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

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

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

On all metrics, token consumption is exploding. According to OpenAI’s recent enterprise report, API token consumption for reasoning increased 320-fold year-over-year. In the meanwhile, the current shift to AI agents—which communicate with each other over multiple steps—is again compounding reasoning volumes. Goldman notes that China enterprise token usage grew 162% from June to December, while China is even more short on compute than the West:

We can see on OpenRouter’s data that token consumption growth has strongly accelerated again as from January of this year:

While optimization techniques and hardware improvements have been driving sharp declines in the cost per token, demand is ramping at such a fast pace that capacity remains the binding constraint. This is UBS on OpenAI’s presentation at GTC:

“The cost of intelligence continues to fall rapidly, but demand growth still appears likely to keep the world compute constrained. The speaker highlighted sharp declines in token costs (99.7% decline in 3yrs) driven by both hardware and software improvements, while also suggesting demand is ramping fast enough that long-term compute availability remains a binding constraint. The broader implication is that “gigawatts” alone are too coarse a metric, with the shape and composition of compute becoming increasingly important as workloads diversify.”

A key focus in the coming years is on reducing latency. You don’t want a software engineer that costs $300-500k per year to sit idle for too long as he waits while Claude is generating code. This is again UBS on this trend after watching the presentation from Nvidia’s Bill Dally and DeepMind’s Jeff Dean:

“Models have improved materially on coding, math, and longer-running agentic workflows, with inference speed now emerging as the next key bottleneck. The panel suggested that agents can already sustain materially longer autonomous workflows than a year ago, with the next step less about raw capability and more about scaling throughput and lowering latency so agents can operate at practical speed (that far exceeds human speed). In that new paradigm, certain software tools developed for use at a human speed may need to be reimagined for much faster agentic workloads.”

This is the reason why Nvidia is now bringing to market a unique and low latency inference solution. Whereas Nvidia’s Rubin GPU will handle the memory-intensive prefill phase with its massive capacity of HBM, the new Groq LPX chip will accelerate the token generation in the decoding phase. Essentially, Groq LPX is a chip that heavily makes use of SRAM, resulting in an extreme amount of bandwidth and thus low token latencies:

This is Jensen on this new solution for high-performance AI workloads:

“So we acquired the team that worked on the Groq chips and licensed the technology, and we’ve been working together now to integrate the system. At the most valuable tier, we’re now going to increase performance by 35x. I would add Groq to maybe 25% of my total data center, and the rest of my data center is all Vera Rubin.

The reason why Groq was so attractive to me is because their computing system is a deterministic data flow processor. It is compiler scheduled, meaning the compiler figures out when to do the compute as the data arrives at the same time. All of that is done statically in advance and scheduled completely in the software.

The architecture is designed with massive amounts of SRAM, it is designed just for inference. And as the world continues to increase the amount of high-speed tokens it wants to generate, the value of this integration is going to get even higher. We’re going to fuse Groq and Vera Rubin, and use Groq to process the very last stage of auto regressive language models. That last stage is extremely bandwidth-intensive.

It would take a lot of Groq chips to be able to hold the parameter size of Rubin as well as all of the KV Cache that has to go along with it. So that limited Groq’s ability to really reach the mainstream and take off, until we had a great idea. What if we disaggregated inference altogether with a piece of software called Dynamo so that we could put the work that makes sense on Vera Rubin and then offload the low latency, decode generation part of the workload on Groq. And so we unified processors of extreme differences, one for high throughput and one for low latency.

It still doesn’t change the fact that we need a lot of memory. And so we’re just going to add a whole bunch of Groq chips, which expands the amount of memory. And so for a 1 trillion parameter model, we have to store all that in Groq chips. However, it sits next to Vera Rubin, where we hold the massive amounts of KV Cache that’s necessary in processing all of these agentic AI systems. So, the attention part of decode is done on Vera Rubin and the token generation part is done on the Groq chip. We run Dynamo, this incredible operating system for AI factories on top of it, and you get a 35x increase for token generation the world has never seen.

We’re in production with the Groq chip, and we’ll ship it in the second half, probably about Q3 time frame. I want to thank Samsung, who manufactures the Groq LP30 chip for us, and they’re cranking as hard as they can. The sampling of Vera Rubin is going incredibly well. And in fact, Satya texted out already that the first Vera Rubin rack is already up and running at Microsoft Azure. We have now set up a supply chain that can manufacture thousands a week of these systems, essentially multi-gigawatts of AI factories per month inside our supply chain.”

Groq LPX in combination with Rubin especially boosts the bandwidth and thus the interactivity per user:

UBS came away thinking that the risk-reward is towards the upside of their revenue estimates:

“In a very well attended session, management highlighted the potential sources of upside to the updated backlog slide presented yesterday and we walked away a bit more optimistic as to how much LPX (the new fast inference solution from Groq) could add in C2027. We took from the discussion that ~$50B could be a reasonable estimate of how much LPX would add on top of the ~$460-470B implied baseline for data center revenue in C2027. On top of this, of course, would be additional business that comes in within lead times – which could be as high as ~$50-75B or even more. This is because we estimate data center will end up at least $40-50B higher in C2026 than what was implied when the company first provided this backlog last year – ergo, with 21mos still left until the end of C2027, the company should be able to add at least this much to C2027. So net of all this, we could see a data center revenue number potentially pushing $600B in C2027 (UBSe ~$520B).”

The golden question is how AI demand will come in over that fiscal ‘29-’31 period for Nvidia, as we already know that this year (FY27) and also next year (FY28) will come in strong. However, the market is assuming we’re now heading close to a peak in terms of AI capex spend which we can also see in UBS’ numbers:

As a thought exercise, we tried to estimate Nvidia’s future annual revenues based on where AI demand will go five-six years from now in 2031. We start from current AI capacity, which is basically the sum of Nvidia’s revenues over the last three years. So, this should be roughly the current installed base excluding Google’s TPU-based data centers. This results in $407 billion of current AI capacity.

We know that this is by far not enough—the market is still undersupplied while token consumption will continue to explode. If we assume that current AI capacity satisfies 10% of future AI demand in 2031, we need an installed base of just over $4 trillion of Nvidia equipment by 2031. If we assume a lifespan of 6 years for this equipment, that demand will be fulfilled by the equipment Nvidia ships over the 2026-2031 period (6 years in total). This would mean Nvidia would sell on average $678 billion of equipment per annum in this time period:

AI demand going up 10x by 2031 is our personal best guess. Of course, it could easily be 15x, or it could only be 5x. There are a lot of moving parts—on the one hand, models are growing larger but also optimization techniques for inference are driving the cost per token down. What we feel very confident about however is that we’re very early in the adoption curve. This can easily be witnessed in the steep growth in token consumption, as well as the continuously expanding use cases of AI—from coding and AI agents in SaaS to realistic video generation, robotaxis, humanoid robotics etc. This is UBS on how AI agents will become an important driver for SaaS revenues for example:

“Also particularly intriguing to us was the CEO’s perspective that the future for software industry with AI, might shift toward monetization focused more on AI token reselling instead of historical software tool licenses/SAAS. In this new paradigm, future software companies’ COGS increase and come in a form of token consumption. At the same time, NVDA expects cloud business model to evolve from instance rentals to monetizing tokens per second (with various tiers offered at different prices). NVDA sees hyperscalers aiming to optimize tokens-per- second-per-watt.”

However, a potential problem for Nvidia is that all hyperscalers are developing their own ASIC accelerators and will gradually shift more demand towards this in-house silicon, similarly to what happened at Google already. Or what has been happening in the data center CPU market with the shift towards Graviton ARM-based CPUs at AWS. This is William Blair on Meta’s recent progress with MTIA as an example:

“In a blog post yesterday, Meta provided greater details on its custom ASIC roadmap. Meta noted that the current generation of MTIA 300 chips is already in full production, with three additional generations targeted for release in 2026 (MTIA 400) and 2027 (MTIA 450 and 500). All four generations leverage Broadcom as a key design partner, with the company contributing significant IP for I/O, networking, and chiplet connectivity. Recall that on its most recent earnings call, Broadcom noted it expects Meta to deploy several gigawatts of its next-generation XPUs in 2027 and beyond. The announced roadmap provides greater details on the design priorities for these new chips (with a refocus on LLM inference) and helps instill greater confidence that Meta will be able to scale its ASIC deployments over the next few years.”

