud’s revenue earned above that floor.”

We fully understand why Nvidia is doing this. The problem for Nvidia is that they don’t want a too concentrated customer base. If they would be purely selling silicon to the top four hyperscalers and the top two frontier labs, it’s likely that these big 6 will increasingly shift share to their custom ASICs over time. All are already doing this. This would make Nvidia a melting ice cube at some stage.

So, Nvidia’s answer is to set up a competing network of AI data centers comprised of neoclouds around the world, which supply AI compute to a wide number of AI startups and also innovative enterprises— like Shopify, Figma, etc. As these neoclouds often don’t have access to capital—they basically don’t have the money to build these data centers—Nvidia now provides a minimum revenue guarantee on this capacity so that neoclouds can get the necessary loans in capital markets.

We definitely think this is a rational strategy by Jensen and that he has to roll the dice here. The additional benefit is that Nvidia gets bigger upside—once from the hardware sales and a second time from recurring revenues on its infrastructure via the revenue sharing agreement. However, this works both ways. If in ‘29 or so there would be an overcapacity of AI compute, then Nvidia would see its hardware sales go into a cyclical downturn while it could also be on the hook for neoclouds that aren’t generating substantial revenues (for example, in the event that certain startups which purchased neocloud capacity go bust).

So, it’s clear that Nvidia will become an even higher beta play on AI demand, with its risk profile moving up, closer to that of a neocloud. This is great as long as there is an AI shortage, but not so great if there’s large overcapacity at some stage.

Nvidia is similarly providing guarantees for frontier AI labs to build out capacity, and from the 10Q we can see that these are substantial:

“In August 2026, we entered into guarantees with SB Energy Corp. to provide credit support on the land, power, and shell buildout at SB Energy’s PORTS Technology Campus in Pike County, Ohio, covering leases for approximately 4.25 gigawatts of IT load. The campus will exclusively host our compute under 20-year leases to OpenAI, subject to limited exceptions, with our obligation capped at $105 billion in the aggregate, subject to certain conditions including SB Energy, the lessor, satisfying applicable ready-for-service conditions. Each guarantee generally becomes effective upon commencement of the applicable lease, with corresponding guarantee amounts increasing as each of nine data centers is placed in service, which is expected to begin in fiscal year 2029. Our exposure declines as OpenAI fulfills its lease payments. Our guarantees are limited to defined portions of lease and power payments and not the full cost of the site or all of the tenant’s obligations. The guarantees terminate upon certain events, including OpenAI achieving a satisfactory credit rating or after each respective lease term has completed. We also hold an option, exercisable in our sole discretion, to provide additional credit support in phases for approximately 3.8 additional gigawatts as the site scales.”

Nvidia is guaranteeing the bulk of the cost of this site. Jensen mentioned that 1GW of Vera Rubin generates $40 billion in revenue for Nvidia, and thus the 4.25GW campus is worth about $170 billion in Nvidia hardware. With Nvidia guaranteeing $105 billion of this investment, this is about 62% of the cost.

Again, we don’t have a huge problem with this—it’s clear that both OpenAI and Anthropic are generating stellar revenue growth, with both companies combined adding $75 billion in ARR since the start of the year. And as we’re still in the early phases of token consumption, it’s obvious that AI tokens are going to become a multi-trillion industry over time. Note that OpenAI also reports revenues on a net basis, i.e. stripping out the revenue share that hyperscalers take in return for them hosting the models, while Anthropic reports revenues on a gross basis. So, OpenAI will see a huge boost to revenues by hosting their models in their own data centers.

These companies need cash to scale AI compute, and so Nvidia steps in as the guarantor so that AI labs can get the necessary loans from the banks. Again, we think this is the right move by Jensen. Nvidia makes sure that the AI labs who need to scale compute can do so, and by doing this, Nvidia cements its central position in the supply chain and continues to drive stellar revenue growth. This strong position then allows Nvidia to secure a large share of available wafer supply and advanced packaging at TSMC, memory capacity at SK Hynix and Micron, and optics at Lumentum and Coherent. In a supply constrained world, this makes it physically impossible for competition, such as AMD, to grow large amounts of share.

But again, everything hinges on AI compute not going into large oversupply. If in ‘29 or ‘30, we have a large oversupply of capacity, that massive site in Ohio isn’t worth that much. Available GPUs for which there is no demand aren’t really worth anything actually. And if OpenAI isn’t generating the revenues they expected, Nvidia might have to start paying out to cover the loans of that data center.

However, we continue to be extremely bullish on AI demand. And the key reason is that we’re using the most advanced models to write large code bases. So, being a programmer gives us a competitive advantage versus the rest of the market to really understand the progress and capabilities of the latest, most advanced models. And they truly are beyond the wildest imagination. Benchmarks don’t properly capture what’s going on. We were watching the Lex Fridman podcast yesterday with well-known programmer DHH, and his conclusions were exactly the same as what we’re seeing:

“We have AI in the pre-agentic era. I was excited about that, but it was not fundamentally rewriting the rules of the game for me. It was not completely changing how I worked. I was still chiseling code, I just had a little helper, a little sidekick- who could bounce ideas off, and I could look up this information online and so forth in a more efficient way. It was just a more efficient way to do what I was already doing, and it didn’t change the emotional connection I had to the computer.

Then we get to November 24th, 2025. Opus 4.5, to me, was the dividing line, where suddenly I gave it a couple of tasks, and I realize that the quality of the output is uncannily close to what I would’ve written. And I remember just leaning back and thinking, “What just happened?” And again, Opus 4.5 now looks like a retarded model.

I had to be in the driver’s seat. I had to tell it what I wanted, and I had to be the reviewer, the auditor of what was coming out. And then finally now, this summer, with Opus 5, Fable, and Sol, GPT Sol, and to a lesser extent, some of the open weight models, we’ve arrived at a new era where I’m not telling it where we’re going. I’m telling it the problem I have. I’m telling it the fuzzy, vague idea I have. It tells me where we’re going. It tells me which path to take, and I will still look at it because I’m a curious person and I like computers and I like the outcome of it, but I really kinda don’t have to. I’ve become optional in the part that produces the code that picks the route.

You could really get to the 100% of code written by AI and really almost– if you’re a good programmer and you have a good intuition about what’s happening behind the scenes, you can legitimately not look at the code, at least that’s been my experience. I don’t know how much prior experience with programming you need to have to kinda know that things are working correctly behind the scenes, like intuitively by observing the ripple effects, the symptoms of the system.

I’ve been working on Omarchy for the last three months, this version that just dropped a few days ago called Quattro. And almost right from the beginning I’m working on that version, the agent acceleration neared 100%, and in the last two months it has been 100%. I have not written any of the code that’s shipped in Quattro by hand. I’ve reviewed the shape of all of it. I’ve reviewed the individual lines of anything that’s critical in the model layer of the system, and I’ve not looked at a bunch of the UI code. I have not looked at a bunch of the auxiliary code, and I’ve not written any of the new functionality entirely by hand.”

It’s clear that models are now matching the capabilities of the best engineers in the world, and across more and more fields. These models are not only becoming state-of-the-art software engineers and at 100,000x the speed of the best human programmer, but also top notch mathematicians, physicists, biologists, etc. What looks like a certainty here is that there is going to be scientific progress in the coming decades at a pace we’ve never seen before. And thus, the big conclusion for us—we’re going to need a huge amount of data centers. Therefore, we continue to be very bullish on AI demand and that Nvidia is making all the right moves here.

Long term, another key attraction in the Nvidia story is that revenue per GW deployed keeps moving up. Initially, the company was making like $3-5 billion per GW. That grew to $18 billion per GW with Hopper. $25 billion with Blackwell, and now $40 billion with Vera Rubin. And after that, it’s going higher again. Note that this is positive for the entire semi supply chain—from TSMC all the way down to ASML, as this rise in the revenue per GW is driven by more advanced manufacturing processes.

So, if you install 1GW of Rubin, it gives you 30x more tokens versus Blackwell, and at a 35x lower tokens cost. All AI capacity that’s coming online in the coming years is going to be massively more productive than in the previous years, and provide much cheaper tokens.

One big concern in the market were Nvidia’s gross margins, given the explosion in memory pricing, however, the impact of this looks to be very manageable:

“As you are already aware, we are experiencing extreme pricing conditions in memory. The magnitude of the price increase has exceeded our prior expectations and are headed even higher into next year. As a result, we are resetting expectations today. For Q3, we expect GAAP and non-GAAP gross margins to be 74% plus or minus 50 basis points. We expect margins to bottom in Q4 in the 71% to 72% range before settling at 72% to 73% in fiscal year ‘28, as executed price increases take effect in Q1.”

The three key takeaways for us from Nvidia’s call were:

• Revenue growth is accelerating again and AI capex estimates for next year are likely too low.

• Nvidia is becoming a higher beta, and higher risk play, on AI demand as the company is becoming a large guarantor in the global neocloud and frontier AI buildouts.

• Memory prices will continue to move higher next year.

We continue to stay long Nvidia here. We like the acceleration story in revenues, the progress in model capabilities—which are far ahead of what many in the market are thinking—and the stellar revenue growth at the frontier labs. All while the valuation of Nvidia shares remains highly attractive at a 19x forward PE:

Read more

Broadcom has substantially lagged AI semi peers such as Marvell and AMD year-to-date, driven by fears around the company’s market share for Google’s upcoming TPUs. However, we’ve seen this movie before. Up to a few months ago, this is the exact reason why Marvell was trading at a discount—the company was ‘losing the Trainium business to Alchip’ and also ‘a shitco’.

In reality, Marvell has a comprehensive portfolio of advanced data center tech—ranging from leading edge semis and IP to optics. For example, Marvell recently closed a broad custom-silicon agreement with Google covering TPU-adjacent semis (XPU-attach). The consensus view was long Broadcom - short Marvell last year, but the reverse trade turned out to be the smart move.

Also with Broadcom, we think the market is now missing the bigger picture here. Once you start looking past the exact market share Broadcom will have in upcoming TPUs, the long term attraction is that Broadcom is the key ASIC partner for the best frontier labs—with Anthropic and OpenAI—as well as other key hyperscalers looking to compete in frontier AI—such as Google and Meta.

Our thesis is that as both Anthropic and OpenAI continue to produce stellar revenue growth in the coming years, we will gradually see more announcements around their ASIC capacity. This will help Broadcom’s multiple as the market gets increased visibility that Broadcom will indeed be a long term growth story.

However, also in the Google TPU business, JP Morgan sees continuous growth in the coming five years after meeting company management last month:

“While it was widely known that AVGO’s revenues would scale each year under its 5-year contract with Google, management incrementally confirmed that it expects to retain a majority volume share through the course of the agreement and that this was a key reason AVGO entered into the partnership. In our view, this confirms that AVGO is anchored as the primary supplier across the 5-year term, which should ease widespread investor concerns around longer-term share-loss.”

However, the real long term bull case is Broadcom’s partnerships with the leading frontier labs. JP Morgan explains:

“Management indicated that true compute demand (ex-China) is concentrated among roughly five frontier model builders that account for the vast majority of demand. While hyperscaler infrastructure can appear GPU-skewed at the aggregate level, management suggested that, when honing in specifically on leading frontier model builders, the mix could be roughly 50/50 between XPUs and GPUs next year (in line with our view that the shift from training-dominated workloads to inference monetization is pulling forward custom accelerator/XPU adoption).

The team confirmed that customers are attempting to accelerate next- generation ASIC XPU tape-out schedules and are pursuing a wider variety of chip designs. This reflects the rapid evolution of inference workloads from one-shot inference to reasoning and agentic workloads, as well as customer focus on lowering cost per token (which in our view increases the strategic importance of workload-optimized custom silicon).

Management discussed OpenAI’s first-generation XPU custom accelerator, Jalapeno, co-designed by Broadcom, and emphasized that the complexity of developing these cutting-edge XPUs is often underestimated. These products “push the envelope” across multiple design and technology dimensions, including bandwidth, SerDes, power and packaging. Broadcom expects to tape out the follow-on/next-gen OAI chip very soon.

Management pushed back on concerns that hyperscaler customer-owned tooling (COT) efforts will materially displace Broadcom’s XPU/TPU share. The key point was that modern XPUs are no longer simple single-die chips; they increasingly require multi-die configurations, advanced packaging, high-speed SerDes and deep productization expertise. Management referenced 6-die and 8-die configurations as examples of the complexity involved and suggested that internal hyperscaler efforts are likely to face significant technical hurdles.

On rising component costs, including memory, the team is aiming to keep ASIC margins intact “one way or another,” emphasizing that technology leadership is the key source of pricing power. Management also pushed back on the idea of customers consigning HBM, noting that HBM is an integral part of the XPU design and must be tightly connected to the core logic die for the product to perform properly.”

Broadcom’s head of semis showed just how massive these next-gen XPU packages are going to get:

This is Charlie Kawwas, Broadcom’s head of semis, explaining:

“This is an XPU that was built and shipped last year. As you can tell, it is made of different chiplets. Some are memory-based, some are networking-based, and some are compute. What we are building and shipping this year is this chip. As you can see, it got bigger. You can see from last year to this year, you are now getting three times more memory and the compute that is in the middle is almost three times stronger.

What ended up happening is we said for next year, we actually have to build this. So now, you have even more memory. In two years, it is almost an order of magnitude (10x). When you look at the compute, in the earlier version you had one large block, and here you have four of them. And then, this is what we are working on for 2028.

I want you to understand two things. One is the speed at which this technology is moving, and hence the innovation that it needs to go through, is extremely rapid. Two, this is how we build these chips—these are actually modular. This is why people come to us, because we actually take these platforms and pre-build them. These little chiplets that go in the chip, we pre-build them, make sure they are production-ready, work with partners—like Samsung and others—and create these memory blocks. This memory integration is one of the reasons why the prices are going higher, because we are putting so much more in them.

So, you will have 16 times what we had two years ago, 16 times in a single chip. This is ‘The Beast’—yes, actually, that is what I call it at work. We are building four of these for the largest four LLM players. It is modular.

Think of the 2025 chip as being about 1 million units a year. Think of the 2026 chip as being about 2 million units a year. But remember, the 2026 chip is three times the capacity of the 2025 chip, so that is a six-fold increase in compute capacity in one year. Then you go to the 2027 chip, which is probably 4 million to 5 million units, and this is nine times the capacity of the 2025 chip. Now you are at 40 to 50 times the scale of the first one. And then The Beast in 2028 is three times that again.

In the next couple of years and beyond—Yes, there is more than Jalapeno. There are definitely more spicy chips and a lot more XPUs that will be coming out—both from the large labs and a lot of startups. What I see is an insatiable demand in gigawatts.

For example, we have announced for one of the top five customers that this year we are putting in about 1.5 gigawatts with one of the chips I showed you. Next year, we are going to put in over 5 gigawatts, which we have already contracted with that same customer. So, we are going to go from 1.5 gigawatts to 6.5 gigawatts total. The year after that, I am pretty sure it will be over 10 gigawatts incremental. If you look at the next three years for each of these lab players, they are going to go from 1 to 2 gigawatts this year, to 5 gigawatts, to over 10 gigawatts incremental. Over that three-year journey, they will go from 2 to 5 to 10 gigawatts, which is 17 gigawatts total, starting from just 1 gigawatt today.

This means there is a massive challenge in terms of the amount of power that is needed, the amount of silicon wafers that are needed, and the amount of memory that is needed. It is actually going to pose a challenge for the rest of the non-AI businesses, which we also do at Broadcom. Remember, I have 16 franchises. Five of them are in AI, and 11 are in non-AI. AI is not just going to live in data centers—AI has to get to the edge, and these other businesses have to actually enable this edge AI. Just like the internet started from data centers and moved all the way to smart devices today. The same thing has to happen with AI.”

On The Beast, Charlie said that “we are building four of these for the largest four LLM players.” This implies to us that they are also working on an ASIC for Anthropic, and that Anthropic won’t purely rely on TPUs for their ASIC strategy. This makes sense, as you don’t want to source your hardware from a key rival long term.

From Charlie’s comments, it is clear that both 2027 and 2028 will again see strong sequential growth. Based on a recent MOU with Samsung, JP Morgan works out that also 2029-2030 should again be growth years for Broadcom:

“On Saturday, Samsung (covered by Jay Kwon) announced it had signed a memorandum of understanding (MOU) with Broadcom to expand their strategic collaboration across memory and advanced foundry technologies, supporting the next generation of AI infrastructure. We believe that Samsung is and will continue to remain Broadcom’s main supplier of high bandwidth memory (HBM DRAM), and we estimate that around 75-85% of all of Broadcom’s HBM DRAM requirements going forward will be supplied by Samsung.

The companies expect the collaboration to result in MORE than $200B of Samsung memory/ foundry products (we estimate Broadcom’s Samsung purchase mix as 90-95% HBM, 5-10% foundry wafers) over the next 5 years through 2030, and based on our calculations, we estimate that this implies > $1T of cumulative Broadcom AI revenues (ASICs + Networking) and around $40-50B of cumulative Samsung foundry-based chip sales (we believe that Broadcom will utilize Samsung for datacenter networking, broadband, and telco service provider chips).

Our estimate of >$1T in cumulative Broadcom AI revenues (2026-2030) implies a ~60%+ CAGR on Broadcom’s forward AI revenue through 2030 (>20% growth per year even after 2028) — a significant growth profile, but not a surprise given the team’s #1 market share position in AI ASIC XPUs and #1 market share position in switching/routing silicon and its Tier-1 line-up of ASIC XPU customers (Google, Meta, OpenAI, Anthropic, Asia Customer #5, Asia Customer #6, Softbank/ARM, Apple, and SambaNova).”

XPUs will continue to take share in the AI accelerator market, especially in the scenario where frontier labs such as Anthropic and OpenAI remain dominant in terms of AI revenue market share. Conversely, the more workloads move to open-source models, the better it will be for Nvidia.

When there are a large variety of models and model providers, it makes sense to optimize all of these for a key platform which is available on all clouds, i.e. Nvidia CUDA. Conversely, when there are a few frontier labs which are printing cash due to their dominant positioning in terms of AI revenue market share, this gives these labs the resources to optimize their AI infrastructure with their own XPUs. Names such as Broadcom and Marvell will see large tailwinds from this.

Note that the two scenarios are not “either-or”—both scenarios are playing out today and they will continue to play out. More simple or pre-defined background processes for which open-source and cheaper models are ‘good enough’, will continue to be deployed on these types of models. At the same time, the most advanced models will continue to win the most advanced workloads, such as writing large code bases or doing advanced engineering work.

One reason that we have the $200 per month Claude plan is that Claude can do anything very well—so, you’re writing a large code base, but Claude also understands how to translate your app into 20-30 other languages, knows how to architect the back-end system in the Google Cloud, knows how to optimize the app for marketing, etc. For example, when we tried to translate our app with GLM 5.2, ‘the Claude killer’ according to social media, it crashed. Apparently it can’t do translation. Engineers don’t like to lose hours with models that crash or can only handle a narrow subset of tasks—they like to iterate quickly and build solutions. So, even if we got GLM 5.2 for free, we still wouldn’t use it and just pay the $200 for Claude. It’s a real bargain given its capabilities.

It seems likely to us that for advanced workloads, such as engineering, companies will just continue to pay up for the most advanced models. You don’t want to work with an inferior model on a problem for 1-2 days, and then figure out it isn’t able to do it. It makes a lot more sense to purchase the best models and then get the best solutions in an instant. Engineers cost far more than tokens, and so in our view, companies will pay up for the best models to maximize the productivity of their engineers. This is where the real value is created. Not by purchasing cheaper tokens in the cloud.

We regularly see highly confident claims on social media that the “model layer will commoditize”. However, this is not how things end up in tech. Typically, there are 1-2 big winners when it comes to software. Let’s just quickly go through the winners from the last decades to illustrate. Client OS—Windows. Text editing and spreadsheets—Microsoft Office. Server OS—Linux. Browser—Chrome. Mobile OS—Apple and Android. Programming languages—Python and Javascript. Cloud—AWS and Azure. Search—Google. Code repository—Github. And you can go on like this.

In our view, when you look at revenue market shares in the LLM market 3-5 years from now, this will likely be a market where two players are fairly dominant and capture 60-70% of the market or so. Then, there will be two or more players which capture another 25% or so. Open-source and cheaper models will only capture a small revenue share of the market, given that these models are indeed commoditized. While these will process large volumes of tokens, not much money will be made in this part of the market as these will be extremely cheap tokens.

This is good news for Broadcom as these dominant AI players will allocate resources to scale their ASIC solutions. The only caveat is that—when we look at Charlie’s comments above—it’s also clear that by the end of 2028, a huge amount of AI capacity will have been brought online. The key question is whether we’ll still be in a shortage then, or whether we’ll actually have overcapacity. In case we get overcapacity in AI at some stage, this will lead to a cyclical correction in data center capex for a few years—until AI demand grows again into the available capacity. And then we’ll start a new semi upcycle again.

However, when we look at Broadcom’s valuation, we see the risk-reward as very attractive here. The company is exposed to one of the most attractive areas in AI as the key ASIC partner for the leading AI labs, while trading at a very non-aggressive valuation of 23x forward EPS:

And a valuation of 14x 2028 EPS:

Read more

We watched a podcast with Professor Robert Pape from the University of Chicago this week, where he detailed how the Iran conflict continues to worsen when it comes to natural gas and oil supplies.

There are basically three choke points around the Arabian Peninsula. One is the Strait of Hormuz, which is now almost fully closed again. Second, Iran and its proxies could also shut down the other two choke points, on the western side of Arabia, in the Red Sea.

Currently, the Saudis and Emiratis are diverting a small but significant amount of oil via these Red Sea routes:

The problem is that after the heavy fighting of March-April of this year, ‘Operation Epic Fury’, the Iranian regime hasn’t collapsed and is still, in fact, in control of the Strait of Hormuz, and with an ability to also potentially close off the other two choke points in the Red Sea. This is Professor Pape on the Mario Nawfal podcast explaining how Iran continues to escalate its demands:

“The first thing to understand here is that Iran’s demands have been expanding, not day by day, but period by period, since the war began. What you can see is in March through May, is that essentially, this is the period where Iran mostly wants peace, mostly wants a ceasefire. They have a little bit of edginess to their demands for money and reparations, but not a whole lot.

Then it transitions to the memorandum of understanding, or MOU. Once we get into the MOU, now we are in the actual bargaining phase. This is where there is negotiation between the United States and Iran on a series of these issues, and also negotiations with Iran and Oman, and Iran wants a little bit more upfront cash in this particular period.

Now we have transitioned, and I put the date as July 7th because that is when the Ever Lovely vessel was attacked. You could have put it a little bit later as well, maybe the 12th, but that would not change the fundamental analysis here. This is the period of unilateral demands. This is the period that we are in now, and I think we are going to be in for quite some time, where the demands have escalated across the board.

This can be explained fairly directly by the declining oil inventories and other buffers. This is basically the chart of the 32 countries that agreed in the middle of March to release 400 million barrels over 120 days. They are still releasing, by the way. But notice the clear trajectory here, it is going to keep going down as the days and the weeks go on because we are going to run out of the cushions. As this goes down, Iran’s leverage has been going up, and its demands have been going up.

The big thing that happened in the spring, even before the MOU, is the West came up with a set of ideas to bypass the Strait of Hormuz. I know because I have been in discussions with a lot of people from the corporate world, in April and early May, saying “No, there is not a problem here with Hormuz, we are just going to bypass it over two years.” Basically, Scott Bessent just revealed the plan, which is that in two years, Hormuz is not going to matter anymore. The reason it is not going to matter is because the Middle East is going to be riddled with pipelines.

There are already some there now, but it will be riddled with pipelines to get energy and other resources out of the Middle East through waterways that do not connect to the Persian Gulf at all. In that situation, Iran then developed, and this became public at the end of May, what the IRGC calls its ‘resistance security belt’. That is when the IRGC generals, the top leaders, started to explain that Iran did not just want peace; it wanted the security resistance belt. That was control and influence militarily from the Mediterranean, which is where the Suez Canal is, to the Red Sea, where the Bab-el-Mandeb is, and through the Gulf.

Basically, the cat got out of the bag because you cannot start talking about building all these pipelines and have this not become public knowledge. That is why Scott Bessent is publicly saying “Do not worry about us conceding on Iran, we are basically going to put Iran and the Strait of Hormuz out of business in two years.” Well, what that has done, is it has defined the goalposts for Iran’s demands. This is not a negotiation. Iran is not negotiating to get economic relief. That is a secondary issue, because if it gets relief today, that can be taken away in two years.

We are just going to put Hormuz out of business, and then we are going to be able to go back to business as usual, pounding Iran’s leaders. We are going to be able to attack Iran whenever we want to, and we will be able to come in and do more damage to Iran. So, this idea of getting economic relief is temporary at best. Power is what dictates what Iran’s security and future are going to look like, not promises by the United States.

What you are seeing here is, since essentially July 7th, where have the attacks been? This is not just a notional concept of a resistance security belt the way it was at the end of May. They are enforcing it; they are doing it. In other words, Iran has moved from survival in period 1, to ambition during the MOU, to hard power in period three. The current situation is a hard power situation where they are not looking to negotiate. They are looking to establish their security resistance belt.

What would Donald Trump have to give Iran to open Hormuz? Well, you see them in the list of demands from the Supreme National Security Council. Number one, the United States has to pull out all of its forces. So, the United States cannot stop Iran’s military control of the security resistance belt. Number two, the United States would have to give Iran money. In other words, fund Iran’s expansion and rising hegemony. By the way, there is also no discussion here at all about Iran’s nuclear program.

Notice that we are about to end the 60-day period of the MOU in just about six days, on the 17th of August. Well, there has been no extension of that MOU. What that means is there is literally not going to be any constraint whatsoever, on Iran’s nuclear programs in about six days. I think this is why we are on a collision course. This is not just a minor set of issues. We are not negotiating on the edges of how to end a war nobody wants.”

President Trump has left the US, and the world, facing a Gordian knot. Iran appears to have a massive set of demands, which look impossible to agree on—for example, giving Iran control of the Strait of Hormuz and the Red Sea, removing all US troops from the Middle East, giving Iran tons of cash, and allowing them to develop nuclear weapons. It looks like this conflict is currently in an escalation trap where things are only going to get worse.

Former Iranian ministers have already mentioned that Iran might well continue the fight into ‘29 and then negotiate with the next US President. That said, reports are mixed—Iran’s own president and central bank chief reportedly told the Supreme Leader the blockade was crippling the economy, and senior negotiators are openly arguing in public that Tehran needs sanctions relief.

The current US plan is a new blockade in the Strait of Hormuz, with the goal of cutting Iran off from the world economy. Secretary of Defense Pete Hegseth has announced that the US Navy intends to maintain this ‘steel wall blockade’ indefinitely by continuously rotating ships into the region. And US Central Command (CENTCOM) recently established ‘Task Force Falcon Strike’ to enforce the blockade using unmanned drone systems in the air, on the surface, and underwater.

The problem is that if this doesn’t get resolved in the coming months, which now looks likely, oil and natural gas prices will start spiking at some stage—either later this year, or early next year. This will likely force the Fed to raise interest rates, which can be a negative factor weighing on equity markets, especially given that valuations now are reasonably high. In case the Fed doesn’t raise rates and decides to let inflation run, well, it’s probably a good idea to buy some gold as well, which is still the best hedge against high inflation.

However, our view is that Kevin Warsh would raise rates under that scenario. He’s made it very clear that he will bring inflation down to 2%, and so we’re not holding any gold at this stage:

“There is no soft inflation target. There is no soft implicit target, not on this committee’s watch. There’s only a target, and it’s 2%.”

Memory Outlook & The DRAM Market

Memory stocks have been selling off heavily over the last few months. The fears are that DRAM and NAND pricing are now peaking. This also isn’t helped by the fact that a lot of news outlets, and other publications, want clicks and so they will publish sensationalist headlines. One of these is that SK Hynix will give a 50% pricing discount on HBM4, JP Morgan comments:

“Multiple investors have asked whether media reports about the SKH team’s pricing HBM4 for 2027 at a 50% discount vs. competitors are accurate. While we believe investor expectations of a 100% y-y HBM ASP increase are stretched, we believe the reports of a 50% pricing discount are inaccurate.

Our relatively conservative pricing assumption (below a 40% y-y increase from 2026E) is associated with three key points:

(1) We believe it is in memory suppliers’ interest to prioritize LTA products for DDR5/LPDDR5/NAND, which are at a significant margin premium over HBM.

(2) NVDA is the largest customer for SKH, with multiple product partnerships, and the company can approach the customer from a multi-year sourcing perspective.

(3) HBM could be re-priced every year, and there is room for price adjustment again in following years (the SKH management view implies a multi- year shortage and short-term HBM contract pricing might not be the top priority if the company prioritizes three to five years’ worth of LTA volume).

If SKH successfully raises HBM ASP above our expectations, this could be a positive upside risk to our EPS projection.”

Clearly there is high dispersion in pricing estimates, Goldman reckons that SK Hynix will be able to raise HBM pricing by 100% in 2027:

“We expect HBM blended ASP for SEC (Samsung Electronics)/Hynix to reach close to US$ 2.9/Gb in 2027E, which would be roughly an 87%/100% yoy increase, respectively. We expect the like-for-like basis price increase to be roughly 60% for major products, and the remaining growth to come from the mix improvement impact.”

We agree with the bullish outlook on pricing, as AI demand growth continues to surpass all expectations. Elon Musk made the best bull case for the DRAM/HBM market on the recent SpaceX call:

“The limiting factor currently is memory. The memory output is increasing by around 20% per year. Now normally, that would be fantastically fast and amazing for any large mature industry. But, the demand is increasing by 200% a year, maybe higher. So if you’ve got demand increasing much faster than supply, then Economics 101 would suggest that the price increases. It does not decrease.”

Goldman also highlights that the HBM:DRAM trade ratio will continue to increase, which will weigh on both DRAM and HBM supply growth:

“We expect the strong HBM demand led by AI servers to continue to outpace supply, and highlight the rising difficulty in ramping the latest generation HBM. This comes from the rising complexity in using a more advanced node for both core die and base die, as well as moving to a higher layer HBM stack, which results in a lower production yield. The higher trade ratio between HBM and conventional DRAM also leads to incrementally less supply of HBM as we move forward to the new generation of HBM. As such, we expect 2027 Supply/Demand to be even tighter compared to this year, which should help drive a favorable pricing environment.”

This is a nice chart from HSBC and shows that while server DRAM pricing has dramatically increased over the last three quarters due to the supply shortage, HBM pricing on the other hand has been flat as contracts are typically negotiated on an annual basis:

We can also see that $/GB pricing of conventional DRAM is now exceeding that of HBM. This should suggest still a lot of upside on HBM pricing in the coming year. The modeling from HSBC looks reasonable, in our view, especially as supply shortages continue to increase. This is Micron’s CBO at the recent Keybanc conference describing the industry landscape:

“Based on these increased signals of demand from our customers, we now expect that 2027 calendar year will be even tighter than 2026 because the growth in demand that is taking shape is faster than the growth in supply for calendar 2027 on an industry basis. On top of that, we have this issue of HBM growing very significantly. And HBM is needed in these AI systems because in a lot of the workloads, the processor—whether it’s a GPU or an ASIC or a CPU—is sitting idle for 50% of the time, because it is waiting for data from the DRAM. And that is a huge underutilization of an important asset.

And so you need to really get a much higher performance memory bandwidth and a much higher capacity of memory in the system, so that AI can be deployed on an efficient and scalable basis. There is so much demand coming from that. And this HBM increase consequently is creating this 3:1 trade ratio that we have mentioned in the past—where to produce 100 bits of HBM, we have to reduce 300 bits of DDR supply,because there is that 3:1 trade ratio between HBM 3E and DDR.

And when you go to HBM4 and HBM4E, by the time you get to 4E, that trade ratio has worsened to closer to 4:1. So all of this growth in HBM pressures the supply that is left for everything else, which means that the wafer supply has to increase dramatically, and that’s just not easy to do. It takes a long time because the whole industry got itself into a state where most of that expansion was needed to be done in greenfield expansion, which means you go to an empty space, where there’s nothing but trees, and you have to create an entire massive fab cluster there.

So that’s a very challenging endeavor. And while our customers have to build data centers and secure power and real estate; building a data center, no offense to our customers, is considerably easier than building a sophisticated leading-edge technology front-end fab. It’s one of the most complex engineering projects in the world from a construction perspective. And so it takes a long time for these fabs to come online, and this leading-edge technology that gets deployed takes a long time to ramp.

And so because of all these reasons, we don’t have line of sight as to when that supply vector will intersect the demand vector, which continues to escalate with every passing year. And they’re all telling us that the #1 constraint they have today is DRAM. The aggregate demand across all of the segments is at much higher levels than it has been.

Our customers are telling us that despite the fact that the prices are at very high levels, that they’re eager to get more supply. And it’s not just focused on the data center, it’s across different parts of the market. As we engage with customers on these longer-term agreements and near-term supply, their constant feedback is that they are not happy about the extent of volumes we are making available to them. They are signing up in these agreements, but they feel like they’re leaving considerable opportunity on the table across market segments.

Of course, it’s most acute in the data center where quite often, we are not able to meet any more than half of the demand our customers have. And so when we look at all of that, these are like multiyear signals that we’re getting from customers.”

Subscribe now

Another concern in the market is despecking—if customers are reducing DRAM content, some in the market reckon that DRAM is not that important after all. However, when you can’t get enough DRAM to ship your products, it’s logical that you reduce the DRAM content to whatever is available. Micron mentioned that they’re only able to meet half of DRAM demand from the data center market. Micron’s CBO explains despecking:

“In our interactions with our customers, it’s not really driven by pricing on the server side as much as it is driven by just lack of adequate supply. Really, our customers are struggling to get their hands on adequate DRAM supply because the extent of supply that is available to them is not going to enable them to ship the units that they need to ship. Consequently, they are focused on balancing how much capacity of DRAM to put in the system with the number of units that they want to ship. And so that is where the modulation of the average capacity in the system is happening, whether it’s in servers or some consumer product.

Despite all of that, when you multiply the units with the average capacity that our customers are planning, the aggregate demand coming from even those new levels of average capacities is rising despite that. Also, there is this phenomena about system performance getting impacted if there isn’t adequate memory capacity. And that kind of change is going to further reduce the potential utilization that the processor is going to have. That means that there is tremendous opportunity to improve the system performance simply by increasing the amount of memory in the system. Once the system is configured and qualified with a certain amount of capacity, increasing the capacity of memory in the system is an easier lift for our customers, and they would be able to introduce a higher-performing SKU and with a higher level of capacity when they see the additional supply becoming available.

Our customers also find that as they make AI systems do more reasoning and get into agentic modes, the context windows lengthen. That requires more DRAM and it ends up spilling over into NAND flash. And so we continue to see an environment where the capacity of DRAM is heavily stressed in these next-generation AI systems.

So, there is all this latent demand out there that comes from being able to increase the average capacities and create SKUs that are higher performing, able to have longer context windows, and much better capability from an overall system performance perspective. But again, it’s not clear on a multiyear time frame when that additional supply will be available.”

If more DRAM would be available currently, customers would be doing the reverse and upgrading the DRAM content of their AI servers. The problem is that the DRAM just isn’t there, and so this creates huge potential for additional demand as next-gen AI workloads scale up.

Read more

We’re always amazed how many market commentators look at these rankings and treat them like they’re the gospel:

Currently, we do all our coding with Claude inside VS Code—the popular IDE from Microsoft—and occasionally, we try out some new models for evaluation. So, one month ago, when GLM was all the rage on social media, we once again decided to do our own test to see if we can actually use this extremely cheap model.

So, we took a task which Claude was doing for us and then we sent it to three additional APIs—GLM 5.2 Max in Openrouter, the most advanced Nemotron model from Nvidia in Openrouter, and Gemini 3.1 Pro in GCP’s API. Then, we let ChatGPT evaluate the output as a neutral arbiter—Claude Opus scored 8.5/10, Gemini Pro scored 8.0, Nemotron scored 6.5, and GLM crashed so it got a score of 0/10.

This is basically the reason why AI spend—where consumers and enterprises are actually spending their dollars—is currently close to a duopoly between Anthropic and OpenAI:

Palantir made a similar conclusion, and noted to treat all the benchmaxing which is happening with a healthy dose of skepticism, this is the company’s CTO:

“We’re excited for a new era, not one of benchmaxing but of benchmaking. The assumption that the frontier is actually the best performing is just not borne out in practice. Within 24 hours of bringing Nemotron Ultra into our stacks, we found 5 production tasks where a standard Nemotron Ultra model without post-training beat frontier models. That shouldn’t even be possible. The benchmarks are right for what the benchmark is measuring, but those are not the tasks my customers were trying to solve. And so, moving this to an empirical basis is, is how we’re going to accelerate the realization of tokens to real economic value.

This underscores that a handful of common benchmarks can be gamed. That era, benchmaxing has ended. We’ve been beholden to a small number of benchmarks that people have been designing models to and then releasing models saying like, “look how well it does on this benchmark”. But the benchmark has actually almost nothing to do with your business. So how do you figure out how to make the benchmark that represents your reality, what you’re trying to succeed at, what you’re trying to get better at and then see what model makes sense?”

So, we haven’t been able to replicate Palantir’s results with Nemotron far outperforming Claude, but naturally, Palantir is also talking their own book. They want open source models to win so that new solutions from Anthropic and OpenAI won’t be a threat to their platform.

We are positive on open-source models taking share for simple or predefined workloads—e.g. an AI model analyzing a customer invoice, classifying which comments are appropriate for a social media app, analyzing images, customer support, etc.—however, for the really advanced engineering work such as writing large code bases, this will continue to be a consolidated market in our view. The reason is simple: only the largest players will have the capital to purchase and create the largest data sets—including paying professionals to create highly customized and advanced data sets such as investment bankers building financial models, lawyers explaining how to argue a case or select a jury, etc.

Subscribe now

AI is basically a scale game where both the quality and quantity of the data the model is being trained on are two of the most crucial factors when it comes to model performance. Therefore, the thesis that the model layer in the AI stack will fully commoditize will continue to be wrong. Bears on frontier models also cite this token usage chart from OpenRouter, arguing that open-source is clearly winning:

However, it doesn’t make sense to use frontier AI models via OpenRouter as you would have to pay a premium. So, everyone either connects to the Anthropic and OpenAI APIs themselves, or uses the big cloud platforms such as Amazon Bedrock. The OpenRouter chart is only useful to see directionally how fast token usage is growing, and which open-source models are being used. However, to draw any conclusions on frontier model usage, its value is basically zero.

The OpenRouter chart is a reason to take Palantir’s comments on Nemotron Ultra with a grain of salt—despite the model being fully free (Nvidia subsidizes your token consumption in return for you allowing your data to be used in training), people just pay for the other models instead.

Another flaw is that market commentators typically only look at the cost per token when assessing the economics of each model, however, one should also take into account the opportunity cost. Every time you give coding work for example to non-frontier models, the risk increases that it will do something wrong and you just lose a lot of time. A model which makes you lose hours because the code has bugs, or even worse, a bad fundamental design, has a high opportunity cost, no matter how cheap the tokens are.

Token cost is a rounding error against engineer time. A senior dev costs somewhere around $100–150 per hour, whereas an agentic day might burn $30–80 in tokens. So the frontier model only has to buy back 30 minutes a day to pay for itself entirely.

We’re on the $200 per month plan with Anthropic, and it definitely feels like a steal. The value of the plan is like having a small engineering team under your control at any time which delivers high-quality work, and extremely rapidly. We will occasionally continue to try other models—for example, we’ll try out Grok 5.0 Build when the new model releases, and the next Gemini Pro, however, we think that frontier LLMs will be a good business as there won’t be much competition. In addition, engineers also get used to working with a particular model, creating a stickiness effect like in software.

Currently, we simply code everything with Claude, and might consult Gemini Pro if Claude proposes something that doesn’t look like a great solution. Recently, in one of our apps, Claude proposed writing a very cumbersome piece of code, and when we gave the problem to Gemini, it came up with very elegant and simple code to implement the functionality. Claude agreed that Gemini’s code was much better.

Google—Taking Some Profits

As discussed above, Gemini Pro does feel like a smart model to us and we regularly consult it when Claude gets stuck, and occasionally, it comes up with very smart solutions. However, as we’ve been using both Anthropic’s and Google’s products extensively over the last 6-9 months, it does feel to us that Anthropic is massively taking the lead.

A key problem is that Gemini Pro just regularly crashes, whereas Claude is extremely stable. We can work with it on a large code base in long sessions. Claude also just gives extremely robust results—it doesn’t touch anything it isn’t supposed to touch, and just implements the changes you want.

Another key problem with Google is that Search is still 50% of revenues (and more of profits), and this business is likely going to get disrupted. At the same time, Google hasn’t released a new model in ages, and now also some of the best people in AI have just left the company.

The parts of the business we like are GCP, Youtube, and Android; but these are only around 40% of revenues combined. Given the disruption in Search, at 27x forward PE, this is an expensive name:

We’ve kept half of the position for the moment on the hope they’ll release a cool model soon, but Anthropic is just on a much faster innovation cadence. Our impression is that Anthropic is running away with this and that the gap with competition will only increase from here. Anthropic has the cash to build custom data sets, has the talent, and the compute. Google seems to be lacking in execution. We don’t want to write Google off fully, but it certainly doesn’t feel like high conviction.

Optics vs DRAM

Read more

What we liked in the Manhattan Associates earnings call—the company is a leader in warehouse management software—was a 100% conversion success from AI pilot to AI subscription. This is again a nice data point that enterprise AI adoption is progressing, while we’re still in the very early stages:

“Since our launch in Q1, Active Agents have progressed from an early adopter program to now touching over 10% of our active installed base, either through a pilot or a subscription. And while it’s still early, so far, we have experienced 100% conversion success from AI pilot to AI subscription.”

The bear case on software is that enterprises will code complicated software such as warehouse management and supply chain management systems themselves. However, we think that the market is underestimating the expertise that this requires. Vibe coding a Python app to run on your PC is very different from coding an advanced SaaS that can scale in the cloud to handle Walmart’s volumes, while being 100% secure and reliable.

For example, Manhattan also talked about how customers need the company’s forward deployed engineers to implement Manhattan’s agentic capabilities. So, the idea that enterprises will build all these systems and agentic capabilities from scratch themselves looks unrealistic to us. As we’ve predicted over the last year, bookings and top line growth for quality software names are actually accelerating:

“Some examples of these investments include building seamless agentic AI capabilities driven by Manhattan forward-deployed engineers. Three consecutive quarters of record bookings has given us confidence that our go-to-market approach is working. So regarding some of the specifics on our Q2 bookings, sales to existing customers have accelerated. And in Q2, conversions from on-prem to Manhattan Active represented over 40% of our new cloud bookings. Renewals continue to be in line with our full year plan and net new logos represented over 25% of new cloud bookings in Q2, while our win rate metric remained consistent above 70%.”

So, the consensus thesis in the market that vibe coding is going to eat SaaS, there is no evidence yet. Obviously, Claude is a state-of-the-art coder and our productivity is up 50x or so with Fable. What could be a one-two weeks coding project, Fable can code the whole thing in half an hour or an hour or so. However, this is very different from coding a warehouse management software that can scale in the cloud and run Walmart’s warehouses.

Manhattan ARR per customer is also only around $1 million. Basically, if a company would hire a small engineering and support team to replicate the system with Fable, the cost is still way more than what you’re paying Manhattan. So, the idea that companies are going to code supply chain management software themselves doesn’t make economic sense at this stage. It would only work if Fable can do everything—code it all out and then provide support to end-users, with maybe one or two engineers for oversight. Still, we don’t think it’s worth the risk even in that scenario as switching warehouse management systems are complicated multi-year projects that are a massive risk to the business and P&L. Are you really going to risk your business to save $1 million by moving to your own, unproven, vibe coded supply chain management software?

We think that 90% plus of existing customers will decide to stay on Manhattan, and perhaps 5-10% of the rare, top enterprises will write these types of systems themselves with an engineering team assisted by Fable. So, we think that the call the market is making will be the exception rather than the rule. At the same time, automating workloads for customers with agentic capabilities represents a new opportunity for software. This is also what the recent Bernstein survey found:

“CIOs increasingly expect AI workloads to be cloud-based and prefer packaged solutions from established software vendors over building in-house applications. Importantly, most do not expect AI to replace enterprise software, reduce long-term IT budgets, or drive a shift from software spending toward hardware investments.

Although GenAI is a leading investment area, CIOs do not expect to increase spending on LLM vendors such as OpenAI and Anthropic, reinforcing the view that enterprises prefer consuming AI through established software platforms rather than building capabilities in-house.”

Obviously, Claude is a phenomenal model but enterprises will prefer to fine-tune open source models themselves for their particular workloads, while also simply levering agentic capabilities within existing software systems. There is still a big engineering skillset needed to build all these agentic capabilities, and most companies simply don’t have that talent. So, Manhattan is doing well and this business will continue to do well in our view.

The market has started to realize that selling these shares down to a 20x PE didn’t make much sense as bookings and revenue are accelerating:

Subscribe now

Data Management Systems as a Sweet Spot

Manhattan Associates is not the only software name seeing a tailwind from AI. As Needham’s customers checks confirm, data management systems in general are seeing an acceleration in spend:

“Last week, our customer checks indicated Enterprises are definitively using PostgreSQL for Agentic-based, external-facing applications. However, organizations are expressing a more nuanced view of the Database market (compared to investors’ black or white extremes), with the understanding that we are in a polyglot environment. Each database is capable of delivering a portion of the Agentic AI stack.

MongoDB was positively reviewed as capturing a majority of net-new, Agentic-first applications; with customers expecting a yr-yr acceleration in CY26 spend. MongoDB is favored for its JSON document model; offering a flexible schema, promoting Developer velocity, and well-suited for persistent memory use-cases (where the current state of the workload requires memory beyond a single session - like multi-step Enterprise workflows, as an example).”