We know that hyperscalers will be aggressively ramping ASIC capacity from Broadcom’s calls. This is from JP Morgan:

“We expect Broadcom to generate close to $65B+ in AI (ASIC + networking) revenues in FY26, led by its primary TPU customer, Anthropic ramp, the first- generation ramp of SoftBank/ARM XPU ASIC programs, and incremental revenue contributions from other AI ASIC customers such as Meta and Bytedance, along with continued strong AI networking revenue. For FY27, Broadcom anticipates close to 10 GW of total compute capacity being deployed across its six XPU customers. Specifically, OpenAI has now become a qualified customer and is expected to deploy over one gigawatt of compute capacity. Anthropic’s order deployment is projected to increase 3x Y/Y in FY27 to 3 GW. Meta remains on track with its internal ASIC program, with multiple gigawatts expected to be deployed in FY27. Customers 4 and 5 (which we believe to be SoftBank/ARM and Bytedance, based on a process of elimination given Broadcom’s disclosures) are expected to have strong shipments this year and more than double in 2027. All of this is in addition to the continued strength of Broadcom’s lead customer and the ramp of the TPU v8 (Sunfish) program. Overall, we estimate Broadcom can drive $12–15B of revenue per gigawatt and conservatively see AI revenues at $120B+ in FY27. Reflecting increasing AI customer diversification, we estimate Google TPU business will decline to less than 45% of overall AI revenues in FY27, down from 50–52% in FY26 and 60% in FY25—even as total AI revenues grow to over $120B in FY27.”

Based on these numbers we can do some modeling. We know that Nvidia makes around 50-60% of their revenues from hyperscalers, so if we then compare these revenues to Broadcom’s AI revenues, we know that Broadcom will grow its market share at the hyperscalers to around 27% by next year:

Will ASICs disrupt the Nvidia investment case? Next, we’ll dive further into developments in AI and three trades we’re making.

Additionally, humanoid robotics should be one of the big growth stories in tech in the coming decades. While the robotics field is competitive, we recently came across a well-positioned components supplier with good engineering prowess that started signaling high growth from both humanoid and quadruped robotics in the coming years. Typical investment criteria we look for are—high market shares and margins, combined with a long runway of growth so that we can simply buy and hold for the next 5-10 years with the aim of seeing shares compound into attractive multibaggers.

Read more

HBM (high-bandwidth memory) is a workhorse in the current AI boom. In HBM, DRAM dies are stacked on top of each other and connected through TSVs. This design allows for loads of fast working memory to be located close to the GPU. Bonding these DRAM dies together is a key step in back-end manufacturing (advanced packaging) as it determines the interconnect density, and thus speed and bandwidth.

This bonding process was dominated by Hanmi Semi over the last few years as the key supplier to SK Hynix (and Micron). However, the SK Hynix relationship soured when the latter started evaluating equipment from other suppliers. Currently, HBM die bonding is dominated by a technique called thermo compression bonding (TCB), and Hanmi Semi’s erratic behavior towards their key customer has now allowed ASMPT to get a strong foothold in the TCB market. ASMPT’s growth has exploded in the TCB market, and is now guiding for a 35-40% market share in this high-growth TAM. This is the CEO on the recent call:

“TCB momentum strengthened further in 2025 with significant new orders across logic and memory, solidifying our TCB technology leadership. We established deep engagement with both logic and memory customers and saw encouraging traction in areas such as HBM and C2W (Chip-to-Wafer) ultrafine pitch applications. This continues to reinforce our position as a leading provider of advanced packaging solutions as customers move to more complex chiplet-based and high-density architectures.

This time last year, we expected the TCB TAM to reach around USD 1 billion by 2027. Since then, the landscape has evolved meaningfully. The acceleration of AI-driven investment—especially in advanced logic and high-bandwidth memory—has expanded the market significantly more than our earlier assumptions. Based on our latest projections, we now estimate the TCB TAM to grow from roughly USD 759 million in 2025 to USD 1.6 billion by 2028, representing a CAGR of 30%. This reflects sustained adoption of 2.5D architectures, higher HBM stacks, and the industry’s move towards final pitch interconnect. These are all areas where TCB is increasingly the preferred solution.

Our target market share remains at 35% to 40%. With our breakthrough into the competitive HBM market, we grew TCB market share significantly, achieving record TCB revenue growth of 146% year-on-year. In 2025, our advanced packaging (AP) revenue growth of 30% year-on-year was driven by TCB.

Now, let’s look at TCB more closely. In logic, our C2S (Chip-to-Substrate) solution maintains its dominant position as a process of record with a steady flow of orders from key OSAT customers in 2025. Extending into early 2026, we are pleased to share that we have secured additional orders for 9 more TCB tools from the same customer. We are well positioned for further order wins as the market shift towards larger compound lines. At the same time, our C2W ultra-fine pitch platform secured orders for 2 tools in February 2026 from a leading customer. Since the announcement, we have secured 2 more such TCB tools from the same customer. As the industry transitions from mass reflow technology to TCB, ASMPT stands to benefit significantly as the preferred C2W solution provider offering plasma enabled capabilities.

In memory, we deepened our engagement with several customers and continue to expand our share with shipments in Q4 2025. Our tools have demonstrated superior performance with industry-leading production yields and interconnect quality. We were also the first to secure HBM4 12-high orders from multiple players, and we are now leading HBM4 16-high development with our flux-based TCB tool deployed for sampling, and our fluxless AOR-TCB process under qualification. These are important milestones for our technology leadership as HBM architectures scale further.”

During the Q&A we got more details on ASMPT’s positioning for the coming years:

“Currently, the HBM portion in the TCB TAM is definitely much larger than logic. But as we go a few years out, this dynamic will actually shift where logic, especially CoW, should take a larger share. But we cannot share the specific split for confidentiality reasons or competitive reasons. ASMPT is very strong in CoS and we are also making wins on CoW. Recently, we won two tools for C2W application and then we won two more. So I think it is a signal that we are also being recognized as a solid solution provider for the C2W space.

Now on the memory side, I think the competition landscape is different. We have a strong incumbent in the memory space, but a year ago, we broke into the HBM market and so we have gained market share there as well. In 2024, we had practically zero share in HBM. In 2025, we have a strong foothold in the memory market. For 2026, orders will come as our customer rolls out 16-high HBM in volume. As the industry moves from 12-high to 16-high, I think all equipment suppliers, including myself for TCB, are waiting anxiously for that particular customer to allocate TCB demand. So at this moment, we feel that 2026 will be a year whereby there will be new true demand for TCB for 16-high, but exact timing, I cannot give you any visibility at this point in time, but it cannot be too long before we know in our opinion.”

For years, Hanmi Semi (a South Korean equipment maker) was the dominant supplier of mass reflow bonders to SK Hynix, the leading player in HBM. However, diversifying the supplier base is fairly common practice for all leading fabs. For example, Intel for a long time kept sourcing a smaller share of its less crucial lithography tools from Nikon, even though those had far inferior performance compared to ASML. The reason is that they wanted to keep Nikon alive in the lithography race and not become fully reliant on ASML. So, it’s fairly common to give 70-80% share to the best equipment maker, and then 20-30% to a secondary player so that the fab can keep multiple options open. This also provides funds to the second player to stay alive in the innovation race.

So, in our view, it’s fairly normal that SK hynix started placing small orders with Hanwha Semi and ASMPT to test out their TC bonders. However, according to media reports, Hanmi subsequently recalled 60 engineers from SK Hynix’s fabs which were servicing the bonding equipment. Additionally, Hanmi notified SK Hynix of a 25% price increase. This led SK Hynix to accelerate its shift towards another strong supplier.

Because ASMPT’s equipment had apparently been successful in testing, this allowed the company to capture substantial TCB order flow in 2025. The Elec mentions that ASMPT has captured half of HBM4 TC bonders so far:

“SK Hynix is expected to place a large order of a hundred set of TC bonders for the production of HBM4 in March. This is when it plans to expand 1b DRAM production at its M15X fab at Cheongju. It remains to be seen how the chipmaker places orders for these bonders to ASMPT, Hanmi Semi and Hanwha Semi. SK Hynix uses around fifty sets of TC bonders for the production of HBM4 at the current stage, and half of these have been supplied by ASMPT. Hanwha Semitech’s bonders were not in use, sources said.”

According to Electrical Engineering World, ASMPT’s bonders are outperforming those of Hanmi Semi:

“A person familiar with SK Hynix said, ‘During the process simulation of HBM3E 16-layer, we found that the performance of ASMPT equipment exceeded that of Hanmi equipment as the number of stacked layers increased.”

So, it’s not really surprising that Hanmi Semi has been reporting disappointing revenues. This is JP Morgan on the company’s recent results:

“Hanmi’s 4Q earnings were a miss, with the OP coming in 40-60% below JPMe/ Bloomberg consensus with TCB sales recording the lowest since 4Q23. After two quarters of earnings misses from HBM4 qualification delays, we expect TCB earnings to recover in 1H26 from memory makers resuming TCB orders. Despite memory capex entering an upcycle, however, we see unfavorable risk/reward with Hanmi trading at a 30%-35% premium to peers with Samsung TCB optimism priced in. We remain UW with a PT of W150K based on 29x FY27E P/E and recommend investors trim at strength.