So, enterprises are increasing their use of various data systems to supply AI agents with the data they need. For smaller applications where engineers are comfortable with SQL, Postgres is a no-brainer. However, if you need vast, unlimited data scaling capabilities and with flexible schemas, MongoDB is the top choice. And if you need advanced analytics capabilities, Databricks and Snowflake are the top choices.

Jefferies notes that Databricks raised more capital as they’re short on GPUs—once again a good indicator of enterprise AI demand. One of our readers had the opportunity to invest in Databricks at $70 billion last year and we told him he’d easily double his money. It’s going to be much more than that. From Jefferies:

“Databricks is raising $3B, led by Coatue, at a $188B valuation – a 40% step-up from the $134B mark set in Dec 2025. Assuming a 65% revenue CAGR from FY26 to FY28, this implies ~17x FY28 rev. SNOW currently trades at just ~13x FY28 rev on a $100B EV. We see SNOW as a key beneficiary of this read-through, as a rising tide in data cloud demand lifts multiple boats; even a discount at 15x would imply a $310 stock and a ~$115B EV. Reiterate Buy.

Databricks’ CEO recently told CNBC that surging demand for both proprietary and open-source models has left the company running out of GPUs across multiple regions. “We nearly exhausted our GPU capacity in Asia, and demand is rising in countries including Japan, South Korea, the United States, and India. We therefore need to acquire a large number of additional GPUs, which requires significant funding. That demand was what triggered our latest fundraising round: we were inundated with customer requests and needed more GPU capacity”.

We continue to believe the leading data analytics vendors, Databricks and SNOW, are best positioned to help organizations make sense of their business data and apply AI to run analytical workflows faster and more efficiently, and we expect momentum to build across the board. Databricks’ Genie and SNOW’s CoCo/CoWork are both gaining traction, with early adoption metrics inflecting. At the same time, we view the two companies’ aggressive investment in AI-driven analytics as making the competitive landscape incrementally tougher for PLTR. In particular, Databricks’ Genie Ontology adds an enterprise context layer and knowledge graph that could encroach on PLTR’s core value proposition.”

We disagree on the final point and don’t think that Snowflake and Databricks are really threats for Palantir. Palantir is an operational decision layer for lower tech-savvy end-users, whereas Databricks is basically an analytics platform for data scientists. At the same time, Snowflake is heavily focused on business analysts with a SQL background. So, we think that all three will continue to see strong growth, and clearly Databricks and Snowflake are hugely benefitting as enterprises are consolidating their data in the cloud, and increasingly using AI workloads within those platforms.

So, we can see that Databricks has built a UI for lower tech-savvy users to compete more head to head with Palantir (image below), but Palantir has been building out their platform for 10 years. So, we don’t think it will be that easy to actually start winning deals against them.

It’s something to keep an eye on though, as this is clearly a UI that all employees in an organization can use to derive business insights from and then make key decisions, while being powered underneath by the powerful Databricks analytics platform.

We like both MongoDB and Snowflake. We continue to hold Snowflake and wouldn’t mind owning Mongo, as both these businesses will see tailwinds from enterprise AI adoption.

Palantir is starting to look interesting as well—EPS will double this year for a company that’s now trading on a 77x forward PE. This is a name we could pick up if the market keeps selling it.

Next, we’ll go through more investment ideas, data points, and insights from across the AI space—including semis, software, and other names. We’re only one-third through this post.

Read more

We’ve written extensively on EDA (electronic design automation) software in the past. For newcomers on the topic, a recent expert call with a VP at Analog Devices gave a good overview of the space:

“EDA varies if you’re doing digital or analog design. Let me start with digital because that’s where a lot of companies are focused. In doing digital design, you start with a specification of what you want to build. That specification is normally in some form of natural language. That specification is then converted into the front-end logic design, where engineers use the different logic components to implement that design. After this design phase is the RTL generation phase.

RTL is actually the code to implement the design—this can be in something like Verilog (a hardware description language used to model electronic systems). That RTL then goes through a verification step. Essentially, we test that logic to make sure it’s working properly. After that phase is complete, the back-end or physical design phase happens, where you physically map the logic design to the chip area that you want to implement. Once the physical design is complete, another round of testing is done to make sure all the logic, all the timing is all still working.

Only after that final phase of testing the physical design is the design ready for tape-out in manufacturing. Those are the high-level steps to build a microcontroller or some form of digital design. Most semiconductor companies will leverage a lot of software tools to help them through these different phases. The two most common partners to work with on this whole workflow are companies like Cadence and Synopsys, and Siemens as a third. They make this much easier by providing an integrated workflow to accomplish these different phases.

We use both companies. Cadence we leverage for a lot of our analog design and Synopsys for a lot of digital design. We’re leveraging their flows pretty much end-to-end—all the way from the front-end tooling for RTL creation, their compilers, their DV (Design Verification) test benches, all of it. These companies provide tools like simulators as well, so that once you start building your design, they provide a lot of capability to simulate that. That’s part of the testing phase.

I’d say that the most painful step, or the one that probably takes the longest and is the riskiest, is design verification (DV). It’s essentially testing the design and like most technology—whether it’s software or hardware—getting all the corner cases, generating what are called behavioral assertions, this is probably the hardest part of any design. We struggle with this constantly to get the proper test coverage. The RTL development simulation is not that complicated compared to the DV test bench at the end. There’s also going to be some challenges on the physical design, the place and route, and all the timing closure, getting it physically to fit within your constraints.

We also leverage different types of IP from these companies, so we license parts of the design. Those individual IPs have existing test benches, that already exists. What becomes more bespoke is when you take their IP and combine them with a bunch of our internally developed IPs, and then you integrate all that into a SoC. The corner cases at the SoC-level are new and you have to create new test cases for each individual device. Think of it as the components are individually tested, but when you integrate them it becomes a challenge.”

Although Cadence was historically stronger in analog, they’ve been closing the gap in digital semi design with Synopsys. This was another good recent expert call with a semiconductor veteran who’s now on the board of Baya Systems (a startup in Silicon Valley working on chiplet connectivity):

“I did start-ups, four of them. Two of them ended up ultimately becoming a part of Synopsys, and then one of them IPOd. You can’t think of Cadence without thinking about Synopsys because it’s a duopoly. If you are designing chips, you are going to use both companies. They have similar histories of how they were started. Some of the same people went back and forth between the companies. Think of them as utilities or tools that, as semiconductor design was starting to use PCs and mini computers, these little tool companies started merging to become what is now Cadence and Synopsys. Combined, they have 80%+ of the market in most sub-segments of EDA. Even semiconductor IP, they’re very big there.

As far as the business model, both of these companies, when you’re dealing with the larger companies, they have a license that might be three years or five years. For smaller companies, it might be something that’s more token based, you generally don’t get all-you-can-eat for all your engineering employees to use the tools. You get a more restrictive license—and probably a higher margin license for the EDA companies—than what the big companies have.

The EDA companies work together with the fabs to create PDKs, process development kits. When TSMC, or Samsung, or Intel, come out with a new process, they work with the EDA vendors to create the kits that the end customers will use to make their designs that target a particular semiconductor process. There’s a ton of engineering work that goes into creating these PDKs. Intel has always sucked at this. Intel has gotten better, but the first time they tried to become a fab, it was impossible for anybody to use Intel’s process, except Intel, because they didn’t have a PDK.”

The EDA Crash

The EDA space has been a very attractive space for investors over the last decades. A duopoly meant strong pricing and high margins, while the recurring nature of software revenues meant that these companies weren’t exposed to the brutal cyclical swings of the semiconductor industry. At the same time, as the cost of advanced semi design exploded in the EUV era, both Cadence and Synopsys enjoyed a massive tailwind to grow their revenues with advanced semi design costs surging node-on-node:

The above tailwinds will remain valid in the foreseeable future. However, Cadence and Synopsys shares are down around 14% since Moonshot’s Kimi K3 autonomously designed a functional semiconductor chip in 48 hours, using exclusively open-source tools. The chip was a 4mm² die running at 100MHz on the freely available Nangate 45nm Open Cell Library—designed, verified, and simulated entirely without human intervention.

A few points to note here. First, the Nangate 45nm Open Cell Library isn’t a PDK and has no foundry behind it. Nangate was absorbed into Silvaco years ago—a small and low-end player in EDA—and the library is basically the “hello world” of open EDA research (constructing a “hello world” program is the typical introduction for students to any new programming language).

The problem with open-source EDA tools, however, is that these tools don’t have access to any advanced PDKs. The EDA duopoly has a large moat—not only is decades of semi design know-how ingrained into these tools, but these tools are also the only ones with access to advanced PDKs. For example, TSMC certifies tools through its Open Innovation Platform, and that list of tools is: Cadence, Synopsys, and Siemens. Then, these PDKs are also only available to TSMC’s approved customers and under strict licensing agreements. Obviously, foundries don’t give away their state-of-the-art IP.

So, what open-source EDA tools actually have access to is a much older tier of the stack—TSMC 65nm, GlobalFoundries 12nm, and Intel 16nm processes. Even if you want to design a TSMC 65nm chip, you still need a TSMC license, so it’s not open to everyone. The only fully open PDKs are very old processes, such as SkyWater 90nm and 130nm, and GlobalFoundries 180nm.

Most of the money in semis is made on the advanced nodes, and there really is no other way than using the established EDA tools if you want to design chips on those nodes. For example, TSMC requires a Synopsys PrimeTime static timing analysis before tape-out. Looking at TSMC’s revenue split, older nodes—where open source tools are competing—are a very small part of the business:

Given the semi industry’s large skew—both in terms of revenues and R&D—towards advanced nodes, we suspect that a high proportion of Cadence and Synopsys revenues are coming from the advanced nodes.

The Massive AI Opportunity

We’ve written extensively on the opportunity for EDA vendors to play a larger role in the semi design flow in the future with AI automation. Goldman recently published an in-depth report with a similar conclusion, and calculates that this represents a large opportunity for the EDA names that the market is missing:

“Consensus has viewed Cadence and Synopsys as “defensive” 12%-15% growers at best, and potential victims of AI disruption at worst. However, our new analysis suggests the shift toward custom AI silicon has exacerbated a structural shortage of chip design engineers that EDA companies are uniquely positioned to monetize with Agentic AI—an incremental opportunity we estimate at ~$3.7bn/year by 2030, which is not reflected in Street estimates and may start to be evident as early as 2H26. We believe consensus reflects the “pre-AI” version of the industry as growth is about to inflect, creating an opportunity in the stocks today.

Chip demand could drive a global shortage of ~72,000 chip designers by 2030. A key industry benchmark study (SIA/BCG, 2022) projected a ~23k shortfall in US chip designers before generative AI. Since then, hyperscalers and AI labs have launched custom silicon programs requiring at least ~2,500 designers. Meanwhile, the supply of human designers grows at just 2%-3% annually. We estimate AI design tools can add ~48k agentic engineer-equivalents by 2030, which can partially close this gap.

We size the Agentic AI revenue opportunity for EDA at ~$3.7bn annually by 2030 based on modest adoption assumptions. We estimate “Copilot-style” AI EDA tools could price at ~$25k per engineer annually, relative to the ~$90–$100k customers spend in EDA per seat. We also think autonomous agents, which both EDA vendors will launch in 2026-27, could handle up to 30% of work by 2030 at ~$50k per agent—or ~25% of the fully-loaded cost of a human design engineer.

Street numbers reflect essentially none—and we believe initial proof points could be evident within months. Consensus models both companies decelerating revenue growth from here, implying zero credit for this new Agentic layer. We estimate that this opportunity could lift Cadence’s revenue growth from ~13% to the mid-20% range (+$2.30 of EPS by 2030) and Synopsys to ~20% (+$2.20). 2H results could provide the first evidence of this growth opportunity. If agentic AI tool adoption gains traction, we believe CDNS/SNPS could re-rate in terms of faster growth rates and higher multiples.”

Is Goldman correct that AI will represent a massive opportunity for Cadence and Synopsys? Or are they missing something? Fortunately, we have a lot more insights that we came across, we’re not even halfway yet through the analysis. We’ll go through these findings next.

Read more

Our view is that we’re in the midst of a long term semis bull market driven initially by the AI data center buildout, which will be followed in the 2030s by the rise of physical AI such as humanoid robotics. TSMC’s CC Wei has a similar view:

“I believe from this day on all the way to probably 2029, 2030, the demand is very strong. Whether in between there’s a dip or not, I’m not very sure. But the trend is so robust that I believe we are witnessing a kind of a new industry. I would like to say the new AI industry, which is so common in our daily life, it’s going to affect automotive, humanoid robots, and also impact all the industry. The amount of money put in by the CSPs alone is a very important new industry to the world, so the demand will be there. And the fundamental thing about semiconductor chips, most of them are TSMC.”

While scaling laws continue to be strong in training:

Most of the growth in AI demand in the coming decades will be in inference:

Due to the combination of scaling laws and then the ramp in inference, demand is currently off-the-charts for TSMC’s leading edge nodes. The company showed this nice chart at their recent symposium which is indicating a steep increase once again in AI accelerator demand this year:

In the years ahead, Wei has to estimate where the real demand is going to be. This is how he goes about it:

“Because all the customers are very aggressive, and that’s the CEO’s job. CEO got to be aggressive. They give me the number of their demand, and I believe they try their best to tell me the truth. So, I put all together, and all the truths together is not a truth. Mark down that word. So, yes, we do a very careful judgment. May not be correct, but we did carefully because this is big money, right? So, we are checking the AI data centers progress, the building, the location, the demand, the racks, we’re checking all that to make sure that TSMC chips will not be put in inventory. This year, we say we increased the CapEx from $52 billion to $56 billion, now $60 billion to $64 billion. And you bet, that will continue to increase. But, we do it carefully.”

Even with the newly announced capex increase, TSMC’s capex-to-revenues ratio remains at fairly modest levels:

Thus, while TSMC is making new heights in capex, there actually remains plenty of room for growth in the coming years:

Despite TSMC being capacity constrained, the company won’t be doing any exuberant price increases like we’ve seen in Memory. Wei took a few jabs at the Memory players:

“Yes, one of my competitors in South Korea, they make a huge amount of money, and I’m jealous about it. What is the wafer pricing strategy for TSMC, and what kind of gross margin we should have? The higher, the better, of course. But we are a partner, meaning that our customers got to be successful. I don’t want to squeeze them out from the market. And besides, we are a very trustable company with our customers. So we don’t suddenly increase our price by 4x or 5x. Your customer cannot survive for that kind of a price increase. So we earn our value, and we make sure that our profit, our gross margin is enough for a long-term sustaining expansion. That’s to the benefit of my customer and TSMC also. That’s our philosophy. So yes, I’m really jealous about the memory companies, 86% gross margin. I would be happy with that. But, anyway, we are very trustable.”

Wei also knows that the risk with massive price increases is that it will attract new competition over time. While manufacturing DRAM is an advanced set of processes, a key reason that the industry is highly consolidated is that it’s both highly capital intensive and, historically, extremely cyclical. While it has been very pleasant to be a shareholder in these names over the last year, the longer pricing remains at these levels, the more competition will get oxygen as well.

We can already see this with CXMT, China’s DRAM manufacturer, which is seeing a boom in revenues as well given that they’re a price follower. This is also a key reason why the SpaceX-Tesla Terafab will focus on manufacturing DRAM. While neither of these two can really pose credible competition in the coming 3-5 years, exuberant memory margins will attract capital to fund more competition in the industry, and these new players can become more credible competition in the 2030s. While CXMT obviously doesn’t have EUV, the more cash that the Chinese fabs can generate, the more cash they can deploy into the Chinese tooling ecosystem for R&D purposes as well.

Chinese tool manufacturers have gradually been climbing the technology curve over the last few decades, and players such as Naura, AMEC, and ACMR are increasingly manufacturing more impressive tooling. While China is currently far behind in lithography and metrology—the more cash the Chinese chip ecosystem has, the more likely it is that China can pose a threat in advanced semi manufacturing 5-10 years from now. While DRAM manufacturers are optimizing for near term profitability, long term this can come back to bite them. NAND is even worse as this industry is much more commoditized with five larger players plus now also China’s YMTC, which has managed to grab a 13% global market share. So, NAND has already effectively become a six player market.

Advanced semi manufacturing is increasingly considered to be crucial for national security. This is why the Japanese government in recent years has been funding the country’s return to leading edge logic manufacturing with Rapidus. With the current astronomical margins in DRAM, we wouldn’t be surprised to see similar announcements in DRAM, with geopolitical powers such as Japan and Europe potentially calculating that they will need to secure supply in the form of a national DRAM champion. Both these regions were still strong in semi manufacturing only a few decades ago, but were eventually disrupted by Korean and Taiwanese rivals. Europe’s last major DRAM manufacturer, Qimonda, only went under in 2009 following the financial crisis.

Also from a geopolitical perspective, dramatic memory price increases are narrowing the advantage the US has over China, as these will slow down the data center buildout, and raise token pricing. As already mentioned, huge price increases are also making Chinese fabs cash rich, which will benefit the entire Chinese semi supply chain due to higher R&D investments.

Subscribe now

So, TSMC’s strategy is really to be the trusted technology partner for all leading edge chip designers, prioritizing long term relationships over short term profits. This is really the beauty of TSMC’s business—while silicon design is much more competitive, TSMC simply takes care of all manufacturing, whether it’s CPUs vs GPUs, or even all the various types of CPUs. Wei explains the company’s strong position:

“The AI market continues to be very dynamic. The emergence of Agentic AI is leading to a resurgence in the role of CPUs in AI data centers, which drives more silicon demand in addition to AI accelerators. We believe this is positive for TSMC as no matter what CPU approach is taken, whether it’s x86, ARM-based, or RISC-V architecture, they are almost all TSMC’s customers. We are already collaborating closely with our CPU customers and working to support them with the most advanced technologies and necessary capacity.”

Customer interest in N2, TSMC’s latest node, is strong—with TSMC having more than 20 customers and over 70 in the pipeline:

TSMC is planning a steep ramp of N2 capacity in the coming years:

Both Fab 20 in Hsinchu and Fab 22 in Kaohsiung are ramping strongly, with capacity this year already surpassing the N3 ramp in its first year:

At the same time, the company continues to grow N3 capacity, with tools now being relocated from N5 to N3 fabs:

Then, in two years from now, the A14 ramp will start, which will again be followed by more advanced and more specialized variants of this node:

A14 offers again higher speed with higher chip density:

Wei mentioned that the lead time for a new technology such as A14 is 5-7 years—the technology has to be developed, tested, improved, and then capacity has to be ramped.

Next, we will go through much more findings and insights on TSMC, leading edge semis, and what’s next for both. Finally, we’ll also review the current semis sell-off, which names are starting to look attractive here, and which names we’d still avoid.

Read more

MEMS used to be a fairly dull corner of the semiconductor industry and so historically, we’ve paid very little attention to this market. However, this is changing as optical MEMS are the backbone of one of the most cutting edge markets in the AI data center with optical circuit switching (OCS). In addition, MEMS will see long term growth from physical and edge AI in end-markets such as robotics, drones, satellites, IOT, etc.

In the AI data center, servers typically rely on layers of electronic ethernet switches to communicate with each other. The problem is that converting signals each time from the optical domain to the electrical one, and then back into optics—i.e. optical-electrical-optical conversion—takes a lot of power, and so it makes sense to keep that signal fully in optics in the first place. This is where OCS comes in.

At the heart of OCS are MEMS-based optical switches, which are basically movable micromirrors. These have a very simple function—redirect an incoming optical signal from one channel into another to reroute traffic. Yole illustrates what this looks like:

Each input-side mirror can be oriented toward a corresponding output-side mirror, effectively creating an MxN optical switching cell. The win is that OCS cuts power consumption by up to 40% to 50% compared to equivalent electrical switches. It also introduces near-zero latency because light moves through the switch at the speed of light in glass/air.

There’s one big drawback however—switching speed. MEMS mirrors are mechanical devices. Moving a physical mirror takes time—typically between 1 and 10 milliseconds, compared to traditional electronic switches that can route data in nanoseconds. Because of this millisecond-scale switching speed, OCS cannot be used for dynamic, packet-by-packet routing. Instead, it is used for a particular topology reconfiguration, for example, for re-wiring the data center fabric to connect huge pools of XPUs for a training run that will last hours or days.

Yole expects the optical MEMS market to grow at a 23% CAGR—OCS is the market in the light blue bubbles:

Silex Microsystems & MEMS Manufacturing

Silex Microsystems is the largest MEMS foundry in the world and was recently IPO’d in Stockholm. The company manufactures the optical MEMS for Google’s OCS. From the founder:

“I founded the Company together with four other PhD students from KTH Royal Institute of Technology in Sweden, with the aim of creating a pure-play MEMS foundry fully dedicated to manufacturing MEMS based on customers’ designs. The Company swiftly outgrew its research laboratory origins, shifted to a new production facility in Järfälla, Sweden, and in 2009, transformed the MEMS industry by introducing the world’s first 200-millimetre (8 inch) wafer MEMS manufacturing fab. We have worked hard to drive the technological forefront of MEMS manufacturing since then. Today, Silex is the leading pure-play MEMS foundry globally. We have built a large customer base and long-term partnerships with many of the world’s leading technology companies.”

MEMS stands for Micro-Electro-Mechanical Systems. In simple terms, MEMS are microscopic machines which combine tiny mechanical elements such as sensors, actuators, and electronics on a silicon chip. MEMS devices usually range in size from a few micrometers to millimeters. To put that in perspective—a human hair is about 75 micrometers wide. Many MEMS components—like tiny gears, mirrors, or beams—are often just 1 to 100 micrometers in size.

MEMS are used in loads of end applications. For example, in a smartphone, MEMS detect when you rotate your phone from portrait to landscape, and they can track the number of steps you take. The same happens in Nintendo Switch controllers—MEMS motion sensors track the user’s movements to then translate them into the game.

Before MEMS, these types of devices were bulky pieces of hardware made of springs, weights and wires. However, as MEMS use the same manufacturing process as in semiconductors—i.e. photolithography, etch, deposition etc.—this allows MEMS to be mass-produced onto silicon wafers for cents on the dollar.

The MEMS universe is categorized by what the tiny mechanical parts actually do:

• Optical MEMS: These use micro-mirrors, lenses, or gratings to manipulate light. Optical MEMS route optical signals through fiber-optic networks and drive the spinning mirrors inside lidar sensors for self-driving cars.

• RF-MEMS handle high-frequency wireless signals such as in 5G infrastructure, smartphones, and radar systems.

• Inertial & Acoustic MEMS measure physical motion, forces or sound waves. These track movement, rotation and sound in consumer electronics.

• Microfluidic & Bio-MEMS feature microscopic channels, pumps and valves to move tiny droplets of liquid with absolute precision. These are used in DNA sequencing, “lab-on-a-chip” medical diagnostics and precise drug delivery systems such as insulin pumps.

• Power & Actuator MEMS are designed to physically move or harvest energy at the micro-scale. Silex utilizes a specialized material called PZT (Lead Zirconate Titanate) in its manufacturing. This is a piezoelectric material—meaning when you apply electricity to it, it physically bends.

Silex is focused on the manufacturing of more differentiated and higher value MEMS. They deliberately avoid commoditized, mass-scale products. The attraction of this business is that once you’re manufacturing a high value component for a customer such as Google, that customer will rarely switch as the intensive R&D to set up the manufacturing process would have to be repeated at the new foundry. However, huge customers such as Google do split their orders, e.g. 75% to Silex and 25% to a secondary foundry. Dual sourcing is common practice across the semi supply chain, with few exceptions.

According to a Japanese source, Silex also pioneered Through-Silicon Vias (TSVs) in MEMS manufacturing (Sil-Via™ and Met-Cap™). Normally, connecting a tiny mechanical sensor to a computer chip requires running delicate wires out the sides. Silex brought TSVs—which are common in semi manufacturing—to MEMS. These TSVs allow Silex to stack a MEMS sensor directly on top of a chip, making the final hardware incredibly compact, hermetically sealed, and perfectly suited for high-vacuum applications like aerospace or advanced medical tech.

The prospectus explains the deep relationship between foundries and customers to bring to market a new product:

“Pure‑play foundries serve as critical enablers for Fabless companies and design‑centric customers. Their involvement often begins at the early development stage, where repeated iteration and refinement are required to reach scale production of the final MEMS. Successful scaling from prototype to production involves coordinated work across design, modelling, process engineering, metrology, and packaging.

As fabrication techniques advance, Silex believes that, in the future, MEMS can become even smaller, more powerful, and capable of performing increasingly advanced tasks, which can drive technology innovations further, making MEMS foundational for the future of smart, connected devices. The pace of technological advancements remains high, and foundries that can adapt their processes to meet the needs of diverse applications gain a substantial competitive advantage as compared to those that are unable to do so.”

So, we also have sort of a Moore’s Law for MEMS, where these devices continue to shrink in size to enable more advanced applications. Overall, this looks to be an industry with high innovation potential and a long term growth trajectory. Silex illustrates innovation in this industry over the last decades:

MEMS are used in loads of end applications—including in high growth markets where edge AI will become important such as robotics, drones, satellites, robotaxis, IOT, etc. From the prospectus:

“Future AR display technologies are currently being served by Silex, who as a MEMS foundry is well equipped to serve the demands required to create very small form factor and complex manufacture of chips that interact with the physical domain. Silex believes it is well positioned to serve emerging device applications beyond the current defined MEMS device market, including AR display technologies that are not presently classified as MEMS devices.

The increasing adoption of MEMS across a broad range of applications has resulted in diversified demand for advanced devices. Underlying markets with rapid growth such as MedTech, AI, autonomous systems, AR and satellite communications, drive both innovation and volume in the MEMS market as new applications require novel solutions. The specialised requirements and innovative nature of these applications are expected to primarily strengthen the growth of Silex’s Focus Market.”

Subscribe now

MEMS manufacturing is still largely an IDM (integrated device manufacturer, i.e. product design and manufacturing within a single company) market:

However, foundries are gradually taking share in this market, and this trend is likely to continue in the coming decades as new technologies will need specialized MEMS design and manufacturing:

Next, we’ll detail further insights into Silex Micro and the MEMS industry, with our thoughts on investing in this space. We’re currently halfway through the analysis.

Read more

While there are many end products that make use of rare earths, permanent magnets are overwhelmingly the key application driving the current geopolitical rush. According to International Energy Agency (IEA) data, permanent magnets account for roughly 95% of the monetary value of global rare earth consumption. The key REEs for magnets are neodymium (Nd), praseodymium (Pr), dysprosium (Dy), and terbium (Tb). These four are also called the ‘magnet rare earths’.

A neodymium-based magnet is the strongest type of permanent magnet available on Earth. However, when you mine rare earths, you have to dig up all elements together. This creates a massive market imbalance—elements such as cerium and lanthanum make up the vast majority of what is physically dug out of the ground. Because these are in oversupply, they are very cheap and often trade for just a few dollars per kilogram. These low value REEs dominate volumes for low tech applications such as catalysts, glass, and polishing powders.

Pure neodymium is a soft, silvery-grey metal. On its own, pure neodymium is not a useful magnet. NdFeB (Neodymium-Iron-Boron) is the specific alloy created by combining neodymium with iron and boron. This combination locks electrons into a specific structure with high magnetic energy. As a result, geopolitically crucial end markets such as robotics, aerospace, electronics, data centers, and EVs will be dependent on NdFeB magnets:

The big problem, however, is that up until recently, all rare earth refining was done in China. For example, in 2024, the US accounted for 16% of global rare earth production, however, these rare earths were then shipped to China for refining and magnet production. Needless to say, this gives China strong leverage over the US as they can halt, for example, manufacturing of US military equipment, and also block new advanced industries such as humanoid and industrial robotics from scaling up.

Goldman shows how China in 2024 had a 59% market share in rare earth mining, but an over 90% market share in both rare earth refining and magnet production:

Goldman estimates that NdFeB magnet demand could easily triple in the coming ten years. There is clear upside to this estimate if new industries such as robotics, edge AI, AI compute in space, and EVs take off. Due to environmental and other regulatory hurdles, limited access to high-grade deposits, high capital costs, and long lead times, it’s likely that there will only be a few players in the West that can provide rare earth mining, refining, and magnet production—i.e. fully vertically integrated mine-to-magnet operators.

Subscribe now

As Western governments and big tech players such as Apple are now willing to pay premiums for non-China sourced magnets, and sign long term supply contracts, we see the long term outlook as attractive in this space. This is becoming less of a commodity market, and more of a high growth market with floor pricing guaranteeing attractive margins.

Next, we’ll go through key assets in the magnets and rare earths value chain, and we’ll review the key players involved. We’ll also discuss future TAMs, profitability, and potential returns.

Read more

We enjoyed Google DeepMind researcher Yao Shunyu’s interview on the Zhang Xiaojun podcast. Previously, Yao was at Anthropic where he helped train Claude 3.7, now he’s involved in the Gemini 3 series. We’ll give the highlights below:

“The pace of model improvement is not slowing down at all. It is hard to quantify the velocity curve because if you only look at standard benchmarks, progress will naturally appear to slow down as scores approach one hundred percent. However, the qualitative difference felt by the user does not scale linearly with benchmark gains. Going from seventy percent to seventy-five percent on a complex task can feel like a much larger leap in usability than going from fifty to sixty percent. As a researcher, my personal observation is that the models’ fundamental capacity to learn is getting stronger. Tasks that used to require massive engineering effort to teach a model can now be achieved with far less friction. The bottleneck is no longer the model’s capacity to learn, but our ability to clearly define the problem and construct high-quality data environments.

My experience is that pre-training scaling has not hit a wall, and I do not see any signs of it plateauing. When people claim the Scaling Law is reaching its limit, it is usually due to one of three things. Either they believe the physical limits of the paradigm have been reached, or they believe we have run out of usable data to scale further, or they simply have an unnoticed bug in their own implementation. In the vast majority of cases where teams hit a wall, it is due to the third reason. In this industry, fixing a single subtle bug in your scaling science or data pipeline can yield progress that far outpaces any fancy algorithmic trick. When you run into a bug, if you believe the paradigm is limited, you declare that you have hit a wall. If you believe it can be fixed, you keep digging. This is why systematic problem-solving, which both Gemini and Anthropic do exceptionally well, is so critical. You must design clean ablation experiments to isolate and test every variable.

Coding has two massive advantages that make it perfect for AI. First, it has a highly objective and well-defined reward signal. If you ask a model to write code to implement a specific feature, you can easily run tests to see if the input matches the expected output. The feedback loop is immediate and unambiguous. Second, coding has an incredible, decades-long data foundation in GitHub, which provides a massive repository of high-quality code written by excellent programmers. This allows researchers to build highly robust simulation environments. From a product perspective, the user requirements for coding are relatively singular and standardized. Unlike social media or gaming, where user tastes are highly subjective and diverse, excellent programmers generally agree on what constitutes clean, concise, structurally clear, and well-abstracted code. This consensus makes developing coding products much more straightforward. AI can write specific code blocks incredibly well, but breaking down a massive, ambiguous project into actionable, modular steps is still a human-dominated skill—at least for now.

A conservative estimate is that over ninety percent of the raw code I use is model-generated. If I am being less conservative, it is closer to ninety-nine or one hundred percent. However, I still spend a significant amount of time reviewing, editing, and verifying the logic to ensure it aligns with my exact design. The core of software engineering has shifted from writing syntax to designing system architecture, defining logic, and providing the model with the right context and reference files. The portion of code that the model cannot write is shrinking rapidly. Compared to eighteen months ago, the efficiency of implementing ideas and running research experiments has increased by twenty to fifty times. You can run multiple experimental ideas in parallel, and let the model monitor the runs and analyze the results. This has actually made my working hours longer and more intense; because the iteration cycle is so fast, there are constantly more ideas I want to test immediately. In the generative AI field, no one is coasting. If you are self-driven and passionate about the technology, the pace is incredibly demanding.

The absolute capability gap between Chinese and Western AI models is narrowing, and Chinese model developers have gotten incredibly good at building competitive models despite facing severe hardware and compute constraints. However, whether they can completely close the gap or surpass Western models remains an open question. The compute disadvantage has forced Chinese labs to become highly innovative in other areas, such as model distillation and multi-agent training systems. Model distillation is an open secret in the industry, but there is a major difference between brute-force distillation and smart, soft distillation.

Brute-force, or hard distillation, involves simply taking a massive batch of outputs generated by a model like Claude and training your own model directly on those tokens. Commercially, this is questionable, and intellectually, it is somewhat foolish because it shows that a team does not know what they want to build; they are merely copying someone else’s distribution to score higher on standard benchmarks. On the other hand, soft distillation involves using other models as assistants or evaluators within your own proprietary data generation pipeline. This is a fascinating scientific approach. It has effectively turned some Chinese labs into pioneers of multi-agent training systems, as they are integrating diverse model distributions into a single cohesive training ecosystem. ByteDance’s Doubao is quite distinctive; while it may not match the raw reasoning intelligence of Gemini or Claude, its real-time voice generation is arguably the best in the world. They have optimized heavily for their specific consumer user base, prioritizing low latency and engaging user experience over raw benchmark performance.

Before Claude 3.7, post-training was largely a process of patching up pre-trained models rather than a scalable paradigm of its own. The breakthrough came when the industry figured out how to scale reinforcement learning by utilizing clean, objective environments with clear feedback signals—specifically software engineering environments. The core development of Claude 3.7 took about four to five months. My work was focused on the algorithmic stability of large-scale RL and building the modular training environments. However, AI development at this scale is a massive, highly collaborative systems engineering effort. The era of individual heroism in AI is over; the technology is an unstoppable wave, and researchers are simply surfers trying to catch the crest. If one individual does not make a specific discovery, another team in a well-funded lab inevitably will a few weeks later.

Anthropic has a highly effective, top-down decision-making structure. Because the core technical leaders like Jared Kaplan and Sam McCandlish are also co-founders of the company, they possess immense technical credibility and corporate authority. When they receive a clear technical signal, they can instantly align the entire company to go all-in on a specific bet, such as prioritizing coding capabilities. This is very difficult for other companies to replicate. At OpenAI, internal political struggles disrupted this alignment, and at Google, the organization is historically much more bottom-up and distributed. Anthropic’s co-founding team has a deep level of mutual trust built from fighting battles together since their days at OpenAI, which keeps the organization incredibly reactive and focused.

I left Anthropic due to several factors. First, Anthropic grew very rapidly, more than doubling its headcount to nearly two thousand people while I was there. This rapid growth introduced significant cultural dilution and internal friction. I personally dislike cultures where people spend their days on Slack talking about grand, abstract principles rather than doing the hard, meticulous engineering work required to implement ideas. Second, I disagreed with the increasingly extreme and emotional anti-China political stance pushed by Anthropic’s leadership. Finally, Anthropic is highly focused on a narrow product set, meaning they do not invest in low-level infrastructure research or multimodal generation. I wanted to learn from a broader set of human minds and have the freedom to explore diverse research directions, and Google DeepMind offers an unparalleled environment for that scale of research freedom.

If your goal is to immediately force your specific idea into a consumer product, Google’s massive size and complex release processes can be highly frustrating. But if your goal is research freedom and accessing world-class engineering talent across a massive spectrum of technologies, there is no better place. Google’s technical reserves are unmatched. Additionally, OpenAI’s early success actually saved Google by waking the giant up. It forced Google to reorganize and streamline its decision-making, transforming pre-training into a highly disciplined, predictable engineering process that operates completely within Google’s comfort zone.”

This bodes well for the future of AI investment—scaling laws are alive and well, models are becoming intrinsically more intelligent, and so AI will continue to be able to solve a wider diversity of problems, and more complex ones. However, going forward, capex growth rates will moderate from the hyperfast expansion we’ve seen over the past few years:

Besides the law of large numbers, another key reason for moderating growth rates is that the four largest hyperscalers are now hitting their limits in terms of what they can afford without having to raise large amounts of debt or equity:

It’s likely that in the coming years, hyperscalers will continue to expand capex as operating cash-flow continues to grow, as raising large amounts of debt and equity to go free cash flow negative will likely impact their share prices, as we’re seeing with Meta for example.

Meanwhile, also neoclouds are getting stretched in terms of what they can spend on data centers:

Due to the huge surge in AI capex over the past three years, the strongest outperformers in the market have been the AI capex beneficiaries. However, in our view, the party isn’t over yet for semi investors as once the growth in AI data center capex slows down, it’s likely that a new S-curve will start to emerge in edge AI. One obvious application of edge AI is inserting intelligence into a smartphone.

Apple, which was still TSMC’s largest customer last year, knows that it has to get its AI strategy right. Missing the boat on edge AI—for example, with Google or Samsung launching a much better AI smartphone—could pose a big risk. In order to run a smart AI model on the edge, loads of DRAM are needed. This means that Apple will have to use its cash reserves to book new capacity at the three DRAM players. From the Wall Street Journal:

“Compounding the issue is Apple’s need for additional DRAM to support more AI features, including a rebooted Siri announced last week. And the company has long used NAND storage upgrades to boost profits, charging $100 to $200 for extra increments that cost it just a fraction of that. In the interview, Cook said Apple stands ready to use its cash reserves to boost memory supply. “We’re willing to use our balance sheet to help be a part of the solution,” he said. “Obviously, more capacity is needed.”

Mobile is still massive for the DRAM manufacturers—at 32% of revenues, Mobile & Client make the same contribution to Micron revenues as the cloud:

For investors wondering where to position in memory, our view continues to be that DRAM is much more attractive than NAND. DRAM is really the fast memory that enables AI, whereas NAND is basically storage where you store the chat history, or offload the KV cache as it grows too large.

In addition, DRAM is a highly consolidated three player market with more differentiation in HBM, whereas NAND is a more commoditized product with five significant manufacturers, increasing the risk for overcapacity at some stage. So, we expect DRAM to continue to enjoy the largest price increases:

Besides mobile, also physical AI (robots) is likely to become a large semis market in the 2030s. If, ten years from now, 100 million humanoid robots are sold per annum (vs a smartphone market which is currently at 1.25 billion units annually), this would translate into another semi TAM of $600 billion. For comparison, Nvidia will be doing $390 billion in revenues this year.

Yao Shunyu explains the current challenges still to develop humanoid robots:

“I have looked at some of the humanoid robots and their pricing; the hardware supply chain in China is incredibly mature, making these physical platforms surprisingly cheap. However, on the software side, robotics is still very much in its feature engineering era. Robots can be optimized to perform incredibly well in highly specific, deterministic environments, but they lack generalizability. If you train a robot to fold clothes or pour water, that capability does not naturally abstract and generalize to other tasks. Language models have crossed that critical threshold where scaling up the base model lifts all downstream capabilities simultaneously. Robotics has not yet reached its GPT-1 moment; we have not figured out how to truly scale the training of physical action models. Many robotics labs are now experimenting with using multimodal language models as a base, training vision-language-action models to bridge this gap. Visiting these labs is incredibly fun—they are far more dynamic than language model offices, with researchers constantly collecting physical data and controlling machines.”

XPeng’s founder goes deeper on the various challenges in robotics, although he estimates the company can start mass production of its humanoid robot ‘Iron’ next year already—these are the highlights from his interview on the Zhang Xiaojun podcast:

“What works for digital AI cannot simply be copied to physical AI. It is a completely different world. When digital AI companies try to define the physical world without ever having run a physical business, they build physical AI in a very narrow sense. The actual physical world involves complex human interactions, environmental variables, regulatory compliance, and commercial viability. You have to think far broader than just having a couple of strong features. That is why digital AI CEOs cannot easily explain the transition to physical AI; they are still discovering if they can even succeed.

XPeng’s robotics journey actually spans three distinct stages. The first stage was from 2018 to 2020. It was an independent team of about four or five companies in China exploring quadruped robotics. The second stage was from 2020 to 2023. Over those four years, we built three different milestones. We tried making robots using traditional robotics methods and even tried making robots the way we make cars, stitching various elements together with mixed success. The third stage began after 2023. When we saw the progress of foundation models in late 2022, our entire logic changed. Previously, we believed it was impossible to build a successful robot brain because the complexity of the cerebellum—maintaining physical balance and motion control—was too high.

Today, many companies claim they have developed a robot cerebellum because their robot can walk forward slowly at a monotonous pace. That is not a cerebellum; that is just a basic spine or brainstem maintaining balance. They are far from achieving true cerebellum functionality. Many of the robot demos you see from other companies today are still utilizing what we consider third- or fourth-generation technology stacks. They are just running basic tests. At XPeng, we have the patience and courage to invest for the long term. A quick demo means very little. It is just like back in 2017 when China had hundreds of autonomous driving startups showing off level 4 data; very few of those technologies ever translated into real, commercialized value.

Cars have a highly mature process spanning from planning and design through start of production. By the end of this year, we hope to transition our robots into an automotive-grade start of production process. By 2027, we expect high-level robots to enter their first true year of commercial mass production globally. At that point, traditional motion-controlled robots will begin to decline as advanced physical AI robots take over.

From the end of last year through the first half of this year, our robotics department alone hired nearly eighty master’s and doctoral graduates from top-tier institutions. They are incredibly expensive, but we are fully committed to supporting their long-term exploration. We believe that we must use super-smart people to solve super-difficult problems, rather than relying on rigid processes and predefined tools.

Starting a robotics business is vastly different from starting a car company. I believe starting a robotics company is twenty to one hundred times more difficult than building a car company. Even with XPeng’s established manufacturing and AI capabilities, our success last year only increased our overall probability of success by a small margin. Today, there are over two hundred robotics startups established in China, which is double the number of electric car startups we saw during the peak of the automotive boom. However, unlike cars, which are primarily categorized into passenger, commercial, or special vehicles, robots will have countless classifications, ranging from medical and freight transport to cargo inspection. Many of these specialized robots do not need to be humanoid. But on the path of universal bipedal humanoid robots, ninety-nine percent of companies will fail. The software complexity is simply too high, and there is no open-source platform capable of helping a robotics startup build high-fidelity physical AI software. We are focused on navigating these pitfalls ourselves because you only truly understand the engineering challenges once you step on the traps.

Traditional automotive motion control has been largely commoditized over the last century, with car companies purchasing motor controllers from Tier 1 suppliers. But to make a robot truly capable, its motion control must be far more integrated and responsive than a car’s. In a car, different control domains are isolated. This works for standard driving, but if a car has its left tires on snow and its right tires on grass, executing a forty-seven-degree turn with a person suddenly appearing in front of the vehicle is incredibly difficult. Managing four-wheel traction, movement balance, and latency under those conditions remains a challenge.

For a robot, the complexity is multiplied. A human has over two hundred joints, allowing for an infinite loop of potential movement combinations. Replicating this full-pose personification using AI-driven motion control rather than rule-based software is incredibly difficult. We want to give our robots the same physical instinct a human has when walking on snow, ice, or grass, sensing friction and adjusting balance instantly. This is why we develop eighty percent of our robot hardware in-house, including our actuator joints and hands, while cooperating with upstream Tier 2 suppliers to scale quality.

Once a robot achieves generalized software capabilities, its market adoption and production scaling will happen far faster than cars. Cars took a century to scale because we had to build global road infrastructure, establish traffic regulations, and manage highly complex manufacturing logistics. Robots can be deployed instantly into existing human environments. If the software is ready, physical AI will scale rapidly.

Over the next few decades, universal humanoid robots will become deeply integrated into human lives. While digital AI is limited to assisting in a few dozen white-collar roles, physical AI robots can address hundreds of physical roles, especially in aging societies. I believe the two most critical factors for the future of humanity are medical AI to help the elderly live longer, healthier lives, and physical AI robots to provide care and companionship. For many elderly individuals, a robot may eventually become their primary support system.

We chose this difficult path because we analyzed the fundamental flaws of other forms. For example, we previously developed quadruped robots—dogs and horses. But if you bring a robot dog that is over a meter tall into a standard home, it cannot function. It cannot turn around next to a bedside table without scratching the walls or damaging the bed. Unlike a real golden retriever whose tail and body are soft, a rigid robot dog will inevitably cause damage. If you make the robot dog smaller, its battery life becomes too short, and its utility drops to basic companionship.

If you build a massive, 1.8-meter bipedal robot covered in heavy metal armor, even its own designer will feel an intense sense of physical oppression standing next to it. You will naturally worry about it falling, overheating, exposing high voltage, or simply being dirty. If adults feel that way, how will children and the elderly react? How do you resolve those safety and legal regulations in a household setting? While industrial robots can utilize that design because they operate in controlled environments, household robots must be designed differently. They must be comfortable for humans to interact with. That is why our current generation of robots is designed to be around 1.69 to 1.70 meters tall—a height that is comfortable for both men and women. They are designed to wear clothes and can even have hair, but they must not have a realistic human face to avoid the uncanny valley effect and other complex sociological issues.”

Next, we will dive into a small company that’s an advanced silicon player in edge AI. We see the long-term risk-reward as very attractive in this name, and it reminds us somewhat of investing in Nvidia in 2015, when AI and data center GPUs were only at their very early beginnings. Eight years later, demand fully exploded. It’s likely that a similar scenario will play out when it comes to edge AI and robotics.

Read more

ASML’s tone has shifted from estimates where demand will be in the coming years to how much capacity they can provide (source: JP Morgan conference). Due to manufacturing optimizations, the company can probably ship 110 EUV tools in the coming years without building new clean rooms.

We can see that EUV shipments showed lackluster growth as from 2021. This was driven by two key factors—the industry digested the capacity which had been added during the covid semi boom while TSMC also moved to the 2nm node, which didn’t see an increase in litho intensity. Because TSMC didn’t dramatically increase the number of EUV exposures per wafer at 2nm compared to 3nm—as it moved from the FinFET architecture to gate-all-around (GAA)—the litho intensity on this new node flattened. Bears assumed that if new nodes won’t require steep increases in EUV anymore, ASML’s high growth era would be over.