Hanmi shares have outperformed peers YTD which we attribute to: 1) memory capex entering an upcycle; 2) Samsung TCB optimism; and 3) the market’s perception of Hanmi being a higher beta to memory stocks. We agree that memory capex is entering an upcycle, but we emphasize that the upside risk is on the front-end capacity and conventional DRAM more than HBM where we do not sense any upward revision to HBM capacity in the next 12M. Samsung is testing both HCB and TCB for 16-Hi HBM4E, and while we expect Samsung to converge to TCB, we see limited reasons for Samsung to diversify TCB vendors.”

Samsung is heavily relying on its internal equipment maker, SEMES, to supply its TCB tools. This means that it is unlikely Samsung will adopt Hanmi’s equipment, unless Hanmi’s equipment strongly outperforms SEMES tools. To make matters worse for Hanmi, Macquarie points out that Hanmi could also lose share at its other key customer, Micron, which has been testing out ASMPT as well as Besi:

“Hanmi continues to face Hanwha Semi and ASMPT in SK Hynix’s value chain. At Micron, Hanmi remains dominant, but Micron has tested ASMPT and Besi, increasing competitive risk. While Hanmi entering Samsung Electronics’ HBM bonding is a catalyst that the market awaits, we see Samsung prioritising its subsidiary SEMES and considering alternatives rather than adopting Hanmi tools.”

Asian broker CGSI models in revenue growth for Hanmi with as demand source Chinese packaging fabs:

However, this is a somewhat unreliable source of demand as increasing sanctions pose a looming risk, especially as Huawei is the end customer which has been a constant target of sanctions. From CGSI:

“As China pushes for advanced chip localization, centered on Huawei, we expect Chinese outsourced semiconductor assembly and test players (OSAT) to place large-volume orders for critical packaging process tools to Hanmi and key Japanese players, driven by Huawei’s production ramp-up for artificial intelligence (AI) chips amid the depletion of pre-secured HBMs. We expect Chinese HBM players to start mass volume shipment in 2H26F and project their monthly HBM through silicon via (TSV) capacity at 60k by end-2026F and 127k by end-2027F. Assuming 50% thermo-compression bonder (TCB) market share of Hanmi in China, we expect its China-bound bonder exports to jump to W164.6bn in FY26F and W183.9bn in FY27F from W58.9bn in FY25F, contributing 38% and 40% in total TCB sales in FY26F and FY27F, respectively, up from 17% in FY25F.”

Overall, the TCB semicap market is an interesting space for long term investors. Not only does it play a crucial role in the HBM market, but it’s also seeing increasing adoption in the logic market for 2.5D chiplet integration which is crucial to pack more compute into a single system. As Moore’s Law has slowed, the obvious way to pack more transistors into a single computing system is to simply cobble more dies together. AMD has been at the forefront of this trend:

ASMPT is forecasting a 30% CAGR in the TCB TAM, while ASMPT is the player with the momentum in this market. Obviously, the company is well-positioned to capture a strong market share of this future TAM.

Is it time to invest in ASMPT? Or are other names even better positioned to capture the value from the attractive outlook in advanced die bonding for logic and HBM end-markets? We’ll dive further into this space next.

HBM/Die Bonding Semicap Outlook

Read more

MRAM (Magnetoresistive Random Access Memory) combines the best of both worlds in memory—the permanent storage capabilities of NAND and the high speed read and write capabilities of DRAM. Its permanent storage capabilities are even better than NAND, as NAND flash physically degrades every time you erase and write new data to it. So, SSDs have a limited lifespan. On the other hand, MRAM has no physical wear and tear which results in a near-infinite endurance. MRAM also has been called ‘universal memory’, as it theoretically could tackle every use case.

However, the MRAM market remains small today, with as key reason that the cost per bit remains high. Nevertheless, the MRAM market should see attractive growth rates in the coming decade(s) as it plays a crucial role in edge intelligence such as satellites, autonomous vehicles, industrial robotics etc. FMI forecasts the MRAM market to grow at an 18% CAGR over the next 10 years:

“The global MRAM market is expected to grow from an estimated USD 912 million in 2025 to USD 4,773 million by 2035, reflecting a robust CAGR of 18.0% over the forecast period. Growth is primarily driven by the increasing demand for high-performance memory solutions in automotive, aerospace, industrial, and consumer electronics applications.

MRAM offers distinct advantages over conventional memory technologies, including non-volatility, low power consumption, high endurance, and faster read/write speeds, making it an attractive option for next-generation computing systems and IoT devices. Adoption is particularly strong in embedded applications where energy efficiency and reliability are critical, such as in automotive ECUs and industrial controllers.”

Aerospace (satellites) should become a key end-market as MRAM is radiation immune. Traditional memory, like SRAM and DRAM, stores data using electrical charges. In the harsh environment of space, cosmic rays and high-energy particles can alter charges in traditional memory causing bit-flips, i.e. accidentally changing a 1 to a 0. Because MRAM stores data magnetically rather than electrically, it doesn’t have this problem. So, you can use MRAM on satellites without the bulky shielding required for standard memory. Additionally, LEO satellites have a short lifespan of around 5 years or less, making this a business with regularly recurring revenues as satellites have to get replaced.

Everspin (ticker MRAM, enterprise value $159 million) is a leading company in the MRAM space. The CEO guided for their revenues to nearly double in the coming years:

“Before I close, I would like to discuss our long-term strategy, which entails reaching $100 million in annual revenue over the next 3 to 5 years. We believe this growth will be driven by the ramp of new products, most notably our new XY parts in our PERSYST product portfolio, such as the 64-megabit part I described earlier and continued solid growth in our Toggle MRAM and licensing business. Our new PERSYST X5 parts are getting solid traction. They are offered in densities from 4 megabit to 256-megabit and include the power-optimized SC families and the high reliability or HR families with quad and optical SPI interfaces.”

As we know, memory is currently supply constrained with manufacturing capacity heavily being reoriented towards HBM and DRAM. Can this have downstream effects where excess demand starts flowing over into the MRAM market next? Is Everspin next to benefit from supply shortages in the overall memory industry? We’ll take a deep look at all relevant topics next.

Everspin & the MRAM market

Read more

We’re interested in the power semi space for a number of reasons. First, these types of semiconductors will start playing an increasing role in both the AI data center and their connected power sources. Second, these are not expensive stocks, as these companies are still crawling out of the semi downturn post the covid electronics boom. Thirdly, these companies are exposed to a variety of attractive end-markets, ranging from robotics to EVs, drones and green energy. Robotics and increasing levels of automation in the wider economy can provide a big tailwind over the coming ten years, making mature semis in general interesting plays on the long term growth in edge AI.

William Blair notes that ON Semi can make $100k per MW of AI compute:

“Power is the key bottleneck to AI infrastructure, making power density and efficiency critical. Onsemi now estimates its power semi content per next-gen 1 MW AI rack has doubled from $50,000 to $100,000, supported by its broad SiC and GaN portfolio. The company generated $250 million from these applications in 2025 with high- teens sequential growth, with much larger upside expected as higher power architectures scale toward 2030.”

Infineon is another key player in power semis and guides for $12k–$15k of content for a current 130kW AI rack. Taking the midpoint, $13.5k of power semi content per 130kW rack means $104 per kW, or $104,000 per MW.

If 30GW of AI compute will be added per annum, this means a power semi TAM of around $3.1 billion from AI compute alone:

Based on Infineon’s recent comments on the conference call, it looks like the German leader in power semis will grab a large share of this market:

“Our unmatched portfolio breadth, system understanding, speed of innovation, quality and delivery capability make us the partner of choice for all relevant GPU makers and cloud providers. Our revenues from AI power solutions keep growing steeply. We re-confirm our target of around EUR 1.5 billion for this fiscal year. This number is still supply limited, capped by how fast we and our manufacturing partners can bring up capacity. And let me be clear, in contrast to some peers, which in their disclosure do not differentiate between AI and other enterprise data center business, the EUR 1.5 billion of Infineon are purely related to AI. Separately, we do another EUR 500 million, give or take, of classical data center revenues.

Meanwhile, projections of build-out of AI data centers and related infrastructure keep going up. More and more, AI is showing tangible benefits in real-life use cases. Several of these we witnessed in our own company, for example, improving speeding up chip design, software development or customer engagement. The demand from our customers for the coming years is still going up quarter by quarter. Therefore, we are doubling down on AI. We have decided to pull in EUR 500 million this fiscal year to accelerate AI-related capacity increases in order to fuel growth beyond the running fiscal year.

These investments, including the conversion of existing IGBT capacities are capital efficient and will underpin our market leadership. Large part of them will be used for a faster than originally planned ramp-up of our new power and analog/mixed-signal module fab in Dresden. Fortunately, we are able to accelerate the opening of the facility to the summer. With the additional capacity, we project to achieve AI-related revenues of around EUR 2.5 billion in our 2027 fiscal year. Said differently, we will add another EUR 1 billion of margin accretive business next year, thus expanding our AI revenues by a factor of 10 within just 3 years.