However, as transistor scaling resumes post 2nm as we move into the Angstrom era with the A14 and A10 nodes, EUV usage node-on-node will again see substantial increases. The same is true for new DRAM nodes, and ASML actually expects the steepest growth in litho exposures node-on-node in DRAM:

At the same time, AI is massively DRAM intensive (especially in HBM), which is driving an unprecedented wave of capex growth in the industry:

While also TSMC will grow capex by 37% this year. We continue to think that the sell side is underestimating TSMC capex growth in the coming years. First, TSMC capex growth rates are still below growth rates during previous upcycles like 2019 and 2021:

Secondly, also TSMC’s capex-to-revenues ratio remains at cyclical lows:

We think that TSMC has a lot of gas left in the tank and that the risk-reward is strongly to the upside here for semicap names such as ASML.

Overall, we see a scenario with steep capex growth at both the DRAM players and TSMC playing out, while we also know that litho intensity at their coming nodes will continue to go up. Meanwhile, analysts continue to model fairly tepid revenue growth for ASML in the coming years, with much lower growth rates than during previous semi upcycles:

If we model in unit shipments from the JP Morgan conference, we already get to EUR 63 billion in revenues, 19% above consensus, giving a 27% revenue CAGR as from 2026:

Below is the model we originally built based on ASML’s capital markets guidance, although we added a very high scenario due to the current demand boom in AI. As a result, ASML will likely move towards EUR 64 of EPS much sooner than 2030, possibly already in 2028:

Bottleneck investing has become popular and ASML is an obvious bottleneck in the AI value chain—it is the monopolist provider of both EUV and immersion lithography equipment. We continue to be ASML bulls and hold this name.

Next, we’ll go through further findings in leading edge semis from this week, a number of stocks we’re purchasing at these levels, including new names, and our current top conviction list in semis.

Read more

The AI data center is splitting into two. On one side, you have Nvidia’s completely integrated kit—the extreme co-design that Jensen loves talking about. On the other side is the ASIC world where hyperscalers are designing custom solutions to get access to cheaper tokens, and to not be fully reliant on Jensen’s shop.

Marvell’s VP of Investor Relations is a semiconductor veteran and he tends to give much better insights on what’s going on at the company and in the industry, as opposed to CEO Matt Murphy who’s gotten a bit too salesman-like for our taste. All he does these days is talk about how pumped he is and how big the numbers are going to be in the coming years. So, the IR did a much better job detailing Marvell’s strategy in the AI data center at the Evercore conference, and we’ll go through some key insights below.

First, Marvell’s strategy is to become a bridge between the ASIC world and the Nvidia world in the AI data center:

“If you think about hyperscalers, they’ve really got 2 different infrastructures, and they don’t talk to each other. In the long run, that’s not how you want to build a data center. You want complete fungibility. That’s where Marvell comes in, with our networking IP and our custom IP. We have the bridge between both those worlds. Now, you can have hyperscalers being completely fungible in terms of how they design their network. So, we see that as really opening a much bigger TAM for the two companies combined [Nvidia and Marvell].”

A big theme in AI is scale-up—linking more XPUs together and make them behave like a single XPU—and Marvell’s goal is here to become an integrated shop which can offer all the various tech involved, whichever way the hyperscaler wants to go:

“Scale-up is almost a perfect example where a customer can work with us on the entire rack-scale infrastructure upfront. We can design all the chips for them, the entire signal path, give them different optimization paths, and this is a fairly unique ability. So, when you want to build a scale-up network, there’s 3 key components. There’s a compute engine, the XPU. There’s the interconnect, which is copper today going optical over time. And then the switch. We have all 3 pieces—we build XPUs, we’re investing in switches, and we have the leading photonics technology in the market.

On top of that, each hyperscaler is looking at multiple different alternatives, this is where our breadth of technology and investment lets us do multiple flavors. So we’re not making a bet on any one single thing. In optics, we have 3 different modulator technologies for doing scale-up optics. It’s MRM, Micro-Ring; MZM, Mach-Zehnder; as well as EAM, Electro-Absorption Modulator. We’re also investing in even more exotic technologies, whether it’s microVCSEL, or microLED. So, pretty much all options on the optical side are available. Even on copper, we can do co-packaged copper (CPC) if that’s what customers want initially as they go towards NPO and CPO.

Similarly, on the protocol side, on the switch side, we’re not restricted to offering one type of technology. We have a UALink switch coming out now, it’s 115T product. We have an ESUN (Ethernet Scale-Up Networking) product, which is based on a 100T platform. And of course, with NVIDIA, we have the NVLink platform.”

What should become a crown jewel in interconnect is recently acquired Celestial AI, which developed a photonics fabric that can connect hundreds of XPUs across racks and make them behave like a single XPU:

Celestial AI has a big win with a major hyperscaler (Amazon) and its photonics solution is on track to ramp in ‘27 and ‘28, with already $1 billion of guided revenues for calendar ‘28.

Deepening its portfolio of advanced optical tech, Marvell also acquired Polariton. This is a spinoff from ETH Zurich that has developed a unique solution to increase modulator speeds with 10x.

The bear thesis on Marvell over the past years was that their stronghold in optical PAM DSPs—where Marvell dominates with a 70% market share—would get disrupted by Broadcom. However, we’ll make two key points here. First, this is a high growth market and so even with share losses this would remain a high growth business for Marvell. Secondly, hyperscalers want multiple suppliers and so we were always skeptical on the thesis that hyperscalers would shift even more of their spend towards Broadcom, which is already the gorilla in datacenter networking.

On the other hand, we see hyperscalers keen to work with a variety of niche silicon suppliers, including Astera, Credo, etc. So, our thesis remains that hyperscalers will continue to spread their budget among this wider group of networking silicon providers, as opposed to consolidating their spend with Broadcom. Marvell’s IR mentioned how their interconnect business remains on fire:

“Our interconnect business is growing at 70% plus for this year. Within that, there’s really 2 big underlying drivers. There’s a scale-out business that’s growing even faster, that’s our PAM DSP and TIA driver business. And then you’ve got scale across, which is just starting and is going to get a lot bigger in the outer years. So, the short to medium term from a revenue perspective looks very strong. And then what’s even more exciting is all the engagements we are seeing a little bit further out in time.

On competition in DSP, the basis of competition in these markets is always “were you first to market?” That’s been the case in every single generation of PAM4, and it’s no different at 1.6T. The largest part of our interconnect business is our PAM DSPs, and next year, 1.6T is a huge part of it. The reality is I don’t think we see anything significantly different. We’ve had a leading position, and I don’t really see that changing going forward.”

Marvell is supplying the DSP, the driver, and the TIA:

And with silicon photonics the company can even sell an extra die (a SiPho PIC, or Photonic Integrated Circuit):

The other bear thesis on Marvell was that they didn’t know what they were doing in custom ASICs and would lose those programs to Broadcom. However, also this bear thesis took a massive dent over the last months as Anthropic is currently training the most advanced LLMs on Trainium clusters. At the same time, Amazon is massively scaling up its AI capacity and is doing this on a mix of Trainium and Nvidia infrastructure. The key platform in AWS’ AI strategy is Bedrock and that is largely running on Trainium. Hyperscalers are extremely good at writing intermediary software layers to run workloads over a variety of underlying hardware, something which will only become easier with Claude Code.

Thus, all data points are pointing to Trainium being a massive success, and Marvell will remain a key partner in this program going forward. JP Morgan notes that Marvell is working on the design for the 2nm Trainium 4 chip, as well as a large number of other ASIC programs including the Microsoft Maia XPU:

“In custom AI ASICs, the next-generation Trainium 3 XPU ASIC program is starting its initial shipment ramp this quarter, with volumes increasing in the second half of the year. We believe the team [Marvell] is deep in the design phase of the next-generation Trainium 4, 2nm program. Their next Tier-1 XPU customer (Microsoft, Maia 3nm XPU program) is progressing well, with plans for high-volume manufacturing next year. Relative to the Microsoft demand forecast, we believe Marvell is still being very conservative (FY28) and taking a significant haircut to the actual forecasted numbers. The team has 10+ XPU attach programs ramping next year (DPU, CXL, networking). Overall, we are impressed with the strong multi-year revenue outlook and the diversity of customer program ramps, and we see a solid setup for CY26/ CY27.”

Evercore’s background checks point to Marvell being a key partner for these programs:

“Over the last 6 months, we talked to about a couple of dozen sources amongst the hyperscalers. And very consistently across the group, what we heard is that Marvell is being viewed more as a more strategic partner supplier. And that’s because the broad portfolio of IP, not just on the XPU side, but also on the networking side.”

Subscribe now

Marvell’s IR explains their role in XPU design, versus where competitors in Taiwan such as Alchip play:

“The reason why customers are coming to us is because there’s a lot of our interconnect technology when you build large complex XPUs. These XPUs are no longer monolithic single-chip devices. These are multiple compute dies and HBM stacks. So you need high-speed SerDes, you need die-to-die interfaces, you need custom HBM interfaces, you need much more optimized custom SRAM for much higher packing density, you need advanced packaging, and those are all things we do for our merchant business. So the reason we’re in the custom business is because of the expertise we’ve established from our merchant business.

Now when you build a chip, there’s an IP portion, which is more considered kind of front-end design, but right at the end of the process, you have to do layout, which is physical design. So some of the design service companies have an important role in the ecosystem, but they don’t have IP. And the reason they don’t have IP is because they don’t have a product business. That’s the real distinction. If we are engaged in a project, it’s because our networking IP is what the customer is looking to access and build into the XPU, versus when they’re partnering up with somebody in the design services side, it’s more of a relationship where you need someone to do the last part of the process, which is physical design.”

One example of specialized IP is Marvell’s SRAM with higher packing density, which enables 17x the bandwidth versus off-the-shelf IP according to the company. Marvell illustrates how this will be useful for both 2nm accelerators and CPUs:

And then also XPU attach is a big opportunity for Marvell. XPU attach is basically companion chips that can take background tasks over from the XPU, allowing the XPU to focus on key AI workloads. This is Marvell’s IR on this market:

“Think about XPU attach as an offload device. You’re trying to maximize your central compute device for its core function. This is no different than years back, where you started creating NICs to minimize I/O of the server CPU. Same idea essentially. CXL and NICs are the two large examples, but there’s a couple more design wins going into production. One would be storage accelerators. Whether it’s SSDs or HDDs, instead of having the core CPU or XPU address them, you’d rather do it on a dedicated device. So that’s a storage accelerator.

And another one is the security offload device. This is a unique product for Marvell, and comes from the Cavium acquisition. We’re one of the only companies building dedicated security accelerators. The idea is that I want to make sure my data when I put it into a cloud network is encrypted. You can certainly do encryption/decryption on a standard XPU or a standard CPU, but why would you? You’d much rather do it on a dedicated device. So that’s another example, which is again going to be a fairly large opportunity for us.”

Millions of XPUs are sold annually, and so if you can cross-sell $1,000-$2,000 worth of silicon that goes alongside these XPUs, this is obviously a big opportunity. Marvell is guiding for $3 billion of revenues in FY28 to come from XPU attach. Both CXL and custom NICs are $1 billion opportunities each.

And this opportunity continues to increase with agentic AI—if you need more CPUs, you need more NICs. And as agentic AI creates more KV cache data, this means also CXL will be needed. CXL is basically an interconnect that lets XPUs treat external DRAM as if it were directly attached to the processor. This way, excess KV cache can be moved from HBM to CXL-attached DRAM.

Share

Overall, we like Marvell’s strategy. The company is clearly building out a portfolio of some of the most cutting edge tech in both AI data center networking and custom silicon. Recent acquisitions like Celestial AI and Polariton are good examples of this strategy and seem once again like very smart strategic moves by Matt Murphy. So, we liked the risk-reward in this name a lot over the last twelve months, as the shares were trading in the low twenties on a forward PE basis:

However, after Jensen decided to pump the stock in Taiwan, we’ve now moved to a 70x forward PE valuation. For the bulls, we actually don’t think the shares are massively overvalued now—it’s likely that numbers are still too low given the emergence of agentic AI, and so we can move towards $10 of EPS already in calendar 2028. This gives a current valuation of around 30x 2028 numbers.

However, this valuation now warrants absolutely flawless execution. At the same time, we suspect that a lot of the speculative and fast retail money has now moved into the stock—the kind of crowd that buys based on a meme or a news headline. In our view, these two factors make the risk-reward now unattractive, with likely a substantial correction again at some stage. After shares rose 30-35% post Jensen’s comments, we decided to liquidate our position. We do love Jensen, somehow this guy always figures out a way to make us some nice profits.

Server CPUs Are A Hot Market Again

At Computex, Nvidia illustrated how the CPU has become important again in agentic AI workloads:

However, it’s also clear that the accelerator (GPU/XPU) will remain dominant as this is where all the reasoning happens. On the other hand, when the AI has to make a database call, do a web search, or execute a Python script, this happens on the CPU. So, the GPU writes the Python script, and then the CPU executes it.

Jensen highlighted at Computex how CPU performance becomes crucial in agentic workloads, from a transcript created by Claude:

“Huang argued that all CPUs until now were created for people — humans who live in a world counted by seconds, who rent CPU cores in the cloud at hourly rates. Agents are fundamentally different. They are impatient. They live in a world counted in nanoseconds. When an agent uses a tool or accesses a database, the response must come back as fast as possible. Every moment of waiting keeps the agent from proceeding to the next step. And because CPUs sit in the critical path next to extremely expensive GPU infrastructure that generates token revenue, these CPUs must be both high-performance and highly energy-efficient.

A narrated video elaborated on the Vera CPU’s technical architecture. The NVIDIA Olympus core at the heart of Vera is built for modern data center workloads including branch-heavy Python runtimes, tool calls, and sandbox code execution. The video described Vera achieving 40 percent lower peak memory latency versus x86. Memory-coherent NVLink chip-to-chip connects GPUs directly to the CPU and can also scale Vera up to multiple sockets. Vera delivers 1.8 times the agentic sandbox performance of x86 CPUs.

Returning to the stage, Huang presented benchmark results. He showed SQL running three times faster on Vera—calling it extraordinary since SQL is among the most difficult workloads to accelerate. He also presented real-time stream processing results for the New York Stock Exchange, where Vera CPU runs six times faster, crediting the bandwidth improvements, single-threaded instruction execution, and internal and external bandwidth improvements of the architecture.”

UBS notes that Nvidia will be doing more CPU revenues than AMD this year, which also includes standalone CPU sales for Nvidia:

“Maybe the most controversial item was new disclosure that “standalone” Vera CPU will be ~$20B in revenue this year and opens up a new $200B TAM. On the surface, this is a much larger number this year than we would have expected and is even bigger than AMD which we model to be ~$16B this year and has >50% share in the cloud segment. In speaking with the company, there are some clear use cases that are standalone chips (e.g. CPUs attached to non-NVDA accelerators, for example, META as part of its recently signed Vera CPU deal) but this also includes “standalone” systems that only have a CPU with no GPU. It is hard to say, but in these configurations, the standalone CPU chip might be ~30% of the total value with DRAM being another large component. So NVDA may only be doing ~$5-7B of CPU chip revenue this year - which would imply ~1-1.5MM units - which makes sense to us. Regardless, NVDA’s forward thinking on supply should especially come to bear in this agentic world where newcomers like ARM and even QCOM are competing for capacity and we read this as negative for the aspirations of these other competitors.”

Subscribe now

Similarly, Qualcomm—the leader in power efficient smartphone CPUs—is now looking to deploy their skills and take share in the data center CPU market:

“The second area where we’ll be entering the data center is the CPU. As you might know, we have a custom CPU that we deploy in handsets, where we are the performance leader. We also deployed that same CPU in PC. And so if you compare us to the x86 players, Intel or AMD, we think we have a very significant performance advantage. And we’re going to bring all of that to bear in a data center CPU solution. CPU use is changing in the data center, and it’s a very large market.”

Even ARM is looking to compete with their own customers and has now launched its ‘AGI CPU’:

“As AI is moving from human-based queries to continuous agent-driven workloads, this shift is expanding the role of the CPU. These Agentic workloads require CPUs to coordinate tasks, move data, manage memory, enforce security and orchestrate workaround accelerators. As Agentic AI scales, data centers will require more than 4x today’s CPU capacity, creating a data center CPU market opportunity of more than $100 billion by 2030. The Arm AGI CPU, which we launched at our Arm Everywhere event last quarter, is purpose-built for Agentic AI.

Our first production silicon product for the data center will deliver more than 2x the performance per rack compared with x86 platforms with the potential to reduce AI data center capital expenditure by up to $10 billion per gigawatt. Meta is our lead partner and co-developer, and is working with us on a multi-generation road map to support personal super intelligence for more than 3 billion users.

Customer response to the Arm AGI CPU has been very strong. We now have more than $2 billion of customer demand across fiscal 2027 and fiscal 2028. This is more than double what we stated at launch. We are on track towards our forecast of $15 billion as stated at our Arm Everywhere event. And soon, the data center will be Arm’s largest business.

For example, SAP will move their core database and business application workloads to Arm, starting with AWS Graviton and expanding to the Arm AGI CPU. This represents a significant strategic shift. Cloudflare will deploy Arm across its global network to support traffic management, security and AI inference closer to users. We have also secured design wins with key network infrastructure providers, including F5 and SK Telecom.

AI infrastructure needs CPUs and accelerators working together efficiently at scale. NVIDIA, Amazon and Google are already using Arm-based CPUs as head nodes along the accelerator-based systems. Cerebras, OpenAI, Rebellions and Positron are doing the same with the Arm AGI CPU. This momentum builds on our existing scale in the cloud. That scale is increasingly driven by Arm Neoverse CSS and Arm-based compute, which now represents about 50% market share with top hyperscalers.”

Historically, Intel and AMD had the x86 moat which meant that all server and PC software was written to run on x86 cores. However, due to smart container-based software such as Kubernetes, the large clouds can now deploy apps inside containers which abstract away the hardware underneath. This means that apps can run on both x86 and ARM-based CPUs, which is allowing hyperscalers to move workloads to their custom ARM-based silicon.

We’re only at the start of this trend as only Amazon has scaled its Graviton CPUs so far, but also Microsoft and Google are looking to replicate this strategy. Where x86 still has a stronghold is in the enterprise server market (i.e. on-premise data centers), however, cloud is the high growth market and over the long term, we suspect that most of the public cloud will move towards ARM cores.

Thus, cloud CPUs will become an extremely competitive market. Nvidia has already become massive in data center CPUs, in the blink of an eye. Another giant with Qualcomm is entering as well, which has strong skillsets in power efficient CPU design. Even Arm has cobbled together a number of interesting deals such as with Meta and Cloudflare. At the same time, hyperscalers such as Amazon, Microsoft, and Google will increasingly move cloud workloads to custom silicon.

Share

Despite increasing competition, we still see two key positives for AMD. While all hyperscalers are mainly looking to deploy a Nvidia + ASIC strategy in accelerators, AMD as a third player in the market can still make substantial money. Its GPUs with higher HBM capacity versus Nvidia are an advantage in inference, so we continue to think that AMD will be able to grab a smaller share in the GPU market, which will still be lucrative. Deals with OpenAI and Meta illustrate how the company has indeed secured share.

At the same time, momentum in the server CPU market will be tremendously strong in the coming years, while newcomers such as Qualcomm and Arm will only gradually gain share. Microsoft and Google are also still only at the start of their CPU ASIC journey. Thus, we see both AMD’s CPU and GPU businesses as booming in the coming years.

However, at 60x forward EPS, and 29x 2028 EPS, this strong story is now well reflected in valuation. So, we don’t see the risk-reward as being attractive at this stage. We could be interested again later in the year if there is a strong correction in AMD.

A negative for AMD is that capacity is extremely tight—TSMC isn’t able to meet all demand and this will also limit AMD to secure additional capacity for the boom in CPUs. This is AMD’s IR at the BoA conference:

“Supply is tight. 3-nanometer is tight. There’s a lot of other areas that are tight. I think we’re very well positioned with the relationships Lisa has personally with the executives in that space to get maybe more than our fair share of the incremental. But it is tight.”

It’s just something to keep in mind as it will limit AMD’s potential to surprise on the upside in the coming years.

Next, we’ll go through a variety of names that we see as winners from agentic AI—we’re currently halfway through the analysis.

Read more

SpaceX has three key businesses—Launch, Starlink, and AI. In the first two, the company currently enjoys a near monopoly-like positioning. For example, over the past few years, SpaceX has launched more than 80% of mass to orbit. The mission success rate for its rockets is over 99%. Additionally, Starlink is the only high-speed internet and mobile provider that’s available to consumers from space, making it available anywhere you are—on a boat, in a plane, hiking in the mountains, etc. This business is already printing cash while still growing at hyperfast rates. Finally, in AI, the company is one of the frontier AI model providers with Grok and soon Cursor (which will likely be fully acquired post IPO), with the xAI business setting records to bring up new data center capacity with its Colossus data centers.

Normally, a company would do one of these activities well—e.g. being a neocloud, training a frontier AI model, being a space launch provider, or running the world’s largest fleet of satellites. A key reason is that founder Elon Musk has been able to attract some of the best engineering talent to work at his companies over the last decade(s), as here engineers can work on the hardest set of problems—rockets, space, etc. This allows SpaceX to run a vertically integrated platform—designing and manufacturing its own rockets, satellites, engines, avionics, writing a lot of its own software, and even making the tools to build all these parts. This extreme level of vertical integration gives the company a platform to innovate rapidly and continuously drive the cost down.

Basically, the company has built a platform to enable new economies in space, with Starlink being the first of these. Due to SpaceX’s manufacturing and launch capabilities, the company was able to put by far the biggest satellite constellation in Earth’s orbit. Its more than 10,000 satellites currently make up close to 70% of all active satellites orbiting Earth. And this number can be increased up to 42,000. Deutsche Bank illustrates the “Starlink belt”:

Another growth angle is that SpaceX is currently transforming into a vertically integrated AI company as well. The goal is to manufacture its own GPUs, CPUs, and Memory in jointly operated fabs with Intel, the ‘Terafab’ project. It looks like Intel will bring in the IP and knowhow of 14A manufacturing, and then SpaceX and Tesla will bring in the capital and new engineering talent to scale up these fabs, with an aim to innovate rapidly on chip design and manufacturing. Also the tool makers such as ASML, Applied Materials, Lam Research, etc., as well as research houses in advanced semi technologies such as Imec can bring in substantial know-how. From the SpaceX prospectus:

“Our strategy for Terafab is to vertically integrate across the design of lithography masks, fabrication of logic and memory chips, and design of advanced packaging in a single closed-loop plant. Conducting all these activities end-to-end in a single facility enables rapid testing and iterations, allowing us to improve chip design and scale manufacturing faster. We expect that our speed and cost advantage from vertical integration will allow us to scale efficiently in AI chip manufacturing towards our long-term goal of producing one terawatt of compute each year.”

Then, one of the goals is to put this AI compute into space on satellite clusters, and power them with the cheap and abundant energy that the Sun provides:

“The Sun contains approximately 99.8% of the solar system’s energy and, as a result, we believe it is the only truly scalable solution to terrestrial energy constraints in the age of AI. Harnessing this energy in space is considerably more efficient than on land. Space-based solar arrays can generate more than five times the energy per unit area of terrestrial solar due to continuous illumination, lack of atmospheric interference, and optimal orientation.