We are not only engaged with 1 or 2, but with all hyperscalers and relevant GPU makers across the globe. And for us, of course, the second stage part is giving us particular growth momentum as vertical power delivery modules have a much higher value. It’s the biggest opportunity I’ve seen for the company in the last 30 years, but also technically very challenging with very demanding customers.

From a utilization point of view, you can assume that the 300-millimeter factories will soon be fully loaded as we are converting IGBT to AI MOSFETs. Then, we still have some idle capacity in non AI-related technologies, but also here, we expect a certain spillover or halo effect from this AI boom, meaning that MOSFETs going into telecommunication or automotive will also see demand.”

Based on what the various competitors are saying on their conference calls, we believe that Infineon currently has a fairly dominant share in this market.

Wolfspeed illustrates that power semis are also used in cooling, as well as to convert power from the grid so that a single GPU can work at a lower voltage level. This will mainly be done by SiC (silicon carbide) whereas GaN (gallium nitride) is more heavily used at the rack level.

The reason is that SiC is better at handling high power, whereas GaN is better at high frequencies. Yole estimates the SiC market to grow at a 21% CAGR from 2026 to 2030, but GaN should grow faster as it will take the larger market share in the AI data center.

Let’s go through some of the players in this market. If dominant player Infineon can’t keep up with demand, it’s likely that the other players will see a demand boost as well.

Wolfspeed, the SiC Specialist

Wolfspeed is a pioneer in SiC and has historically held a strong market share in this market. Infineon actually tried to acquire Wolfspeed in 2016, which is how we first gained knowledge of Wolfspeed’s strong SiC capabilities. However, this acquisition got blocked by the US government on national security grounds, as Wolfspeed makes SiC modules for the US military.

Wolfspeed has now completed a state of the art SiC fab in New York State for $6 billion, however, as the EV market went into a downturn over the previous years, the company blew up due to excessive leverage. Basically, Wolfspeed loaded up on debt to build this fab but when SiC demand came in lower than expected, shareholders got wiped out.

This debt load has been restructured now—basically, the old debt holders have now become the new equity holders, and trading resumed a few months ago on the New York Stock Exchange. The good news is that the financial situation is now looking much better—the fab is ready, allowing the company to limit capex in the coming years, while the balance sheet has now $1.3 billion in cash and with a reduced debt load of $2 billion, giving an overall net debt figure of $700 million.

The bad news is that fab utilization is still low, which resulted in negative gross margins last quarter, and so the company will need to see either strong growth coming from EVs and industrials, or alternatively, start winning deals in the AI data center space. The cash burn was $74 million last quarter, so the $1.3 billion in cash gives shareholders more than 4 years of runway at the current state. However, new management is heavily focused on cutting costs while at the same time getting those new SiC wins. The company had a design win with Toyota last quarter for example for their battery electric vehicles (onboard charger SiC chip). At the same time, also Wolfspeed’s AI data center revenues are growing tremendously, although this will be from a very low base.

This is new CEO Robert Feurle, who is semiconductor veteran with previous VP roles at Infineon and Micron:

“We have continued to build solid momentum across the business this quarter from achieving 50% quarter-over-quarter growth in AI data center revenue to producing a 300-millimeter silicon carbide wafer and securing key customer wins. We’re concentrating in a few key areas: strict financial discipline, advancing our technology leadership, and driving operational excellence. A central theme across these priorities is diversifying our revenue base, particularly in industrial and energy, including applications tied to AI-related power demand and grid modernization.

We remain focused on driving costs out of our footprint as we navigate underutilization headwinds. We have officially completed the shutdown of all 150-millimeter device production in our Durham campus, roughly a month ahead of schedule, transitioning our entire device platform to higher efficiency 200-millimeter manufacturing. We continue to improve production efficiency and speed to optimize the earnings potential of the business. The results of these efforts will be even more apparent when demand accelerates and we begin to increase fab utilization.

We have organized our go-to-market strategy around 4 verticals that we believe will drive growth in our business in the near to midterm—automotive, industrial and energy, aerospace and defense, and materials, and we are already seeing strong traction from these early efforts. Our first vertical, automotive, remains a core market despite muted EV demand due to a mix of macro and structural factors, which include higher interest rates in the U.S. and Europe, the elimination of certain government incentives in the U.S., excess supply across the market and intensifying competition globally, including China. Despite weaker near-term demand, our portfolio is aligned with OEMs that provide us efficiency, range and power density. A great example of this is our recently announced partnership with Toyota, one of the most respected and quality-driven automakers in the world, to power the onboard charging systems for their BEVs.

In industrial & energy, we are leveraging our expertise to expand our reach, concentrating on AI data center power, grid storage, solid-state transformers and broader grid modernization applications. We have the expertise to extend into the AI data center, which operates at significantly higher voltages than legacy data centers. As voltages increase, we believe an increasingly larger portion of this addressable market will be better served with silicon carbide technology from grid to rack, compared to legacy silicon-based solutions. Our devices are already embedded in critical AI data center power systems, and we have doubled our data center revenue in the last 3 quarters with 50% quarter-over-quarter growth from Q1 to Q2. Further, we are actively collaborating with a broad ecosystem of partners to support the industry transition from legacy 400-volt architectures to next-generation 800-volt AI platforms.

Data center build-outs and widespread electrification have driven a surge in global energy demand. We’re already seeing silicon carbide adoption across wind and solar applications as evidenced by our recently announced collaboration with Hopewind to advance the next generation of wind power solutions.

Turning to aerospace and defense. We believe there is a growing opportunity due to the tailwinds from defense modernization and electrification, including direct energy platforms. U.S. government has already recognized silicon carbide as strategically significant to national security with both the Department of War and the Department of Energy designating it as a critical material. Additionally, the U.S. government has emphasized the strategic importance of secure domestic semiconductor supply chains for national security applications.”

During the Q&A, Feurle gave more details on what they are working on in the AI data center:

“Today, you’re around 100 kilowatt per rack. That’s moving in 2 years to like 600 kilowatt per rack and into like a megawatt rack in the 2029-2030 time frame. This means you have to figure out how do you power these racks and this is where Wolfspeed can play, stepping that voltage down. And then, as more and more renewables come into the mix, you also need a lot of energy storage systems in between. So, that’s the next portion where we are focused on. Then, you need to get this energy into the data center with transformers, right? And there is a transition happening from traditional transformers to solid-state transformers where silicon carbide is the perfect solution.”

Wolfspeed is currently making around $475 million in annual revenue from power semis and $200 million from SiC wafers. From previous management, we know that the SiC power semi fab has a revenue capacity of around $2 billion, which could result into $1.2 billion of annual FCF. We’re not sure yet whether new management will confirm these numbers, but they provide some guidance on what a bull scenario could look like.

The current market cap is $722 million, and with the net debt of $700 million, this results in an enterprise value of $1.4 billion. So, currently, the shares are valued at an EV-to-Sales of around 2x.

Let’s work out a realistic scenario to see if there is substantial upside. If a cyclical recovery in EVs and industrial end markets doubles revenues, and then the company wins another $500 million in AI data center sockets, it’s well possible that over time we can add another $1 billion in revenues. Assuming a 50% contribution margin, this should result in $500 million of additional FCF. With further cost optimizations, this could well result into $200-300 million of overall positive FCF for the company on an annual basis. Valuing this at 15x FCF would give a market cap of around $3.8 billion.

So, Wolfspeed could be interesting here. There should be a coming boom in power semis driven by AI demand and we’re also at the start of a cyclical recovery in industrial end-markets, going through conference calls from TI, ADI and Microchip. And then if Warsh comes in to cut rates, this should boost automotive demand such as EVs.

We also think that Wolfspeed should see more design wins due to its US-based fab capacity as hyperscalers and GPU makers increasingly secure their supply chains, such as by shifting more advanced logic manufacturing to TSMC Arizona. At the same time, Wolfspeed’s current cash position and cash burn give new management ample runway to turn things around.

The main risk would be a macro downturn in the coming years. Should there be a credit, geopolitical, or other event impacting the global macro, the risk would be that Wolfspeed’s cash burn deteriorates further which will no doubt send the shares quite far south. However, due to the strength of the current AI investment cycle and the demand this is creating for power semis, we do think the risk-reward is towards the upside here for Wolfspeed. Needless to say, this remains a speculative story as the company will need a substantial amount of design wins related to AI data centers to turn things around.

Next, we will dive into opportunities in power semis and the overall semis industry, including a few likely layups for investors. Finally, we’ll discuss our findings on AMD.

Read more

One of the big debates in AI is whether LLMs will be a commodity, or whether these will be attractive businesses where a handful of big companies will dominate like we have in big tech today. The commoditization camp will point to charts like the one from Nvidia below, pointing out that open source models are only six months behind and with the gap in capabilities continuing to narrow:

A few key points though to note here. Whenever we’ve tried open source models, the output we get is of lower quality than the leading models. What is likely going on in the chart above is that open source models are heavily trained to perform strongly on benchmarks, whereas with leading models there is also a strong focus to perform in the real world. This is the classic statistical phenomenon of in-sample versus out of sample performance. Just because a statistical or quant model performs very strongly on in-sample data doesn’t mean it will perform strongly once you test it on real data it hasn’t seen before. For example, in quant investing, a backtested and data-mined performance shouldn’t get confused as a reliable indicator for real world results.

Despite the chart showing a narrowing performance, our view is that real world performance between leading models and the rest of the followers will continue to widen. Before, training was heavily dominated by pre-training, which is the phase where you train the model on the available body of human knowledge. However, over the last twelve months, both reinforcement learning and chain-of-thought reasoning have become crucial to get state-of-the-art problem solving capabilities. With reinforcement learning, models can now be trained on unlimited data in particular fields such as mathematics and coding as the model’s proposed solution to synthetic problems can objectively be verified.

This is similar to DeepMind’s training process of AlphaGo nearly a decade ago, where the model that beat the world champion in the boardgame was not trained on human data, but purely on synthetic data by having the model continuously play Go against itself. So, we believe that scaling laws in current AI training are alive and well, and that the increasing capital intensity will make it hard for new players to enter the field. You either need large revenues or access to large amounts of funding in order to compete in this game.

Going forward, another differentiator is that leading firms will have the capabilities to build high-value data sets themselves. This includes hiring investment bankers to teach models how to do financial modeling, lawyers to set up research workflows for legal cases, scientists for scientific research, etc. As the model arrives at its final answer via chain-of-thought reasoning, competition aiming to distil a leading model over API can only memorize this final result. However, these distilled models will struggle to resolve new real world problems themselves.

A positive for the big tech firms is that they also have access to unique data sets. Good examples here are Meta with Facebook and Instagram, xAI with X, and Microsoft with LinkedIn. While other providers of high quality data sets such as Reddit will likely continue to steeply increase pricing in return for training access. Amazon is also already blocking access to their shopping site for AI agents, but presumably for a hefty fee this could get unlocked as well.

The key point from the above data points is that capital requirements in order to compete in this race will only continue to grow. Historically, these types of markets have not resulted in highly competitive landscapes with a plethora of competition, but rather a handful of competitors at best. While there is a widespread narrative that LLMs are commodities, we don’t believe this is actually the case today and that it will be even less true in the future. OpenAI and Anthropic scaled to $10-20 billion ARRs within a few years—the fastest ever— providing objective evidence that these businesses have clear differentiated value.

The only counterargument to the above we can see is that big Chinese tech firms can legally purchase high-quality training sets via Chinese data exchanges. For example, while US medical data is fragmented across thousands of private hospitals on different systems, Chinese AI firms have access to large scale anonymized medical records giving an obvious competitive advantage in fields such as medicine and biology.

Another advantage is that China has a ton of engineers which could more cheaply assist models to build workflows. A decade ago, China was famous for data labeling farms, where low-skilled workers tagged traffic photos to train autonomous driving models on. Today, companies like DeepSeek and Baichuan use expert-in-the-loop systems where junior engineers and field specialists write chain-of-thought reasoning paths for models to follow. A big drawback for China is that the country remains heavily short on compute. However, in the long term, it’s likely that China will have the advantage in terms of power generation.

Nvidia vs AMD – Apple vs Android?

At CES, Jensen explained how Rubin will give 5x the performance of Blackwell, despite the number of transistors only increasing by 60%:

“This is the Rubin GPU. It’s 5x Blackwell in floating performance. But the important thing is to go to the bottom line, it’s only 1.6 times the number of transistors of Blackwell. That tells you something about the levels of semiconductor physics today. If we don’t do extreme co-design at the level of every single chip across the entire system, 1.6x puts a ceiling on how far performance can go each year. One of the things we did is called NV FP4 tensor core. The transformer engine inside our chip is not just a 4-bit floating-point number, it is an entire processing unit that understands how to dynamically adjust its precision to deal with different levels of the transformer.

So, you can achieve higher throughput wherever it’s possible by losing precision, and go back to the highest possible precision wherever you need to. You can’t do this in software because obviously it’s just running too fast, and so you have to be able to do it adaptively inside the processor. That’s what an NV FP4 is. When somebody says FP4 or FP8, it almost means nothing to us. Because it’s the tensor core structure in all the algorithms that makes it work. I would not be surprised that the industry would like us to make this format and this structure an industry standard in the future. This is completely revolutionary.”

Next, we will dive into Nvidia’s approach and compare this to AMD’s coming solution. Subsequently, we will dive into more topics in semis and software.

Read more

A key reason for the AI supply chain outperforming in ‘25 was token consumption coming in much stronger than expected, which resulted in cloud capex upgrades throughout the year. The chart below highlights how the bulk of capex is being driven by the hyperscalers and especially the major clouds, which is where enterprises run key infrastructure (alongside on-premise data centers).

Despite narratives of a bubble and circular financing, the bulk of investments is carried by hyperscalers who have some of the best balance sheets and cash flow generation in the world. While neoclouds are highly leveraged, these remain small in overall capex spend. However, most of their capex spend is backed by long term contracts and with high quality players such as hyperscalers. In addition, national governments are participating in building out dedicated AI infrastructure, while financing these by tax revenues and government bonds issuance. Again, these are not lower quality participants as they actually have the cash to participate.

If we retrieve quarterly accounts, we can see that the four hyperscalers are actually close to peak FCFs again at $60+ billion per quarter (black line below, axis on the right). Despite elevated levels of capex, the returns on those investments are obviously already healthy. And all these installed AI servers will generate revenues for 6 years, so the real rewards in FCF still have to come as we’re still in the investment phase.

Hyperscaler cash flows of operations have been growing at attractive CAGRs since the start of this investment cycle:

If we take into account rising capex via depreciation, EBIT has been growing at a 26% CAGR over the last five quarters:

And the outlook continues to be extremely good. Cloud revenue growth has already been massively accelerating throughout 2025, while the key leading indicator of cloud bookings accelerated again to 54% YoY growth last quarter:

Concerns around large capex spend have been weighing on hyperscaler share prices, especially Meta. However, the financial returns are clearly there already, while we continue to be confident in the demand picture. We expect AI demand growth to remain strong as inference applications get increasingly inserted in the real economy and model reasoning capabilities continue to increase. The latter is becoming increasingly token intensive, with the ratio of the number of tokens generated in reasoning relative to the number of tokens in the final output being on an upwards curve. We’re moving away from an old world, where there was a 1:1 relationship, to a reasoning world where models consume huge amounts of internal tokens to solve complex problems such as in coding, science, etc. At the same time, advancements in AI agents mean that AI models can now work independently for longer periods on their own.

Overall, we see the hyperscalers as building out extremely valuable infrastructure, where enterprises are effectively locked in at the major clouds due to high switching costs in software services and data gravity effects.

Another fallacy we regularly encounter is the so called ‘circular financing’ narrative, i.e. that this capex cycle is being driven by circular deals. While there is some amount of circular financing, most notably in some Nvidia-neocloud deals or in the OpenAI-AMD deal, these remain a very low proportion of overall capex spend. Circular financing doesn’t even make 5% of Nvidia revenues on our estimates.

The main risk for the AI hardware names is that forecasts at some stage will start showing moderate growth and with the risk of a cyclical pullback. Going over the financials above, we think that the economics are very healthy for the AI bull cycle to continue. Even when capex growth rates start to moderate later this decade, for the major clouds, share prices can actually do extremely well as free cash flow growth accelerates (as we’re post the investment cycle). At the same time, more moderate capex growth still translates into revenue growth for the hyperscalers, albeit at a slower pace. Another factor is that pricing power for cloud software services remains strong, which will further boost margins at the major clouds.

AI hardware names on the contrary need ever more data center capacity being added to show attractive growth rates. A tailwind is that as AI data center equipment is becoming ever more advanced, ASPs will automatically increase. So, AI hardware names can still get growing revenues even by selling the same volumes of equipment, but at higher ASPs. Key cyclical indicators to watch out for in the coming years will be token consumption growth in the cloud, GPU rental pricing and availability, and capex guidance at the hyperscalers. We’ve seen quite a number of semi booms and busts over the last few decades, and so overcapacity at some stage later this decade is a key risk factor to watch out for. Once you’re in the overcapacity scenario, it will likely mean a multi-year downturn in AI equipment sales.

Another risk is that while AI demand should continue to come in strongly, we simply run into bottlenecks in the AI value chain somewhere, which reduces growth rates for the overall industry. To some extent this has already been happening since the start of this boom. First, in CoWoS, then in HBM, and now in Optics.

Overall, we see the environment as very healthy for AI. We expect token consumption to keep growing at high rates while the hyperscalers are reaping the rewards to pay for increasing investments. This brings us to where these investments will be allocated..