In orbit, construction costs are replaced by launch costs and satellite production costs. We expect reusable launch systems and high flight cadence will significantly reduce the cost per kilogram to orbit, enabling more efficient deployment of compute payloads to orbit, and eventually approach the cost of fuel. We believe our advanced satellite manufacturing capabilities enable us to build AI compute satellites at scale and lower cost than competitors. Other terrestrial data center construction costs such as building the shell, MEP, and grid interconnection are not applicable in space. As a result, once Starship and our AI compute satellites are fully deployed at scale, we believe that the initial deployment costs of in-orbit compute in the aggregate will be less than construction costs of others’ terrestrial data centers.

In orbit, chips are expected to be powered by solar energy which is low cost and unlimited, and we expect to leverage radiative cooling architectures, which incur no operating costs compared to liquid or air cooling. Our integrated, space-based Starlink network architecture also enables more cost efficient routing of data between compute clusters and to end users on a global basis. We have already solved the hardest part in the development of AI compute satellites. AI compute satellites represent an evolution of spacecraft engineering already demonstrated at scale through Starlink’s connectivity satellites, and we believe development of AI compute satellites will be easier for us than for anyone else.

Connectivity satellites are primarily designed for communications, with substantial onboard equipment dedicated to phased-array antennas, radio systems, and data transmission. In contrast, AI compute satellites are optimized for high-performance computing. Key differences include significantly larger solar arrays to support higher power requirements, substantially larger radiators for thermal management, different electronics centered on AI accelerators rather than communications processors, and the removal of much of the communications hardware. Our V3 satellite platform already incorporates proprietary chips, providing a strong foundation for the ability to operate AI-focused electronics in space, and we expect to begin deploying our orbital AI compute satellites as early as 2028.

The primary remaining challenge is one of scale. For example, a deployment rate of approximately 10 gigawatts per year would require a materially lower manufacturing and launch cadence, which we believe would still enable a commercially attractive AI compute business with strong economic returns. While our long-term vision includes the ambition of deploying up to 100 gigawatts of power to orbit annually, which would require the deployment of thousands of launches per year, assuming 100 kilowatts of compute power per metric ton and Starship capacity to orbit of 100 metric tons, we believe we can be economically successful at significantly more modest volumes.

Specifically, we expect these satellites to leverage our already-designed V3 satellite platform. The core V3 satellite design is complete, and the AI compute satellites are expected to generate substantially more power than V3 satellites. This performance is expected to be achieved primarily through the use of significantly larger solar arrays. These satellites are targeted to generate approximately 100 kW of compute power per ton, which initially will require approximately five times the solar array output compared to V3 satellite designs. In general, the approach contemplates larger deployable solar arrays on each satellite, with no significant on-orbit assembly currently anticipated.

To ensure optimized thermal management and power generation, we will design each satellite’s solar arrays to face the sun for constant power while its housing radiator panels face cold deep space for radiative cooling. Space based compute also introduces orbital debris risk, which we already manage at constellation scale through our autonomous collision avoidance system across Starlink. To date, we have not experienced any failures of our autonomous collision avoidance system that have resulted in satellite loss.

We have built one of the largest satellite manufacturing operations in the world with standardized bus architectures, rapid iteration cycles, and automotive-style production lines, enabling us to evolve bus architecture and subsystem design with limited reliance on third-party suppliers. Our satellite constellation provides a direct, orbital data path that circumvents the bottlenecks of terrestrial communications networks. This architecture is particularly suitable to support high-speed connectivity for latency-sensitive workloads, which we believe are increasingly valued in certain consumer- and enterprise-facing applications.”

Each of Musk’s businesses scales based on “the algorithm”, which is basically a philosophy that every manufacturing process should be as simple as possible, and then be automated. These are the five steps to optimize manufacturing, crucially, the steps have to be completed in order:

1. Question Every Requirement

2. Delete Any Part or Process You Can

3. Simplify or Optimize

4. Accelerate Cycle Time

5. Automate

Let’s go through SpaceX’s current businesses in more detail..

Business 1: Rockets & Launch

The key to understand about the launch business is that the overarching goal is to drive down the cost of putting cargo into space. There are two ways to do this—increasing rocket reusability and building bigger rockets. The more times you can relaunch the same rocket, the lower the cost will be to put cargo into space. Similarly, if you can build a bigger rocket which can carry more cargo at once, again, the lower the cost per kilogram of cargo will be.

So, these are the two key factors which SpaceX is innovating on—building bigger rockets and increasing their reusability:

SpaceX describes the state of the space industry before its founding and how the company is transforming it:

“According to NASA, until the 2000s and the introduction of the Falcon 9 rocket by SpaceX, global commercial launch activity averaged 25 to 35 launches per year. As a result, the space industry remained a niche domain with limited ability to support large commercial markets or scaled space-based infrastructure. During this period, satellites—which comprised the majority of launch payload—were typically bespoke, expensive systems requiring significant non-recurring engineering that consisted of development cycles that were measured in decades.

Fundamental breakthroughs in high cadence, reliable, and affordable access to space—driven largely by SpaceX—have expanded space from a purely mission-driven activity to a fully industrialized and commercial sector capable of supporting and enabling industries far beyond traditional launch and satellites. SpaceX’s advancements reduced the cost of access to orbit from tens of thousands of dollars per kilogram to just a few thousand dollars per kilogram.

In December 2015, we achieved what many deemed impossible: landing a rocket launched to space back on Earth. By 2017, we were routinely recovering and reusing the Falcon 9 first-stage booster post-launch, delivering another step-function drop in space access costs via groundbreaking reusability. As of March 2026, our Falcon 9 rockets have demonstrated the ability to refly a first-stage 34 times. With the future deployment of Starship, which is designed to be the world’s first fully and rapidly reusable spacecraft, we aim to reduce the cost to reach orbit by 99% or more relative to the historical average launch cost, establishing the most affordable and scalable path to creating new opportunities in space.

Starship V3 is designed to deliver 100 metric tons to Earth’s orbit in a fully reusable configuration while enabling rapid turnaround times akin to commercial aviation. Future generations of Starship are being designed to double this payload capacity. To date, we have executed 11 Starship flight tests. We have also scheduled a 12th flight test, which will debut the next generation Starship vehicle and Super Heavy booster, powered by the next evolution of our Raptor engine and launching from a newly designed pad at Starbase. We expect Starship to commence payload delivery to orbit in the second half of 2026. We have achieved innovative milestones such as catching a booster using “chopstick” arms on the same tower it launched from. We expect this capability will facilitate rapid refurbishment and reuse, allowing for multiple launches per day at reduced costs.

Starship’s core design innovation is its full and rapid approach to reusability: both stages return to Earth for catch and rapid refurbishment. The Super Heavy booster returns to the launch site following stage separation and is caught mid-air by the launch tower’s mechanical arms, also known as “chopsticks,” to facilitate immediate inspection, refurbishment, and relaunch. The Starship upper stage, after orbital delivery or missions beyond, is designed to reenter protected by advanced heat shield tiles, execute a propulsive landing burn, and be similarly caught mid-air by the launch tower’s mechanical arms. We believe that Starship’s full and rapid reusability will enable sub-one hour reflights, causing a paradigm shift in launch cadence.”

Historically, launch costs were high—it used to cost about $8,000 to $64,000 to put one kilogram of payload into orbit. As a result, not much was being put into orbit before SpaceX’s founding. You can see that SpaceX’s Falcon 9 and subsequently Falcon Heavy really drove down the cost, which then allowed the company to put its massive Starlink constellation into space:

Basically, half of all payloads which have been put into space throughout history have now been carried out by SpaceX:

The initial goal of Starship is to reach a cost of $185 per kg to put a payload into space, although Elon has mentioned that $20 per kg is the real goal. Google estimates based on historical learning rates in other industries that SpaceX’s launch costs should fall to less than $200 per kg by 2035, a level at which it would become economical to put their TPUs onto satellites. This is Google’s analysis on the learning rate of Falcon 9 and Falcon Heavy, and how these have been driving down the launch cost per kg:

SpaceX sees its competitive advantage in the launch and space industries as vertical integration, allowing the company to innovate quickly to improve its products such as rockets:

“While conventional aerospace manufacturing relies heavily on fragmented and outsourced supply chains, we operate with extreme vertical integration. By designing and manufacturing a significant portion of our components in-house, we bypass many of the slow, bloated sourcing channels that structurally constrain the rest of the industry. For example, approximately 80% of Starship, SpaceX’s next-generation launch vehicle, is manufactured in-house. Our vertical integration allows us to achieve iterative cycles in weeks, compared to years for some legacy companies, enabling us to build newer, more technologically advanced products faster than many of our competitors.”

A good example here is the Raptor engine, with the latest version being shown on the right below, and which will power Starship:

Business 2: Starlink

Starlink is currently the high-growth cash cow for SpaceX. The service already has 10 million subscribers, up more than 100% from one year ago. The below map beautifully illustrates where those subs are located. Obviously there is still large growth potential as more people become aware of Starlink, more countries will allow the service in the future, and as SpaceX will launch a mobile service as well.

SpaceX provides further details on Starlink:

“We operate the world’s largest and most advanced space-based internet broadband service with median latency at approximately 25 milliseconds. We provide fiber-like download speeds—at a median of 225 Mbps during peak hours for residential users—and the technological capability to provide service everywhere on Earth, including the poles. This service quality is enabled by our vast network of approximately 9,600 Starlink broadband and mobile satellites in Low-Earth Orbit, which accounted for approximately 75% of all active maneuverable satellites in orbit as of March 2026.

We expect to commence deploying our next-generation V3 satellites, designed to offer one Tbps of downlink capacity per satellite, using Starship in the second half of 2026. We expect that a single Starship launch will be capable of deploying up to 60 V3 satellites to LEO, representing a potential twenty-fold increase in Starlink downlink capacity deployed relative to a Falcon 9 launch.

We charge our Starlink Subscribers a monthly subscription fee, which varies based on geographic market and download speed, plus typically a one-time upfront terminal cost. SpaceX is a critical partner to a wide array of enterprises. We offer Starlink to enterprise customers across industries including construction, agriculture, retail, telecom, hospitality, aviation, maritime, and land mobility. Starlink’s unique capabilities are well‑suited for deployments across field offices, remote worksites, research stations, drilling rigs, rural hospitals, aircraft, cruise ships, trains, and hotels. Our enterprise customers include companies such as United Airlines, Carnival, Maersk, and John Deere, among others. We also serve a broad fixed‑site customer base across industries such as retail and financial services that require high availability for critical operations.

Our enterprise contracts are based on a combination of subscriptions, data consumption, capacity, or other pricing models depending on each customer’s particular needs. Since 2023, no Starlink Enterprise customer having contributed more than $750,000 of annual revenue has voluntarily discontinued their service, demonstrating the strong performance and value of our offering.

We provide satellite-to-mobile connectivity. Our current capabilities under our “V1” constellation (consisting of approximately 650 V1 Mobile satellites in orbit) include light data, text messaging (SMS), and over-the-top voice services (e.g., WhatsApp and FaceTime). We are developing more comprehensive satellite-to-mobile services, including broadband data and IoT connectivity, which are expected to deliver resilient, infrastructure-independent connectivity worldwide and enable 5G connectivity. We have partnerships with approximately 30 MNOs on six continents, covering an area that is home to approximately 1.9 billion people. We charge MNOs either a fixed fee or a per-mobile user fee-based amount, which is typically passed through to the customer via the carrier as an “add-on” feature.

According to the Global Satellite Operators Association, terrestrial network infrastructure only covers approximately 20% of global land mass, resulting in significant unserved and underserved regions across both developed and developing economies. This terrestrial connectivity gap spans areas that are remote, difficult to build in, or economically impractical to serve—and also includes mobile “dead zones” within otherwise well-connected areas and in urban markets. According to the J.D. Power U.S. Wireless Network Quality Performance Study, U.S. wireless customers experienced service problems in approximately one out of every 11 mobile interactions, even in well-connected areas. As demand for ubiquitous, high-reliability connectivity continues to rise, terrestrial networks alone are increasingly unable to bridge the widening gap between user demand and available coverage.”

Starlink is an attractive partnership for mobile network operators (MNOs) as they can offer their customers connectivity everywhere as a premium service.

Starlink is also a business with a huge moat—you need enormous amounts of capital to launch all these satellites into space, you have to get SpaceX to agree to launch these satellites for you, and you also need huge capital to buy the spectrum to offer mobile services. SpaceX disclosed they’ll be spending close to $20 billion to purchase spectrum.

The raised capital from the IPO can also be used to advertise this business worldwide, as many target customers won’t even be aware it exists:

“We intend to grow the number of Starlink Subscribers by expanding our consumer distribution network across thousands of authorized retail stores globally and execute region-specific marketing campaigns to increase Starlink brand awareness. By clearly demonstrating Starlink’s superior speed, low-latency, affordability, and ease of installation—not only in rural, remote, and infrastructure-limited areas, but also in suburban and urban areas with wireline broadband options—we expect to drive meaningful subscriber and revenue growth.”

Business 3: Frontier AI & Colossus

SpaceX, due to its xAI acquisition, has quickly become the most impressive builder of AI capacity in the world:

“COLOSSUS and COLOSSUS II collectively provide approximately 1.0 gigawatt of compute power, with additional power capacity available for data center operations. We believe speed is a competitive advantage. In order to bring compute clusters online as fast as possible, we employ a vertically integrated, nimble approach to construction.

At COLOSSUS, we brought online the first cluster of approximately 100,000 H100 processors, approximately 130 megawatts of compute power, in just 122 days, repurposing the shell of an existing factory. At COLOSSUS II, we brought online the first cluster of approximately 110,000 GB200 processors, approximately 210 megawatts of compute power, even faster in 91 days. As an illustrative comparison, an industry benchmark to bring online a 100 megawatt greenfield data center is approximately two years.

Furthermore, in the case of COLOSSUS II, following the initial cluster, we brought online the second cluster of 110,000 GB300 processors and 220 megawatts of compute power in 64 days, demonstrating our ability to rapidly scale our facilities once built. We expect that once fully operational, the next phase of expansion at COLOSSUS II will bring online at least 220,000 additional GB300 processors and over 400 additional megawatts of compute power. We also demonstrated a significant improvement in cost efficiency, achieving data center construction costs for COLOSSUS II that are considerably lower than industry benchmarks on a per megawatt basis.”

These data centers serve a dual purpose. On the one hand, a Colossus data center can be used for inference and training of internal AI models such as Grok and Cursor’s Composer model. Note that SpaceX already has an agreement in place to acquire LLM coding specialist Cursor later this year. With Cursor obtaining access to xAI’s data centers, they quickly have been able to improve the capabilities of their Composer LLM via reinforcement learning. Based on the below benchmark test, Composer 2.5 has similar coding capabilities to Claude Opus 4.7 and ChatGPT 5.5, but at a much lower cost per task:

Artificial Analysis finds similar results:

“Composer 2.5 sits in third place on the Artificial Analysis Coding Agent Index, behind only higher-effort variants of Claude Opus 4.7 and ChatGPT-5.5 which cost ~10-60x more per task. This release puts Composer among the leading coding agent models, something that wasn’t clear for past releases.”

In addition, Grok is also an extremely capable model that is competing at the frontier of intelligence. Now that Elon is completely overhauling xAI—last year he was still leading the DOGE effort in the Trump administration, so xAI was more of a side project—xAI should become more competitive in the coming years.

However, three factors are crucial to compete at the frontier in AI—one is compute power, second is access to the best quality data sets, and third is engineering talent to innovate on algorithms and model performance. As the company has access to a massive training cluster with its Colossus data centers, xAI is obviously extremely well positioned when it comes to compute power. In coding data sets, Microsoft/OpenAI can have an advantage due to Microsoft’s ownership of Github, however, due to reinforcement learning, huge synthetic data sets can be created. Anthropic currently has the lead in coding which suggests that Microsoft’s ownership of Github might only be a limited competitive advantage.

Finally, access to engineering talent is crucial as well. Obviously, Google, OpenAI, Anthropic, Meta, as well as various AI startups, are all well positioned to attract AI talent. This is likely the reason that SpaceX made an agreement to acquire Cursor and is looking to acquire other AI startups as well—i.e. to get access to top talent. In addition, xAI recently formed a specialized recruitment unit made up of engineers themselves reporting directly to Elon Musk. These recruitment engineers are tasked solely with finding unconventional top talent and poaching them. There is a lot of math, physics and engineering talent around the world, and so looking outside the obvious circles where every AI lab is recruiting could give the company some unique capabilities. For example, Albert Einstein was also just an employee at a patent office who started writing physics papers in his spare time. Nikola Tesla never graduated university and gained his insights with self study. There are lots of examples throughout history of finding rare talent in unexpected places.

Given that xAI has the compute power and that they should have sufficient access to quality data sets, its success will likely hinge on factor number three, getting access to top talent that can train models that enterprises and consumers want to use. SpaceX disclosed that Grok 5 is currently being trained in the Colossus II data center.

We expect frontier AI to be a high-moat business with scale and networking advantages. SpaceX makes a similar conclusion and thinks that they will obtain a competitive advantage in hardware and energy:

“Only operators with access to massive amounts of power, very large GPU clusters and tightly integrated training infrastructure can train cutting-edge models, and these systems exhibit non-linear performance advantages that compound over time.

Companies that can structurally reduce energy, compute, networking, and deployment costs per token will be positioned to train faster, iterate more rapidly, and ultimately manufacture greater intelligence, scale models more rapidly, and deliver increasingly powerful and accessible AI solutions. This creates a self-reinforcing advantage in which lower token costs drive greater model quality and user adoption, reinforcing AI leadership. This is because lower cost per token enables more frequent model training, larger and more sophisticated models, longer chains of processing for reasoning and agentic workloads, and significantly higher inference volumes at economically viable prices.

Accordingly, for a given level of intelligence, we expect the long-term economics of AI companies to be driven by the ability to consistently deliver bleeding-edge compute at the lowest possible cost per token. Put simply, we view cost per token as a function of three primary inputs—the underlying AI model, the compute hardware, and energy, and we expect to have a competitive advantage in the latter two cost components. We believe we have a pathway over time that will significantly reduce compute hardware costs through continued vertical integration and development of proprietary chips, building on our experience designing custom silicon for our Starlink satellites. We also expect that the marginal cost of energy for our AI compute satellites will be minimal because our satellites are powered by solar arrays in space. By driving the energy component to minimal levels and pursuing improvements in compute hardware cost, we believe we can achieve a meaningfully lower overall cost per token in the future.

We believe that diversifying our long-term access to the supply of processors, including through our Terafab initiative with Tesla and Intel, will be a key driver in reducing the overall cost of compute hardware over time. By combining internally manufactured, lower cost chips with those we source from third-party suppliers, we expect the overall cost of our processors to decline.”

Our reading here is that SpaceX is banking on Terafab to give them a competitive advantage in hardware development, where a closed innovation loop with their fab will enable fast iteration to drive hardware innovation. Secondly, due to the company’s prowess in launch and satellite manufacturing, the vision is to have a competitive edge on the energy side as well, with the company’s AI compute being powered by continuous and low cost solar energy in space.

SpaceX’s current Colossus data centers can also be utilized as neoclouds with capacity being rented out to other AI providers. The company recently signed an extremely lucrative contract to rent out its Colossus I data center to Anthropic. Anthropic had massively underestimated demand for its Opus model, and so the AI lab has been scrambling to get more access to compute. Thanks to the Colossus I data center, Anthropic can now double each user’s rate limits, which had been severely restricted. The SpaceX-Anthropic deal also includes an agreement to explore building multiple gigawatts of AI compute in space.

SpaceX estimates that $7 trillion in AI capex will be needed through 2030 given that frontier AI remains infrastructure constrained:

“Massive demand for frontier AI models is accelerating the build-out of AI infrastructure at a pace and scale with few historical precedents. Meeting projected AI needs will require $7 trillion in global data center investment through 2030, with generative AI workloads expected to account for roughly 70% of total data-center power demand by the end of the decade. Each new generation of frontier models requires exponentially greater compute, following well-established scaling laws that link model performance to the volume and quality of training data, parameter count, and total compute expected. The rise of agentic AI and the potential emergence of artificial general intelligence are expected to further amplify inference workloads, driving a step-function increase in compute requirements and the corresponding data center capacity needed to support them. Frontier AI has become fundamentally infrastructure-constrained.”

SpaceX is clearly a highly innovative and vertically integrated company that’s working on a set of the hardest problems in engineering. The company dominates in the rockets & launch industry, Starlink is obviously a crown jewel, and the company is now competing at the leading edge in AI, which is already the fastest growth market of all time. We expect AI to grow into a massive industry over the coming decades, where the winners will likely become the biggest companies in the world in terms of market cap.

Below, we’ll go through SpaceX’s current businesses and opportunities to assign a valuation to each of these. At the end, we’ll sum up all these individual valuations in a sum-of-the-parts framework to see at what valuation we’d be interested to purchase the shares. IPOs are volatile. Facebook crashed more than 60% in the months following its IPO—which gave a tremendous buying opportunity for long term investors. So, it’s a good exercise to go through at which levels we’re ready to buy.

Future Opportunities, Financials, and Valuation

Read more

After a weak start for tech stocks this year driven by a panic in software and subsequently the Iran war, markets went into one of the most spectacular rallies on record as hostilities in the Middle East came to a halt and one of the key frontier AI labs, Anthropic, disclosed stratospheric revenue growth.

There have been a variety of AI bear theses over the past few years. Initially, lower quality accounts argued that “Nvidia is just cooking the numbers”. Media reports speculated that “Nvidia’s revenues are simply driven by circular financing”. The most credible bear argument was probably that there wasn’t a killer app in AI, or no substantial revenue.

However, the first killer app was already there as AI was driving revenue acceleration at the largest companies in the world, i.e. the internet giants, by creating better user feeds combined with better ad targeting:

In the meantime, also LLM pioneer OpenAI managed to build up sizeable revenues driven by accelerating growth rates:

Then, in the first months of this year, it became clear that Anthropic’s revenues were fully exploding on the back of its Opus model’s coding capabilities.

Our personal experience is that our coding productivity has gone up 10x, easily, and so this allows a single developer now to do the work of what previously would have required a small engineering team of even 10 people.

Naturally, this will create massive cost efficiencies at all companies that write code. As such, Meta announced this month that it will be cutting 10% of its workforce alongside cancelling another 6,000 open roles. Similarly, Block—led by Jack Dorsey—reduced its workforce by 40%. Dorsey is also looking to cut out middle management layers and have engineering teams report to him directly via an intermediary AI layer.

Therefore, we think that the setup for certain names in software is looking attractive—we’re at historically low valuations in the space, many investors are now spending zero time anymore researching the space, while software companies that can successfully automate workloads for their customers are likely looking at their biggest revenue opportunity ever. Finally, there is a huge potential for software companies to optimize their workforce and cost base (such as share based compensation) to drive steep EPS growth. While low quality software will get disrupted, key systems that form the backbone of enterprises are well positioned here in our view to capture this opportunity.

As the cost of writing production grade code has dropped massively, the demand for code is skyrocketing. Open roles for software engineers keep surging (chart below). Anything that couldn’t be automated before—either because the skills weren’t available or because the costs of writing the code didn’t justify it—now can quickly be written out by AI.