Next, we will dive into the entire semis space as well as key topics in AI. As a bonus, we’ll discuss our outlook for the quantum computing space, and we’ll also review a US-based, state-of-the-art fab trading on a sizeable discount to intrinsic value. Finally, we’ll discuss our top stock picks for 2026 across tech.

The Battle for the Accelerator

Read more

“We revisit why high NA is likely to be adopted and update the latest evidence/data supporting its adoption: Why? 1) new data suggest high NA could lead to 20-40% cost saving vs double or even triple patterning critical layers; 2) We identify 3 specific critical layers for logic likely to adopt high NA and 2 for memory; 3) Advanced logic roadmap confirms that the below 20nm metal pitch at A10 node (tools in 2029-30, mass production 2031) will likely require triple patterning low NA if high NA is not adopted; When? 4) latest data suggests tools availability likely to reach >90% in ‘26 which we believe is likely to trigger high NA orders in H1/H2 26 for insertion for A14 2nd gen (tools in 2027-28, mass production 29).

Our analysis suggests that High NA could account for 15-20% of system sales by the end of the decade, driving a 3-5 percentage point increase in lithography intensity share due to cost reductions achieved in critical layers of other manufacturing processes, such as deposition and etching. As for catalysts, we view the SPIE industry event scheduled for February 22-26 as a key date to monitor progress toward achieving >90% readiness for High NA, which may pave the way for its adoption in 2027 for the A14 2nd/3rd generation node. Furthermore, we believe the introduction of High NA is significantly more impactful for ASML compared to low NA multi-patterning, as High NA results in a greater lithography share, whereas multi-patterning leads to a lower lithography share. This dynamic could meaningfully influence ASML’s valuation premium relative to its peers.

ASML is trading at a c20% premium relative to U.S. peers, compared to its 10y historical average of 86% with positioning near the all-time low range of 5-15%. Given High NA could add 3 to 5 percentage point WFE share relative to US peers we believe high NA is not only important for EPS but could trigger further upside to relative premium multiple to a more normalised level vs history of around 40-60%.”

The other reason we like ASML is that TSMC capex is still fairly depressed. TSMC built up a lot of N7 capacity during the covid semi boom and has been able to shift those tools to strengthen its N5 and N3 lines during the AI boom. However, this tailwind in terms of lower required capex should now be dying down, and with all capacity TSMC is planning, it looks likely that we’re looking at a bull case for semicap. Even modeling in moderate capex-to-revenues levels for the coming years, capex comes out much higher than what the market is expecting:

On Micron’s earnings call this week, it became clear that AI demand continues to strengthen:

“Server unit demand has strengthened significantly, and we now expect calendar 2025, server unit growth in the high teens percentage range, higher than our last earnings call outlook of 10%. We expect server demand strength to continue in 2026.

In the medium term, we are only able to meet about 50% to 2/3 of our demand from several key customers. So we remain extremely focused on trying to increase the supply here and making the necessary investments.”

Obviously this bodes well when it comes to litho demand from the memory market. Especially as the key driver of memory demand, HBM, is extremely capex intensive. This is again Micron’s CEO:

“The dramatic increase in HBM demand is further challenging the supply environment due to the 3 to 1 trade ratio with DDR5 and this trade ratio only increases with future generations of HBM. Additional cleanroom space is necessary to address this increased demand and lead times for cleanroom build-out are lengthening across geographies.”

We wouldn’t worry about China’s Manhattan project to produce EUV tools at this stage. Even if China produces the tools by the end of this decade, it seems unlikely that the US will allow any of the semi manufacturers amongst its allies to purchase that equipment; so that includes TSMC, Samsung, SK Hynix, and obviously Micron and Intel. Similarly, Western enterprises are focused on reducing Chinese components in their supply chains and this trend will only continue to strengthen with the recent Nexperia debacle.

So, even if Chinese EUV tools will see adoption by 2030 or so, which is a big if, we see the risk as fairly limited to ASML at this stage. ASML can’t sell their EUV tools into China anymore anyways, and so Chinese EUV tools will mainly be sold to SMIC and the Chinese memory players. Where it would impact ASML is on their Chinese immersion business, as the likes of SMIC have been purchasing plenty of those to do multiple patterning and still produce fairly advanced silicon of around N7 levels.

ASML’s valuation still remains fairly reasonable, while we should be looking at a bull case in demand in the coming years:

Next, we will dive into market share movements in XPUs vs GPUs, and Broadcom’s sell off.

XPUs vs GPUs

Read more

The beauty of coding for CPUs is that it is extremely simple, you can leverage languages such as C, Fortran, Rust, etc., and even make use of multi-threading to parallelize tasks over multiple CPU cores where possible. However, to move these workloads onto GPUs, this requires a complete algorithmic and codebase overhaul. GPUs demand massive thread-level parallelism with tens of thousands of threads, and a codebase that has to be rewritten for CUDA. This code has been built up over decades, and so rewriting it is an enormous task. This is precisely the reason why it has not been done before despite GPUs being much faster.

What will no doubt help here is that LLMs have gotten extremely good at coding, and so they can now take over large parts of this workload. Basically, we will have GPUs creating code to run on GPUs, creating a virtuous circle of GPU demand.

This is Synopsys’ CEO on how verification is currently a major compute bottleneck in chip design:

“We started redesigning some of our products about 7 years ago on NVIDIA GPU using the CUDA layer and in a number of cases, we’ve seen a significant speed up. When you talk about 10x, 15x, 20x speed up on a work that may take 2, 3 weeks and bring it down to hours; customers will adopt it. The bottleneck of designing these complex chips or complex systems is limited by the computation requirement to verify these systems. The worst thing you can do is assume that you’re ready to go, launch a chip and it doesn’t work as intended. That costs hundreds of millions of dollars. So, you spend a lot of energy in the design and simulation phase. We have a number of products already in use at customers, still very early stages in terms of a broad adoption, but that’s why we refer to an expanded opportunity for both companies.”

Jensen highlighted how EDA and CAD software will play a major role in digital copies of the physical world, i.e. digital twins:

“The expansion of the market opportunity for Synopsys goes from the EDA industry – addressing a several hundred billion dollar chip industry – to now addressing a multitrillion dollar, every product industry. In the future, every product will be designed in digital twins. This partnership essentially enables this industry to address the entire R&D budget of the whole world’s GDP. Everything that’s going to get designed and built will be done first in a digital twin.

In 2016, 90% of the world’s scientific supercomputers running physical and biological simulations were general purpose CPUs. Today, it has flipped completely, CPU-only supercomputers are 10%. NVIDIA-accelerated computing is now 90% of the world’s physical science simulation computers. This is going to happen also in the EDA industry. For industries that are fundamentally based on physics, this capability is really important.

The challenge, of course, is to reformulate the algorithms so that it could be accelerated on CUDA. And that is a multiyear journey. It took some 7 years probably to do cuLitho. Reformulating Synopsys’ applications to take advantage of this acceleration is what this is all about. All of it will be CUDA accelerated over time. And so we’ve got a lot of work to do, but that’s what this partnership is really about, focusing the two engineering teams across the entire companies so that we could take this capability to market as soon as possible.

The partnership is nonexclusive. There’s no obligations whatsoever for Synopsys to only buy NVIDIA. And they’re welcome to continue to work with their rich ecosystem of chip partners, and we’re going to continue to work with our ecosystem of really important EDA and CAD partners like Cadence, Siemens and Dassault. We also partner with Intel, all of NVIDIA’s EDA – the way we do chip design and system engineering today – is still largely based on x86 CPUs. This is really the beginning of that platform shift so that in the future, that’s going to be accelerated by NVIDIA GPUs. And I’d be delighted for all of the chip industry to buy NVIDIA GPUs for designing their chips.”

This is obviously a bit of a sideshow for Nvidia as GPU-based chip design will be a small end-market compared to AI. However, it is of key strategic importance to accelerate their own chip design cycle. Nvidia has recently moved to a one year cadence in releasing a new major chip, and no doubt removing the verification bottleneck will have a major positive impact once this work is complete.

Clearly this is an evolution that can have major impacts on the chip design industry overall. Most notably, it will bring down costs as weeks of chip simulation can be reduced to a single day or less. This opens the door for more chip experimentation and also more custom designs to be done, even by less experienced players who have the capital (e.g. large enterprises), as you can iterate on your chip designs quickly and remove bugs via verification.

The winners from this should be the EDA companies, Synopsys and Cadence, as lower chip design costs should boost demand for chip design software from more enterprises. Additionally, Synopsys and Cadence have solid pricing power. When they can provide a tool that speeds up work by 20x, they will have the ability to raise prices for GPU-accelerated modules. While Nvidia’s investment could give Synopsys an edge here, also Cadence is pursuing GPU accelerated workloads and has historically been able to build out a solid simulation portfolio.