Overall, the current annual value that AI is creating is already in the hundreds of billions. We can see this by simply looking at Anthropic’s and OpenAI’s revenues, the revenue acceleration at the internet giants, the cost reductions in the required workforce, and then GPU revenues from running inference for open source models.

We’re in the very early phases of the AI revolution, similar to the nineties when it comes to the internet revolution. After the launch of the Netscape browser in 1994, we saw multiple waves of innovation in subsequent decades with massive new TAMs being created—search, social media, SaaS, the public cloud, streaming, etc. Killer apps such as TikTok even emerged only recently. It’s likely that with AI, in the coming decades we will see multiple waves of innovation that will continue to drive up demand. Writing code, creating better social media feeds and ad targeting, helping with research, tracking your calories, etc. are only the first applications of AI. There’s huge potential for large parts of white collar jobs to be automated or augmented with AI, as well as the emergence of physical AI in the form of robotaxis, humanoid robotics, and industrial automation.

When it comes to AI capacity currently being built, we see the risk of current capacity expansions being wasted as low as the environment will remain tight at least into 2027.

This is neocloud Nebius commenting on GPU demand on the recent call:

“Our pipeline generation in the first quarter grew 3.5x over the fourth quarter, and this is a record for us. And the demand is broadening across industries. […] The vast majority of capacity coming online over the next several quarters to 12 months is already under contract. We are typically seeing 4 or more customers competing for every GPU we bring online. We have significant expansion planned for 2027, including Vera Rubin’s, and we’ll start selling that capacity as we move into the second half of this year.”

Physical AI is still too far out to really make an impact in the coming years, although this can become a large semis market in the 2030s. Say, ten years from now, 100 million humanoid robots are sold per annum (vs a smartphone market which is currently at 1.25 billion units), this would translate into another semi TAM of $600 billion. For comparison, Nvidia will be doing $372 billion in revenues this year.

Assuming the semi designers will make 60% gross margins, this means that foundries can capture a $240 billion market in physical AI. For comparison, TSMC will ‘only’ be doing $136 billion in revenues this year. And then there’s further potential in physical AI with robotaxis, industrial robotics, etc. We remain long term semi bulls and continue to hold the best quality names in the space.

Small Cap Semis Fever

However, we’re also seeing parts of the market that are starting to look frothy, especially in small cap semis that we would categorize as speculative stocks.

Take Aehr for example, a credible company with a strong growth outlook. However, with a market cap of $3 billion, the company is now trading on 60x its current backlog of $50 million. For comparison, ASML—which is probably more attractive in our view—is trading on 10 times its backlog.

We see a similar story in Wolfspeed—an interesting company with a state-of-the-art SiC fab in New York. However, the market cap has gone up 4x since the start of April while we didn’t get any news of meaningful orders on the last call. Remember, due to low utilization rates, their SiC fab is still generating negative gross margins and is burning around $100 million in cash per quarter, while the company remains in a net debt position. Wolfspeed really needs orders to come in, in a SiC market that remains oversupplied.

We’ve seen a number of bullish reports connecting Wolfspeed to the solid state transformer (SST) market where SiC power semis will be heavily used. However, we’re unsure whether this market will be large enough to really turn this fab FCF positive. Infineon estimated the SST power semi market to be over $1 billion in 2030. If we bullishly assume Wolfspeed can capture 25% of this market, it would mean $250 million in revenues in 2030. Wolfspeed needs a lot more orders than that, and faster. In the meantime, power semi king Infineon is already shipping silicon for SSTs and continues to pile up orders, with SolarEdge announcing they will be sourcing their SST SiC from Infineon.

While it’s good news that Wolfspeed has developed a 300mm SiC wafer, likely at the request of TSMC for use in AI applications, SiC wafer manufacturing is a pretty commoditized market with tons of competition. If TSMC needs 300mm SiC wafers, they’ve likely asked a number of players to develop these. While it can give Wolfspeed a new revenue source, SiC wafer manufacturing is too much of a commoditized and low margin business for us to really get enthused here as well.

In our view, Wolfspeed’s share price is starting to get disconnected from the fundamentals as well and we see the risk-reward now as much more balanced compared to when we first initiated our position in December. Therefore, we’re locking in profits here and will allocate proceeds to stocks with better risk-reward.

Memory As The New Hottest Industry

In the meantime, we’re also starting to see classic signs of a bubble emerging in other areas of semis. Bloomberg notes that South-Koreans are now massively opening new trading accounts and borrowing money to invest in the local stock market, with the KOSPI index being heavily dominated by the two memory giants, Samsung Electronics and SK Hynix:

“In South Korea’s $4.6 trillion stock market, signs of euphoria are popping up everywhere. Enthralled by a 200% surge over the past year that has far outpaced every other market on earth, locals are borrowing record sums to amplify their bets on stocks. Trading volumes have soared to all-time highs, while daily price swings of 5% or more have become more common — making the benchmark Kospi index the most volatile major stock gauge worldwide.

The sense of FOMO spreading through offices, lunchrooms and family gatherings across South Korea has gotten so intense that investors are increasingly buying stocks for their kids, too: Data compiled by Toss Securities show a near 10-fold surge in new account openings for under-18s in the first quarter versus a year earlier.

“The mood in the retail community is very hot, nearly maniacal,” said Jang Eunjung, 37, a Seoul-based video producer whose YouTube show on stock investing grew from a tiny audience to more than 1.3 million subscribers as the market soared. “Will we ever see such a vertical rally again?””

Another classic sign of a bubble is the launch of new investment funds purely focused on the particular thematic, and yes, we’re also seeing that now when it comes to memory. From CNBC:

“‘Biggest bottleneck in the AI buildup’ fuels DRAM ETF to record. The Roundhill Memory ETF (DRAM) just hit $9.8 billion in assets under management in 43 days— the fastest pace ever for an exchange-traded fund.”

The DRAM ETF is purely focused on all memory stocks, i.e. the three DRAM/HBM players and the NAND players.

As readers will know, we’ve been holding the DRAM/HBM names over the past years as this is a fairly attractive three player market while AI is extremely HBM and DRAM intensive, giving a clear growth angle.

NAND on the other hand remains much less attractive as the industry is much more commoditized, and it’s also the least crucial type of memory for AI. Basically, this is where you store the KV cache when your DRAM can’t hold it anymore. HBM is what really powers AI workloads, and then DRAM is where you store data that you need fast access to. NAND on the other hand is much slower.

We’ll go in detail through our updated outlook for the DRAM/HBM space, with our thoughts on whether to continue holding these names or whether we’re locking in profits here.

Memory Outlook - SK Hynix, Micron & Samsung

1. The Impact of Long Term Supply Agreements

Read more

Snowflake was extremely hot. Whereas the initial range was set at $75-$85, the IPO was finally priced at $120. Shares opened at $245 and had wild swings on the day. However, as the hot money exited, shares traded lower in the following week. In this case, investors who bought the shares at the open made a so-so call—as this was roughly the average price level over the following two months:

Astera Labs was another hot IPO. The initial range was $27-$30, the IPO priced at $36, and the shares opened at $52. In the following days, shares rallied, after which we saw share price contraction during the following weeks. However, shares continued to trade above price levels on the day of the IPO. In this case, investors who bought when the shares opened made a great call, with shares still trading 38% higher two months later:

Finally, we have Arm. This was a weaker IPO than the other two, as Arm only priced at the top of its range at $51. The shares opened at $56 and rallied strongly on the first day but during the second trading day, a sell off started. Although investors who bought at the open still didn’t do that badly, losing only 5% in market value over the next two months. Again, buying at the open wasn’t a great call, like in the Snowflake IPO, although the damage was limited.

Next, we will go through our thoughts on the Cerebras IPO and share price levels we would be interested in.

Read more

“A 5-year backlog on grid transformers just killed half of America's 2026 AI data centers. Sightline Climate tracked 12 GW of 2026 US data center capacity announced across 140 projects. Only 5 GW is actually under construction. 11 GW sits in the "announced" stage with no physical progress despite typical build times of 12-18 months. 25% of those projects haven't disclosed a power strategy at all.

Nvidia is shipping. The gating constraint is high-voltage transformers, switchgear, and grid-tie batteries. Pre-2020 lead time on a high-power transformer was 24-30 months. Today it stretches to 5 years. Electrical equipment is under 10% of total data center cost and 100% of the bottleneck.

The winners under this regime are whoever locked in power purchase agreements and electrical equipment orders 3-4 years ago, before anyone was modeling hundreds of megawatts of inference load. Everyone else is waiting in line behind them. Second order is uglier. Hyperscalers buying $50B of GPUs that sit unpowered depreciate against Nvidia's annual cadence while paying carrying costs on empty data center shells. Every quarter dark is a quarter of compounding waste.”

There are three bottlenecks mentioned—high-voltage transformers, switchgear, and grid-tie batteries. Let’s focus on the HV transformer first, that’s the big one. You need a massive HV transformer when you connect the data center’s power supply to the grid, as transmission lines carry electricity at high voltages over long distances:

So, when you don’t have a HV transformer, the solution is to move behind the meter—i.e., a data center that generates its own power on-site. Local generation can produce electricity at a medium voltage (MV), or even low voltage (LV), avoiding the need for a massive HV transformer.

Fortunately, there are plenty of ways for a data center to generate power on-site. A recent HSBC report on AI bottlenecks discusses the trend to ‘bring your own power’ (BYOP):

“Grid equipment lead times are rising with some shortages in medium voltage equipment, but the key bottleneck is in high voltage substations, where lead times are 3-5 years. This shortage, along with long wait times for grid connection, has driven the rise of the BYOP (bring your own power) model of onsite power generation. Whilst many datacentre developers scrambling for electricity intend to use onsite power for a few years until the grid infrastructure can catch up, a few plan to bypass the grid indefinitely. Demand is rising for off-grid “energy islands” – cutting out the grid, under a colocation model. One project opting to forgo a grid connection is a Meta Platforms datacentre campus near Columbus, Ohio. In July 2025, regulators approved a plan by pipeline company Williams to build onsite natural gas power. Also Oklahoma state lawmakers passed a bill earlier in 2025 to let companies build their own power to take advantage of abundant gas supply.”

The problem with power from natural gas however is that now also gas turbines have massive backlogs. Leading turbine manufacturer GE Vernova for example still has some capacity in ‘29 and ‘30:

“We’re directionally at about 3 years lead time today, we’re sitting in the spring of ‘26 and we do still have capacity in both ‘29 and ‘30. What’s happened in the first quarter is we sold a lot of 2030 slots. We still have about 10 gigawatts remaining cumulatively in ‘29 and ‘30 together. It’s important to contextualize that the 100 gigawatts we have on contract today is with almost 90 distinct customers in 24 different countries. So there’s a need for incremental electrons for many different applications in many different countries, which has us continuing to work hard to figure out how in a very capital-efficient way we meet this moment and serve this market.”

Note the focus on capital efficiency, i.e. GE doesn’t look to be planning to add much manufacturing capacity, further increasing gas turbine shortages. Looking at the HSBC analysis, Siemens and Mitsubishi are planning to add capacity however already as from this year, and overall gas turbine production capacity will rise from 70GW in ‘24 to over 90GW in ‘29:

GE is probably reluctant to add capacity in the case that demand pulls back. This is what happened in the early 2000s. There was a boom at the time in orders following US power sector deregulation, after which HSBC notes that ‘key suppliers aggressively expanded capacity but then suffered from low utilisation as orders normalised’:

Besides the long lead times to actually get the gas turbines to a data center, another problem is that gas turbines are loud and cause pollution. Two factors which are especially important for infrastructure near residential areas. Due to these problems, other solutions such as fuel cells are seeing huge order growth. Fuel cells have numerous advantages—they’re clean, don’t make noise, and can output electricity at low voltages directly. This means you don’t need a HV transformer, or even a medium voltage transformer (which steps down MV to LV).

Bloom Energy is the big winner in fuel cells so far and recently saw a massive order from Oracle for up to 2.85GW of power. From Barclays:

“Oracle’s Project Jupiter provides insight into the shift by hyperscalers. Into the print, our channel checks did suggest that there were some turbine orders being canceled/slots being sold with customers switching to Bloom, but we weren’t sure if it was a one-off or an indication of something broader. On the call, BE confirmed that ORCL’s Project Jupiter pivoted from a conventional plan (gas turbines plus diesel backup) to a 100% Bloom, fully islanded microgrid - no grid interconnect, no diesel, no turbines, no batteries. Management emphasized that BE was selected not only due to its speed to power, but also because it is clean and community acceptance (air quality, water usage, noise pollution).

While investors tend to compare the capex cost of BE’s option to CCGT alternatives, this update would underscore that this may be an “apples to oranges” comparison like management stated on the call, as there are attributes and advantages to using BE’s fuel cells that may not be as easily quantifiable. This is especially the case when adding in the benefits of its native 800V DC fuel cells and considering the NIMBY attitude that we are starting to see regarding data centers. As more than half of current data center backlog is now comprised of non-ORCL hyperscalers, neo-clouds, and colo providers, this would also suggest this is not a one-off situation, but more reflective of a broader trend. We would expect future ORCL projects to also find their way into the backlog, solidifying the re-rate higher we have seen in recent weeks.”

Bloom’s Solid Oxide Fuel Cell (SOFC) can be supplied with either natural gas or hydrogen, and the technology is proven—the company already has 1.5GW of capacity installed globally. Another benefit is that you don’t need backup batteries to power your data center, or dirty diesel engines. While Chinese battery manufacturers are stepping up capacity—and we didn’t come across shortages in this area—batteries aren’t necessary to power AI data centers anyways.

Additionally, Bloom has been talking on their call how they can easily scale capacity to meet demand. Recent highlights from the call:

“The marketplace is recognizing and embracing our proposition of clean, reliable on-site power that is community-friendly and deployed at the speed of AI. Bloom is rapidly becoming the standard and go-to choice for on-site power. Last night, Oracle announced a new power paradigm for Project Jupiter, a multi-gigawatt AI factory to be built in New Mexico. We are thrilled to partner with Oracle and applaud them for their visionary leadership. This up to 2.85 gigawatt power block will replace Project Jupiter’s previously planned gas turbines and backup diesel generators with Bloom Energy Servers. It will be 100% Bloom. When completed, it will be one of the largest islanded microgrid power facilities in the world.

At a time where every quarter of delay translates into hundreds of millions in foregone AI revenue and loss of competitive advantage, speed of powered infrastructure development is the difference between leading and following. It’s not going to be a one-off project. Where Oracle is going is where the broader market is headed. More than half of our current data center backlog comes from other hyperscalers, neoclouds and colocation providers. Just like the Oracle Jupiter project, these microgrid installations will use no grid, no dirty diesel generators for backup, no battery banks for load following, no engines, no turbines, just Bloom.

Time to power has gone from a procurement consideration to an existential necessity. The companies driving the AI transformation are racing against each other and bumping into bottlenecks such as permissions, permits and community acceptance. The winner will be the one who can grow and deploy faster. The legacy power industry’s model is long-cycle times and capital-heavy capacity additions. Our model is different at every layer. We innovate and improve continuously—in our technology, how we manufacture, in our capital intensity, how we deploy. That is what allows us to deliver double-digit cost reductions year after year, expand capacity with materially less capital than industrial-era players, and meet our customers’ schedule. We are adding lines and expanding capacity to meet growing demand, and innovating to reduce cycle times, space needs and costs.

Our competitors’ supply to current orders arrives only in 2029 or later, our supply arrives this year or the next, whenever the customer is ready. Based on the demand profile, we have now shifted to adding capacity continuously, hundreds of megawatts a quarter as opposed to lumpy one-off additions. The traditional power industry has spent the past 2 years celebrating its backlog that is 4 and 5 years out. Our ability to expand capacity is our competitive advantage.

Our current manufacturing footprint will allow us to deliver 5 gigawatts of product annually. We will expand that capacity and meet the delivery dates needed by our customers. Today, we are not capacity-constrained. The pace of our revenue growth is decided by how fast our customers can build their greenfield sites, not how fast we can power them. Going beyond the 5 gigawatt capacity, our supply chain and manufacturing strategy allows us to build that capacity significantly faster than any other option in the market using our copy exact model.

We have a 100% attach rate between our product sales and our service. Even with the data center opportunities, on average, it’s 10 to 15 years of servicing, and so it’s a tremendous source of annuity revenue that we see.

Communities care deeply about what kind of infrastructure shows up next door. Bloom preserves local air quality. We do not combust and pollute the air like conventional technologies. We use minimal water at startup and none during normal operations. We are quiet, compact and efficient with land use. We integrate well with environments rather than disrupt them and become an eyesore. As permits and permissions become the gating factor for AI infrastructure, community acceptance matters increasingly. Our solution does not raise the monthly electricity bill for community residents.”

Goldman notes that fuel cells are actually cheaper than gas turbines when natural gas prices go higher:

“Overall, fuel cells currently have a higher LCOE (Levelized Cost of Energy) than gas turbines - roughly 40% higher than that of simple-cycle gas turbines (at natural gas price of 4$/mmbtu). The cost gap is mainly driven by the high cost of fuel cell systems and limited deployment scale. However, as natural ‐ gas prices increase, the cost gap narrows in favor of fuel cells, because fuel cells consume less gas per unit of electricity. At spot natural‐gas prices of $13.3/MMBtu, our LCOE analysis indicates that fuel cell generation is broadly comparable to, or around 7% lower than, gas turbine based generation. Finally, as manufacturing expands, we expect the LCOE of fuel cells to improve substantially, gradually approaching the competitiveness of advanced gas turbine technologies in the longer term.”

While in the US, natural gas is dirt cheap, in Europe and Asia spot prices are much higher. This will make the economics of fuel cells especially attractive for European and Asian AI data centers.

So, Bloom’s fuel cells solve the bottleneck of transformers, which has gotten a lot of attention. The other bottleneck mentioned is the switchgear. However, listening to the conference call of European industrial giant ABB, additional capacity is now coming online and switchgear shouldn’t be the limiting factor in the AI data center buildout:

“The capacity expansions are now coming online, that money was spent 2 years ago and last year. I don’t think we are the limiting factor when we’re talking about build-out. We know the equipment like gas turbines, large power transformers, is more of a limiting factor than we are on switchgear and the component level. We’re also expanding capacity by appointing new partners, what we call the OEM partners, and the ones who do the system integration on ABB’s behalf.”

While Bloom is bullish on adding capacity, they have a few flagship sites when it comes to manufacturing and assembly. So, when those get to full capacity, it is possible that the company won’t be able to meet demand until new sites are added.

Therefore, besides Bloom, there should be room for other players with other competitive technologies to start capturing orders as well. Next, we will go through two small caps with two different technologies that are also well positioned to start capturing orders to power AI data centers. We see the risk-reward as compelling in one of them—with minimal downside and substantial possible upside. Additionally, we’ll also analyze the valuation for Bloom and the company’s outlook.

Read more

Cerebras accelerators have a radically different design than current XPUs. Instead of dicing a 12-inch silicon wafer into dozens of individual chips, Cerebras leaves the wafer intact to create a single, gargantuan processor. The current iteration, the Wafer-Scale Engine 3, is fabricated by TSMC on a 5nm process featuring 4 trillion transistors distributed across 900,000 AI compute cores.

Its wafer scale accelerator is 57 times larger than a single Nvidia B200 die:

This monolithic design addresses the primary bottleneck in LLM workloads: the memory wall. By keeping 44 GB of SRAM directly on the wafer, Cerebras’ WSE delivers a massive 21 petabytes per second of memory bandwidth, resulting into much faster inference (up to 15x according to the company compared to Nvidia B200):

For a GPU to calculate the next token, it needs to multiply the current state by the model’s weights. Because of how autoregressive decoding works, you cannot process the next token until the current token is finished. So, to generate one single token, the GPU must read the entire model from its VRAM (HBM) into its compute cores. Because the GPU is doing very little math per byte of data moved, this means decoding is memory-bandwidth bound. Cerebras’ design minimizes this bottleneck by storing the model weights in SRAM, which is located right next to the compute cores.

Nvidia is aware of the massive bandwidth advantage of SRAM, and so recently announced a unique inference solution that combines the best of both worlds—Rubin for the prefill and the attention phase of decoding (which are compute heavy) and Groq for the final token generation phase of decoding (which is bandwidth intensive). In the first phase, handled by Rubin, the AI figures out what the prompt and its context (stored in the KV cache) are about. While in the final phase, handled by Groq, the model applies its learned knowledge (i.e. its parameters or weightings) to the specific query:

A key drawback of SRAM-based inference systems like Groq and Cerebras is that SRAM doesn’t have much memory capacity. So, you have to connect 45 Cerebras CS-3 wafers together to hold a 1 trillion parameter model. Cerebras can technically cluster up to 2,048 wafers together, so even a state-of-the-art 10 trillion parameter model fits well within this theoretical maximum.

The KV cache stores all the context from the user queries and so takes up a lot of memory. This is why you want to network together an SRAM-based system with a classic XPU server which has loads of HBM, DRAM, and even SSD. As context windows get longer and longer with AI agents doing multi-turn reasoning, the KV cache explodes in size. It quickly exceeds what fits in HBM, and even DRAM. As having to recompute the KV cache is compute expensive, this is the reason why Nvidia recently added additional SSD to their servers. This is also the reason why pure SSD stocks rallied strongly last year, however, this is a commoditized and competitive industry, so make sure you don’t hold these bags for too long.

The second drawback of a Cerebras inference system is the high price tag. While it can hold a 1 trillion parameter model and more, networking a 45 Cerebras CS-3 cluster together can easily cost $100 million or more. For comparison, a single Nvidia GB200 rack can hold more than six of these models in its HBM at a hardware cost of only $3.5 million. The advantage of Cerebras is low latency, whereas the advantage of Nvidia is loads of memory. A single Nvidia rack has so much memory it can handle queries from tons of users at the same time, perfect for a chatbot app or a LLM API. Cerebras is more like a Formula 1 car that provides ultra low latency inference for a user willing to pay up for it, e.g. Citadel hedge fund or Goldman’s trading desk.

So, Cerebras has two key problems—firstly, a lack of context memory, and secondly, an astronomic price tag.

The first problem, a lack of context memory, can be solved when you combine the Cerebras system in a disaggregated inference solution with a traditional XPU server which does have the memory. As Nvidia has already shown, you get the best of both worlds—loads of memory and fast decode. This is exactly what Cerebras is now doing with their new Amazon AWS partnership:

“We signed a binding term sheet with Amazon Web Services for AWS to become the first hyperscaler to deploy Cerebras systems in its data centers. Deployment in AWS data centers will require us to meet strict standards for performance, scale, and reliability. Pursuant to the term sheet, we will create a co-designed, disaggregated inference-serving solution that will integrate AWS Trainium3 chips with Cerebras CS-3 systems, connected via high-bandwidth networking, to partition inference workloads across Trainium3 and CS-3. Each system will perform the type of computation at which it most excels. The approach is expected to deliver 5 times more token throughput in the same hardware footprint, at up to 15 times faster speeds compared to leading GPU-based solutions as benchmarked on leading open-source models.