Marvell’s New Data Center Tech

Marvell made a new acquisition to increase the company’s focus towards the data center and AI in particular. This is a strategy that has been working well, with the company’s revenues growing 41% year-over-year this quarter (excluding the recently divested automotive ethernet business to Infineon), and the data center end-market now comprising 74% of revenues. This is Marvell’s CEO on the tech and team they’ll be acquiring, which sounds pretty interesting:

“I’m excited to share details on the strategic acquisition we announced today of Celestial AI, which brings an entirely new disruptive technology, a photonic fabric platform purpose-built for next-generation scale-up interconnect. This acquisition is the latest in a series of decisive moves to further strengthen our data center portfolio. Since 2019, we have continued to increase our focus on the data center, divesting our WiFi business and acquiring Avera, Aquantia, Inphi and Innovium. These transactions have driven significant revenue growth and scale, and have each proven to be an absolute home run. The acquisition is expected to close in the first quarter of next year, subject to regulatory reviews in the United States.

AI is reshaping data center architecture at an unprecedented speed. Next-generation accelerated systems are no longer confined to a single rack, they are evolving into multi-rack scale-up fabrics that connect hundreds of XPUs in a high-bandwidth and ultra-low latency any-to-any fashion. These advanced fabrics demand purpose-built switches and interconnects, creating a new TAM for companies like Marvell. Industry analysts are forecasting the merchant portion of the scale-up switch market to approach $6 billion in revenue in 2030. On the interconnect side, as it attaches to both the XPU and the switch, the opportunity actually doubles to over $10 billion. These are both very large and incremental opportunities for Marvell.

As we first evaluated Celestial AI, it reminded us of our early look at Inphi and the products we saw in their PAM technology to transform the scale-out interconnect market. We see even greater potential for Celestial AI’s photonic fabric to transform scale up interconnect. Interconnect technology is critical to couple hundreds of XPUs together. Copper-based interconnects used in today’s scale-up systems are approaching their fundamental limits in reach and bandwidth, creating a compelling need for optical solutions. Celestial AI’s photonic fabric technology was purpose-built for this inflection. It enables large AI clusters that scale both within and across racks using a high bandwidth, low latency, low power and cost-effective optical fabric. This breakthrough enables a true optical solution with greater than 2x the power efficiency of copper interconnects, but with far longer reach and significantly higher bandwidth.

Celestial AI’s solution provides nanosecond class latency and the thermal stability is a significant competitive differentiator. It enables reliable operation in the extreme thermal environments created by large multi-kilowatt XPUs. This allows the photonics technology to be co-packaged vertically with high-power XPUs and switches in a 3D package, enabling the photonic connection to be made directly into the XPU rather than from the edge of the die. This stands in sharp contrast to many other CPO implementations where the photonics engine must connect at the die edge. Celestial AI’s approach results in a more compact and integrated solution, freeing up highly valuable die edge beachfront that can be repurposed to significantly increase the amount of HBM within the XPU package.

Celestial AI’s first-generation product is a photonic fabric chiplet (PF chiplet), which integrates all the required electrical and optical components. This delivers an unprecedented 16 terabits per second of bandwidth in a single chiplet, 10x the capacity of today’s state-of-the-art 1.6T ports used in scale-out applications. Celestial AI is deeply engaged with multiple hyperscalers and ecosystem partners and has already secured a major design win with one of the world’s largest hyperscalers who plans to use PF chiplets in its next-generation scale-up architecture. These PF chiplets will be co-packaged into both the hyperscaler’s custom XPUs and scale-up switches to provide connectivity.

This is just the beginning of a broad set of new applications which can be enabled from this technology. Our base case forecast shows Celestial AI’s revenue reaching a $500 million annualized run rate in the fourth quarter of fiscal 2028, doubling to a $1 billion run rate by the fourth quarter of fiscal 2029.”

Marvell’s fiscal ‘28 ends in January of ‘28, so significant revenues will start coming in two years from now (Q4 of calendar ‘27). The main goal of this new tech is to scale up (providing low latencies) across multiple racks. Traditionally, when you add racks to your compute, this is called scaling out (or horizontal scaling). However, Marvell-Celestial AI are calling this scale up here – a term traditionally used to refer to adding more compute into a single server or rack – as due to low latencies, they could make a cluster of 5-10 racks behave like one rack.

Current scale-up fabrics (like NVIDIA’s NVLink or Credo’s AECs) rely on copper, which confines a computing node to a single rack (e.g. 72 GPUs). To grow larger, you currently have to step down to slower scale-out networks such as Ethernet or InfiniBand, which forms a bottleneck especially in AI training today. During the Q&A, it became clear that Amazon AWS will be the first customer, and so this tech will first be implemented for Trainium-based AI training clusters. As shipments start in Q4 2027, this will be either for Trainium v3 or v4 based training clusters.

With this acquisition, Marvell is building out their data center connectivity portfolio with what looks like a very promising next-gen technology – high bandwidth, low latency, long reach, and thermal stability. Marvell has a current stronghold in DSP-based connectivity (800G and 1.6T pluggables), which are the go-to for hyperscalers to wire up AI clusters. The company acquired this expertise with the Inphi deal in ‘21, which was obviously a great investment and it is currently the key driver of growth.

So, the company has a fairly dominant market share in DSPs, but the main concern would be increasing competition from Broadcom which is launching competitive products. However, end-market strength is expected to offset headwinds from competitive pressures. This is HSBC on these dynamics:

“We believe Marvell is maintaining its market leadership in this segment through its 800G PAM4 DSP, which is seeing strong demand, and the launch of the subsequent 200 gig per lane 1.6T PAM4 solution, for which the company anticipates accelerating adoption over the next few quarters. However, we believe its market leadership may be threatened by Broadcom, which has caught up with the launch of its own 800G and 1.6T DSP solutions. As a result, while we do believe that the growth of its AI optical business, which is primarily driven by PAM4 DSP solutions, will decelerate — we believe the growth would still be higher than Street estimates due to the overall TAM expansion.

Amidst the massive AI infra demand expanding beyond traditional hyperscalers and into enterprise, sovereign AI, etc., the demand for DSP solutions continues to rise. Hence, while we expect AI optical business y-o-y growth to decelerate from 54% in FY26e to 38% in FY27e and 20% in FY28e, our FY27e and FY28e estimates remain above consensus estimates of 20% y-o-y growth in FY27e and 3% y-o-y contraction in FY28e. Our bullish estimates are based on our belief that while 800G will continue seeing strong momentum, the 1.6T solution will also see faster adoption driven by the TAM expansion.”

Next, we’ll dive further into Marvell. Will Marvell get disrupted? Is it time to take profits after the strong share price rally since September? We’ll provide much more analysis below..

Read more

Secondly, when it comes to running workloads in the cloud, Nadella described Azure as the “customer’s portal to the entire AI stack”. Microsoft’s strategy is to provide high margin web services in their cloud—like storage, databases, networking, security, and further apps—while they don’t need to own the raw GPU for every job. Nadella explains:

“Azure is the portal... For pure GPU workloads—like massive training runs or inference at scale—we can burst to neoclouds. It’s fine for us because now we have line of sight to demand. If they bring their capacity into our marketplace, that customer uses the neocloud for compute, but pulls storage, databases, all the rest from Azure. That’s a great win-win. They get specialized capacity fast, and we capture the full workflow.”

We see this as a sound strategy. Azure has a wide portfolio of web services in the cloud with its only competitors being Amazon AWS and Google Cloud Platform. At this stage, as future AI growth rates remain hard to predict, Microsoft can leverage capacity from a variety of neoclouds, control their capex expansion and remain free cash flow positive. The more attractive and sustainable AI workloads can selectively be powered by the Azure platform, with excess capacity being outsourced to neoclouds. In the future, more of those types of workloads can always be brought in-house again, with bursty AI demand being offloaded to neoclouds.

Thirdly, while OpenAI’s apps can be powered by other clouds, the APIs will be exclusively on Azure. This is an important point as the latter part of the business will likely be very lucrative – loads of other software and apps connect to OpenAI’s APIs to send their ChatGPT queries over JSON. If OpenAI’s consumer-facing app at some stage can disrupt Google Search, obviously it could be even more lucrative, but Nadella is happy for those workloads to also be powered by other clouds. The reason is likely that Microsoft doesn’t want to send their capex levels to stratospheric heights at this stage as future AI growth rates remain hard to predict. So, Nadella is striking a balance here, Azure can still accelerate growth and achieve highly attractive 30-40% growth rates, while the company focuses on selectively expanding capacity where there should be long term sustainable demand.

Fourthly, Nadella explicitly confirmed that Microsoft has full access to all of OpenAI’s IP, with one notable exception in consumer hardware i.e. any future OpenAI-branded devices. While no doubt Sam Altman will try to shield some of OpenAI’s innovations in algorithms to train and infer their models, Nadella implied here that Microsoft would have access to everything. This means it’s fairly likely that Microsoft will be able as well to build their own LLM capabilities in the long term, which Nadella referred to as MAI (Microsoft AI).

Finally, Nadella envisions their software products such as Office 365 and their Azure platform being utilized by AI agents. So, the business model should evolve from per-user licenses to per-agent provisioning which could balloon the TAM:

“Our business, which today is an end-user tools business, will become essentially an infrastructure business in support of agents doing work... It’s not per-user anymore, it’s per-agent.”

The only counter argument we’d make here is that a lot of these software products’ key functionality is also available via open-source competitors. While these open-source apps are typically lower quality, thanks to AI their code bases could be improved to increase functionality and reliability. Thus, while we can see a bull case for software as also agents might need to buy licenses or usage credits, there is also increased risk as writing code has become more commoditized. Doing coding on a regular basis, we estimate that one programmer’s productivity currently increases by 5 to 10x with the use of LLMs. And this will only continue to increase as now LLM agents are directly available in popular coding environments, so it’s not like you have to copy-paste code anymore from Claude’s web portal.

Subscribe now

For subscribers, we will dive further into developments in AI. We will discuss CoreWeave, as well as some accounting tricks the company is playing. Next, we will give the outlook for AMD and, finally, we will highlight a smaller, hardcore engineering firm in robotic AI that we came across. We’ll also give an overview of our current buy list in the universe of stocks that we track.

Read more

As the industry moved into this oligopoly structure, pricing discipline improved which resulted in better margins and EPS:

However, as the industry remained extremely cyclical, these stocks weren’t Buffett-style buy-and-hold investments, but rather tactical opportunities to play during an upcycle in the semis market. End-markets also weren’t that exciting. Revenue exposure to the mobile and client markets was high, two end markets that had gone ex-growth:

However, with the rise of AI, end market exposure for the memory players has dramatically shifted towards the data center. Last quarter, the two data center related segments comprised 54% of revenues for Micron. Besides being one of the most attractive growth markets in semis in the coming five-ten years, the data center market also has a number of other attractions. Firstly, customers are taking out multi-year contracts, which enhances revenue visibility and stability for the memory players. This is SK Hynix’s CFO on this week’s call:

“It is true that in the memory market, some businesses have shifted to order first, produce later approach with the emergence of HBM which have massive investment and long lead time. In addition, with strong HBM demand coming from our customers, the company was able to secure visibility into the customers’ demand from the contracting stage with long-term agreements. For both the memory industry and the company, this has led to greater market predictability and business stability than in the past when it was much more volatile.”

Subscribe now

One of the historical problems for the memory players was that memory was basically a commodity. You could swap out one vendor’s DRAM and replace it with that from another. This is now changing in the crucial HBM market where growth is exploding. SK Hynix’s CFO explains:

“Custom HBM will gradually increase from HBM4E. So products will be developed in close collaboration with customers from the early stage of design of GPU or ASIC products, unlike in the existing standardized HBM. This will lead to much longer term and strategic transactions between customers and a small number of suppliers, contributing even more to business stability and profitability improvement on the part of the memory suppliers.”

All of the above has led to better dynamics in traditional memory markets as well. Industry consolidation led to better pricing and investment discipline, and the growth in HBM is resulting in supply constrains and long term agreements in conventional memory markets:

“Now if I may add, memory companies are allocating capacity to ramp up HBM supply. And this has led to supply constraints in conventional memory, resulting in supply shortage of conventional memory, for which demand is actually growing. And as a result, we are seeing an increase in customers who want to sign long-term agreements for conventional memory products as well. Some customers are very much actively responding to the current supply shortage by issuing prepurchase POs for 2026. Now, given the customers’ demand and the company’s capacity for next year, not only HBM, but DRAM and NAND capacity has essentially been sold out.”

For premium subscribers, we will dive further into current developments in the memory market, we’ll do some interesting modelling on Nvidia, Broadcom and AMD, discuss the outlook from here, and finally, we’ll highlight a few trades which we will be making (selling some winners).

Read more

JP Morgan initiated on Credo, a niche silicon player in the AI data center:

“Credo is well positioned to capitalize on rising AEC demand, with industry unit shipments forecasted to expand at a 60%+ CAGR from <1 mn in 2022 to 10 mn+ in 2028, supported by increasing deployments from major companies such as Amazon, Microsoft, and xAI as well as broadening adoption by others, including Meta and potentially Oracle, Google, and more, which are currently engaging with the technology. Underlying the increasing and broadening adoption are expanding use cases beyond the traditional NIC-to-ToR connection inside a server rack, as AECs are now being used for short-reach rack-to-rack connections, intra-switch rack connections, and eventually scale-up use cases. These dynamics are creating an ASP tailwind, as early adoption of short-reach 100G “straight” cables by a single company ( ASPs in the low hundreds ) is transitioning to a variety of ranges, more complex form factors, and data rates extending to 800G for multiple companies ( ASPs in the mid-hundreds ), which, when combined with the shipment forecast, translates into a $4 bn+ market opportunity by 2028

Credo became the first supplier of AECs in 2018, creating both the product and the market, and remained the primary supplier for many years. Despite the more recent entry of well-known players such as Marvell, Broadcom, and particularly Astera – which has gained notable traction in the market recently – Credo still commands an overwhelming majority of the share, estimated to be ~80%. This leadership is driven by Credo’s ability to deliver an entire, qualified, and tested solution, where the silicon to finished cable is based on its in-house IP (vs. peers providing only the silicon or module), which gives Credo advantages in time-to-market, ease of deployment, reliability, and cost. While we expect other peers to gain market share over time as large customers multi-source, we anticipate that Credo will maintain its overall leadership, with other suppliers primarily serving as second or third suppliers.

Beyond Ethernet AECs, Credo’s product portfolio also encompasses other interconnect technologies, such as optical DSPs and PCIe retimers. While Credo currently has limited share and will be competing against established peers, these markets represent significant opportunities – estimated to track to $6 bn and $2 bn in 2028, respectively – and even minor share gains could translate into a solid growth opportunity for the company. Importantly, the aforementioned implies an aggregate interconnect TAM expanding at a +45% CAGR and reaching more than $12 bn in 2028 .

We are forecasting a revenue CAGR of 50%+ from FY25 to FY28E, including growth of 100%+ in FY26E. This forecast is underpinned by increased deployments of AECs by Credo’s two large, more longstanding customers, new deployments across three additional customers, as well as strong growth for Optical DSPs and solid growth currently assumed for other solutions ( both from a low base ). Importantly, our expectations are ahead of industry forecasts for AEC unit shipments, which are projected to grow at a 40%+ CAGR over the same period. We believe this outperformance reflects new use cases beyond NIC-to-ToR (e.g., rack-to-rack, switch rack), the ramp-up of new customers, and pricing tailwinds.

We are projecting earnings growth to outpace revenue, with an earnings CAGR of 70%+ from FY25 to FY28E, driven by gross margins tracking toward the high end of the long- term range (63%-65%), supported by increasing scale and an improving product mix as well as operating expense growth at a pace below revenue. These factors should also drive a solid improvement in free cash flow generation, which, against the backdrop of a very healthy balance sheet, could lead to a greater appetite for capital allocation that has been limited to date and would represent upside to our model. We introduce a December 2026 price target of $165 based on valuing shares of Credo at ~48x our CY27 EPS estimate.”

The below image shows what Credo does – the purple cables are Active Electrical Cables (AECs). These are low-power, copper-based interconnects that integrate advanced silicon to extend signal reach. These are heavily used in AI data center networking for shorter distances with the most common use cases being inter-rack and top-of-rack communication.

The attraction with Credo is an 80%+ market share in the AEC market where volume growth is even higher than in the accelerator market. While competitors such as Marvell, Broadcom and Astera are moving in, Credo has further growth opportunities in other interconnect technologies.

Subscribe now

There are not many companies which can effectively compete at the leading edge as semi design costs continue to explode. So, typically, there are a handful of competitors in these markets, or less. This makes these industries and companies attractive for investors to play the exploding growth in leading edge silicon, as these are not commoditized businesses. This can easily be witnessed in attractive gross margins. You can see that both Astera and Credo are generating similar gross margins to Nvidia:

The attraction with both TSMC and ASML is that these upstream companies in the value chain currently face no or only very limited competition, which gives them clear upside to keep moving gross margins up over time. We expect that TSMC will be moving its gross margins up to 65-70% long term, and similarly, it’s likely that ASML will move towards a 60% gross margin as well. Obviously, Astera and Credo have a handful of competitors, which could give some gross margin pressure over time, especially when the pace of growth slows down at some stage. That said, we think these latter businesses are definitely investment grade as well in the context of a diversified portfolio.

For investors in the name, we’d simply stay long as we continue to be bullish on the AI cycle, while Credo is obviously fairly well positioned.

It’s a punchy multiple, but then again EPS are exploding:

For premium subscribers, we’ll dive further into developments in AI with a number of stock picks. Finally, we’ll take a detailed look at a very small but highly innovative company that should be looking at a long runway of attractive growth.

Read more