Inference disaggregation is a technique that separates AI inference into two stages: prompt processing, or “prefill,” and output generation, or “decode.” These two stages have different computational characteristics. Prefill is natively parallel and requires very little memory bandwidth. Decode, on the other hand, is inherently serial and memory bandwidth intensive. Decode is typically the bottleneck. It dominates total inference time, and defines the speed of the user experience. Disaggregated inference would allow Cerebras to operate alongside other architectures, serving as the high-performance engine for decode while other systems handle prefill.”

AWS will make this hyperfast inference available via Amazon Bedrock for a variety of LLMs, including open-source models.

The second problem with Cerebras is the very high price tag. This also becomes more manageable in a disaggregated inference solution as the throughput, i.e. the number of jobs and users that the Cerebras servers can handle, gets a large boost as Cerebras’ wafers can now purely focus on serial token generation. If those users are willing to pay a substantial premium for these hyperfast tokens, the investment starts to make sense. This is what Jensen said on the demand for hyperfast inference at GTC:

“At the most valuable tier, we’re now going to increase performance by 35x, and I would add Groq to maybe 25% of my total data center. With the rest of my data center all Vera Rubin.”

So, Jensen has probably provided the biggest bull argument for the TAM Cerebras will be competing in. The second bull argument is that Cerebras looks to be a strong player in this TAM, as the company recently disclosed another interesting partnership…

The OpenAI Deal

“We signed the MRA with OpenAI on December 24, 2025. On January 23, 2026, we began delivering capacity to OpenAI, and on February 12, 2026, OpenAI’s Codex-Spark model, powered by Cerebras infrastructure, was made available to the public. Spark is OpenAI’s model designed for real-time coding. Using Cerebras, OpenAI’s customers can translate ideas into working software in seconds, enabling developers to create software at the speed of thought. OpenAI has committed to purchase 750MW of Cerebras inference compute capacity over the next three years.”

This 750MW deal is worth $20 billion, and so spread over three years, this would be around $6.7 billion in annual revenue for Cerebras. We can work out that OpenAI is basically paying here roughly $27 billion per GW of Cerebras capacity. There is an option to expand the agreement to 2GW by 2030, meaning that an additional 1.25GW can be provided by Cerebras to OpenAI in ‘29 or ‘30. This would mean another $33 billion potential order from OpenAI. OpenAI and Cerebras also agreed to co-design future OpenAI models for future Cerebras hardware.

OpenAI won’t purchase the equipment. Cerebras will provide the capacity in a managed compute-as-a-service type agreement. As Cerebras has to order the required wafers from TSMC, OpenAI also gave Cerebras a $1 billion working capital loan. If Cerebras can deliver the compute successfully, no interest has to be paid. Otherwise the interest rate is 6% on the loan.

Cerebras also issued a warrant to OpenAI for 33.5 million Cerebras shares, with vesting directly tied to equipment purchase milestones. So, if OpenAI orders the full 2GW of compute by 2030, they’ll be given the full 33.5 million Cerebras shares. As the exercise price per share is basically $0, OpenAI has a financial incentive to order more Cerebras compute. The dilution, if this warrant gets fully exercised, will likely be around 10%.

Bears will point to the circular financing nature of the agreement, however, we argue that for a startup it’s not easy at all to get a hyperscaler or leading AI lab to run inference on your silicon. A large AI player will have to make considerable investment to deploy models on your hardware—compilers will have to be written, the equipment has to be tested and verified, software layers have to be coded to automatically deploy models on a variety of underlying hardware, etc. All of this costs considerable engineering hours.

In the meantime, AI is progressing at breakneck speed. While AI labs will want to optimize their hardware, their key focus is on not falling behind in this race. The value of their business will be determined by capturing sustainable market share and revenues. For example, by winning enterprise AI workloads. In order to do this, you have to train differentiated models, like Claude Code. Improving margins by optimizing hardware will be a secondary objective.

Also on an accounting basis, the value of these warrants will be booked as contra revenues in the coming years. From a US GAAP perspective, these warrants are treated as a sales discount, which lowers the net revenue Cerebras ultimately recognizes from OpenAI.

Business Background & Founder’s Letter

Cerebras does not sell standalone silicon—it packages its wafer into a full system which integrates power delivery, liquid cooling and networking; occupying one-third of the space of a standard data center rack. Revenue so far has primarily been driven by sovereign and university supercomputers in Abu Dhabi.

To make it easy to deploy AI models on Cerebras wafers, the company provides a software platform which includes a Pytorch compiler, an inference stack with APIs, and a cluster management system providing scheduling, telemetry, and health monitoring at scale. This is even integrated with the most popular IDE, VS Code, so you can write an AI model with Pytorch in VS Code and then compile it to run on Cerebras.

The management team, led by CEO Andrew Feldman, brings a proven track record. Cerebras’ founders previously established SeaMicro—a pioneer in high-density and energy-efficient microservers—which was acquired by AMD for $334 million in 2012. Feldman subsequently served two years as VP at AMD, after which he co-founded Cerebras. More than 50 engineers from SeaMicro joined Cerebras, highlighting that Feldman can create a cohesive engineering culture. CTO and co-founder Sean Lie has a longer background at AMD and was also an engineer at SeaMicro:

“Sean Lie is one of our co-founders and has served in various roles since 2016, including most recently as our Chief Technology Officer since April 2022. From April 2012 to June 2015, Mr. Lie served as Chief Architect, Data Center Server Solutions at AMD. From March 2008 to March 2012, Mr. Lie served as Lead Hardware Architect of the IO virtualization fabric ASIC at SeaMicro. From July 2004 to February 2008, Mr. Lie worked as a microprocessor architect at AMD in the advanced architecture team. Mr. Lie holds an M.Eng. and a B.S. in Electrical Engineering and Computer Science from the Massachusetts Institute of Technology.”

It is notoriously difficult to get high yields with wafer-scale silicon—a single defect can ruin an entire wafer. Cerebras circumvents this by outfitting the hardware with millions of redundant compute and memory cells, bypassing defective areas dynamically at the manufacturing level. While this solves the yield problem, the resulting systems are extremely capital intensive. A single physical node is estimated to cost between $2-3 million and can draw up to 27kW of power.

From the founder’s letter, this gives useful background on how they’re thinking about the business and industry:

“The first bet: that existing general-purpose processors would not be sufficient, and that what has always been true throughout the history of compute would also be true for AI – that transformative compute workloads require purpose-built silicon. This is what PCs did for x86, graphics did for GPUs, and mobile did for ARM. The second bet: that modifying existing compute architectures would not realize AI’s potential. We would need to build a new computer architecture from first principles, optimized in every way for AI.

At a computational level, AI is a communication-bound problem. The faster compute communicates with memory, and the faster compute communicates with other compute, the faster and smarter the AI, and the better the user experience. The enemy of speed is communication latency. And since communication is thousands of times faster on-chip than across chips, the best way to reduce latency is to keep communication on-chip.

Our answer: build the largest commercial chip in the history of the computer industry. We used the entire wafer for one chip: a technique called wafer-scale integration. Wafer-scale integration allowed us to bring together quantities of compute and memory never before assembled on a single commercial chip and deliver AI at previously unimaginable speeds.

Nobody knew how to yield a chip 58 times larger than the leading GPU. Nobody knew how to deliver power to a chip the size of a dinner plate without melting the motherboard. Nobody knew how to package such a big chip without cracking it. Nobody knew how to cool a chip of this size, with air or water, without the coolant getting warm before it reached the other side.

We delivered our first systems in 2020 and the second generation in 2022. We had built something extraordinary, but the market wasn’t ready. A few visionary customers in supercompute and life sciences saw the potential, but to most of the world, the benefits were not immediately apparent. Change arrived with ChatGPT. Suddenly everyone was talking about AI, and entrepreneurs were extending the art of the possible.

By early 2025, AI was smart enough to be valuable, and people were using AI everywhere. Inference usage exploded. Suddenly, everyone remembered the lesson from Google search: speed produces more satisfied and more frequent users, while even tiny delays significantly reduce user satisfaction, search frequency, and search revenue. The market had shifted. Everyone realized that fast AI is more useful than slow AI.

By the end of 2025, it was clear: fast inference was powering the highest-value workloads. Fast inference was making engineers more productive because fast AI coding agents could write code, edit it, test it, and get a product to market more quickly. Fast AI made lawyers, analysts, bankers, doctors, and researchers more productive than their counterparts waiting on slow insights. For many workloads, Cerebras is up to 15 times faster than leading GPU-based solutions as benchmarked on leading open-source models. In some more exotic workloads, we have been more than 1,000 times faster.

In March 2026, we started a multi-year partnership with AWS to bring fast inference to an even bigger scale through global distribution. This is planned to give every startup, AI native, and enterprise company easy access to Cerebras’s blisteringly fast inference.

We were among the first semi-conductor companies to have a cloud business. We deliver hardware on-premises to customers concerned about data security and sovereignty. And we reach other customers through the cloud – both the Cerebras cloud and soon the AWS cloud. This allows us to reach customers who want to rent compute by month, year, or to pay by the token, and provides Cerebras with an attractive mix of recurring revenue as well as lumpier hardware revenue.

The fundamental building blocks in computer architecture are calculation (cores), the storage of results (memory), and the movements of results to where they are needed (communication). Our inventions are blurring the lines that have traditionally forced tradeoffs between memory capacity and speed. And the massive size of our processor serves as a foundation for pioneering communication techniques that are only available to wafer scale solutions.”

Financials

The current financials are nothing to get too excited about—the company mainly derives its revenues from AI compute in Abu Dhabi while due to the heavy R&D investment, the business is operating on sizeable operating losses (-40% EBIT margins):

Around one-third of operating losses are non-cash, but dilutive costs, in the form of share based compensation:

As already mentioned, warrants will be deducted from revenues which will give some revenue headwinds:

“When we begin to recognize contra-revenue amounts from the warrants in the first quarter of 2026, we expect quarterly revenue growth rates will decline from recent trends.”

Gross margins are fairly low at around 40% but this is still a small scale business. If the company can successfully scale up to provide compute-as-a-service to AWS and OpenAI, long term gross margins should be fairly attractive.

The two partnerships (so far)—with OpenAI and Amazon AWS—are with two of the biggest players when it comes to AI accelerator demand. This does give validation of the tech Cerebras has built. Another proof point is Nvidia’s acquisition of Groq’s SRAM-based inference tech, swiftly integrating the solution into a single disaggregated, high performance inference platform.

Next, we’ll go through our view on valuation and what we’ll be looking for at the IPO.

Read more

“We are minded to believe reports that TSMC has completely sold out of N2 and A16 capacity for 2028 due to over-whelming demand. The best tell is that we believe even Rapidus is attracting interest from tier-1 clients for leading-edge foundry, not just Samsung and Intel. This means that decisions made some time ago now will have a significant impact on market shares for leading fabless companies and that allocation (and yield) is now the key criterion of success. Those facing unanticipated demand may then be forced to consider alternatives, whether they like it or not. In such an environment, TSMC is likely to face difficulty serving incremental demand in the short-run, leaving pricing as the key methodology for rationing supply, for example via hot runs. It is no wonder that there are reports that sub-5nm nodes will see four consecutive years of price hikes of 5-10% pa, starting in 2026. This brings into view the possibility of TSMC sustaining a mid-60s gross margin (vs cons: 62- 63% and 60% last year) despite Arizona dilution and cost pressures (power, GAC). And yet, forcing customers elsewhere even if just for advanced packaging (eg EMIB) is unsettling and brings risk. As a result, we continue to expect TSMC’s traditional 95% leading-edge share to head down to 90% or potentially even lower, but regard that as normal in a very tight environment.”

We agree with Deutsche that leaving room to competitors to grab market share introduces some risks, as it will give competitors more cash and know-how to better compete going forward. TSMC is aware of this and is now stepping up N3 capacity around the globe coming online in 2027, this is CC Wei on the TSMC call:

“Historically, we do not add additional capacity to a node once it has reached its target capacity. However, as a foundry, our first responsibility is to provide our customers with the necessary capacity to unleash their innovations. Based on our assessment, to meet the strong demand in AI application, we are stepping up our CapEx investment to increase our N3 capacity. Thus, we are now executing a global capacity plan to support the robust multiyear pipeline of demand for 3-nanometer technologies, which are used by smartphone, HPC AI, including HBM-based dies, automotive and IoT customers.

In Taiwan, we are adding a new 3-nanometer fab to our GIGAFAB cluster in Tainan Science Park. Volume production is scheduled for the first half of 2027. In Arizona, our second fab will also utilize 3-nanometer technologies. Construction is already complete and volume production will begin in the second half of 2027. In Japan, we now plan to utilize 3-nanometer technology in our second fab and volume production is scheduled in 2028.”

While many were already screaming bubble in ‘23, consensus forecasts for AI demand have been consistently too low over the last three years. We also know that current installed capacity is not able to meet token demand, despite the boom in AI capex over the last two years:

Our central thesis is that we’re still in the very early stages of what AI can do, similar to 1998 when it comes to the rise of the internet, i.e. three years after the Netscape IPO. In subsequent decades, we saw massive new markets and companies being formed on top of this global network, from search to e-commerce, social media, the cloud, SaaS, the sharing economy, video streaming, etc.

To gauge long term AI demand, one way would be to estimate how much of global corporate expenses will be reoriented towards token consumption. For example, if we’re heading towards 10% of global opex being redirected towards AI workloads, this would mean $6 trillion in annual AI opex. Which can either come in the form of purchasing tokens over API—both from closed source models such as Claude and Gemini, or from open source models such as Qwen and Kimi— purchasing tokens within existing SaaS systems, or purchasing GPUs to host your own customized models.

If long term AI demand goes towards a ratio of 10% of global opex, we probably need about $15 trillion in AI capacity. Assuming this equipment has a lifetime of six years, this translates into around $2.5 trillion in required annual AI capex.

Personally, we think that long term AI demand could easily be higher. For example, 10% of global GDP being accelerated by AI workloads would mean $13 trillion in annual AI demand. This would require $32.5 trillion of AI capacity, or $5.5 trillion of annual AI capex.

This is why Jensen on a recent podcast mentioned that energy will be the biggest bottleneck to scaling up AI capacity. At $30 billion of investment per GW, $15 to $30 trillion of AI capacity would require 500GW to 1TW of power. The entire installed power capacity of the United States today is roughly 1.2 TW. The world combined sits at roughly 8.5 TW. So, we would need another 0.5x to 1x the capacity of the US power grid dedicated to AI compute.

Elon arrived at a similar calculation—although his estimate is that Tesla and SpaceX alone would need 1TW of compute. Roughly 80% of the compute of the initial fab with Intel is destined for SpaceX—which recently acquired xAI—to build space-based AI satellite constellations. The remaining 20% will focus on edge inference ASICs for Tesla Full Self-Driving, Cybercab robotaxis, and Optimus humanoid robots.

Putting AI data centers into space will be the key investment angle to IPO SpaceX at an astronomic valuation. SpaceX does have a strong position here—the company is transforming into a vertically integrated AI company. SpaceX will manufacture the GPUs, CPUs and Memory in its jointly operated fab(s) with Intel, build the satellites, and then launch this compute into space with Starship at the lowest cost in the industry. The $ cost to put a kilogram into space continues to fall and is estimated to drop below $100 with Starship—Elon has mentioned a target of $20 per kg.

To run Terafab, Intel will bring in the know-how of 14A manufacturing and advanced packaging such as EMIB, while SpaceX will bring in raised capital from its upcoming IPO. It sounds like Intel will bring in the IP, and then the fab(s) will be owned and operated by SpaceX. However, Intel has substantial shell capacity available and if SpaceX needs to scale up production down the road, it’s also possible that Intel can finally fill these fabs and then supply silicon to SpaceX and Tesla as a foundry. This is Musk on this week’s earnings call on the Terafab:

“Terafab is not some sort of mechanism to generate leverage over our chip suppliers. It’s just literally, we don’t see a path to having a sufficient quantity of AI chips down the road. As we scale production to high levels, just the rate at which the industry is growing in logic, but even more so in memory, we just anticipate hitting a wall if we don’t make chips ourselves. So that’s the reason for the Terafab.

We do have some ideas for how to make maybe radically better AI chips. And these are kind of research ideas, which means like long shot, but if long shot pays off, it’s maybe a giant improvement. And it’s just easier to do that if we have our own research fab and are developing our own production technologies. If you look long term at having AI satellites, making chips for those, no way the existing industry can keep up with that. It’s impossible.

So Intel is excited to partner with us on some of the core manufacturing technologies. We plan to use Intel’s 14A process, which is state-of-the-art, and in fact, not yet totally complete. But by the time Terafab scales up, 14A will be probably fairly mature or ready for prime time. So we think it’s going to be a great partnership.

In the near term, Tesla will be building the research fab on our Giga Texas campus. This is something we expect to be probably a $3 billion-ish initiative and capable of maybe a few thousand wafers per month, but it’s really intended to try out ideas. We have some ideas for improving the fundamental technology of how chips are made and there is some new physics we’d like to test out, but we also want to test out the ability to see if something is working in production. So you need a few thousand wafer starts a month to make sure that a production process is sound.

Then, SpaceX is going to take care of the initial phase of the scaled up Terafab. Any kind of intercompany thing has to be approved by both the SpaceX and Tesla Board of Directors. It has got to go through a conflict resolution because we’ve got to make sure Tesla shareholders are served and SpaceX shareholders are served, and then strike the right balance there. So it takes a while to work through the kind of independent director reviews on this.

So, Tesla is doing the research fab, SpaceX is doing the initial part of the large-scale Terafab, and then we got to figure out the rest.”

Besides a great new relationship, Intel is also finally seeing demand growth for its own silicon. Agentic AI requires code execution such as Python, web scraping, database querying etc., while GPUs are terrible at these types of sequential tasks. CPUs happen to excel here. This is Lip-Bu on the Intel call:

“Xeon server demand is seeing strong and sustained momentum. Customers are deploying server CPUs along accelerators in the ratio that is moving back towards CPU. The accelerator remains central to Frontier AI, but the backbone of AI computing in production remains a CPU anchored architecture. That is good news for the x86 ecosystem, and it is a structural reason I’m confident that the CPU franchise will continue to be a meaningful growth engine for the company in the years ahead.”

This should provide both Intel and AMD a nice tailwind in the coming years. However, long term, the biggest winner from this trend will be ARM. ARM is the key CPU partner of Nvidia while also hyperscalers continue to move workloads to custom CPUs based on ARM cores. Advantages are that ARM-based CPUs offer better power-to-performance ratios while hyperscalers can fully customize the design for their particular cloud workloads.

However, short term, Intel will see the biggest tailwind from this. As we saw above, TSMC doesn’t have enough capacity while we know that Intel’s fabs are running at low utilizations, by looking at their gross margins of 37% and also their revenues, which are down more than 30% since 2022. So, in the coming years, Intel should have the highest EPS delta to agentic AI benefitting server CPU demand.

Due to the shortage in power generation, any company that can generate electricity for AI data centers will continue to see strong demand growth. This is JP Morgan in their recent initiation on gas turbine supplier GE Vernova:

“The transformation since then has been dramatic as large gas turbine production capacity has emerged as a critical bottleneck in the AI-driven power and grid infrastructure buildout. GEV’s gas turbine order book is effectively sold out through 2029 (~83GW under contract at YE25, up 28GW from 55 GW at June 30) even as annual production capacity ramps to ~20GW by mid-2026, while Electrification backlog grew $11 billion year-over-year to $35 billion, and the Prolec GE acquisition adds captive transformer manufacturing in a globally supply-constrained market. This demand environment translates directly into clear visibility to margin improvement and free cash flow growth over the next 3-5 years as higher-priced backlog converts to revenue.”

We can see the same boom in demand for all other suppliers exposed to data center construction:

Overall, our view is that the market currently continues to strongly underestimate long term AI demand. Nvidia and other AI data center growth stocks have been climbing a wall of worries over the past three years, with continued questioning of what will drive these share prices higher (the answer is simple – EPS growth) and whether AI investments will turn out to be a bubble. The obvious answer is no, as all GPU capacity continues being taken up and is generating tremendous revenue growth for both AI model and cloud providers. Next, we will go through our key stock picks.

Read more

In just 16 months, Anthropic has now scaled from $1 billion in ARR to a $30 billion run rate. Anthropic is making these kinds of revenues in a world that’s still far short on AI compute and where enterprises can also deploy free open source models on their own GPUs.

So, the key bear thesis on AI capex—that AI isn’t making sufficient revenue—is quickly falling apart. Besides direct AI revenues, we already knew that AI is also generating tremendous indirect revenues by improving user feeds and ad targeting on social media. The other large opportunity is enterprise cost savings and boosting employee productivity—i.e. growing revenues on a flat or decreasing headcount. We’re only at the start of this trend, and the recent large headcount reductions by Block will likely be followed by more software companies. We estimate our own coding productivity has gone up 10-20x with Gemini and Claude, and so we would be surprised if headcounts at software companies don’t go down. However, overall hiring trends for software developers remains strong, as historically many companies and institutions simply didn’t have the required access to engineering talent.

Evercore looks at the tremendous amount of enterprise global opex at $60 trillion and the value that AI can create here. One way to think about this is if AI can save 20% in costs, the annual value of this cost saving would be $12 trillion, which dwarfs the $1 trillion in annual AI capex spend, while GPUs can power workloads for six years and more.

Another way to think about the above, if AI can make that global opex spend 2-3x more productive, the annual economic value added would be around $90 trillion. So, if AI would automate a large proportion of employees away in the knowledge economy, the economic value that this creates would allow for some sort of universal income to be paid out for those unable to find a new job. We briefly read through the Citrini piece that made a lot of headlines last month, and the scenario they outlined where large parts of the population become unemployed and a handful of players in Silicon Valley—like Jensen and Dario Amodei—would capture that economic value, we don’t see that happening. For the obvious reason that there isn’t going to be an electorate or political party that’s going to be in favor of this. Sam Altman has previously proposed to levy a tax based on market cap, e.g. 2.5% of the market cap of the S&P 500 would be remitted into a national public fund every year. Another way would be to impose a tax on every GPU compute hour, similar to the VAT in Europe.

Given the substantial economic value that AI is creating, we remain bullish on the outlook in AI both on the semiconductor side, which is heavily exposed to AI capex, as well as on the application side, which can make annually recurring revenues from AI inference. In the semi value chain, we liked the below chart from Evercore showing that ASML is most exposed of the semicap names to where that AI capex will flow—i.e. leading edge logic and DRAM:

On the other hand, we remain cautious when it comes to neoclouds. As the below analysis from Bloomberg highlights, if parts of data center construction get delayed, neoclouds will be heavily impacted as they are highly levered in order to do these data center buildouts, while revenues that bring in cash will be postponed.

Instead, we’re looking for companies with a strong market positioning—as witnessed in a dominant or strong market share in their niche—a long runway of growth, and ideally, a cheap valuation as the market is completely overlooking this opportunity. Next, we will dive further into our findings on five smaller names. We’ve invested in three of these although we’re also considering investing in the other two.

Stock Picks in Tech – Five Niche Leaders with Long Runways of Growth

Read more