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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

Read more

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

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

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

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

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

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

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

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

Read more

Traditionally, publishers such as the Wall Street Journal sold advertising space by closing large deals with ad agencies. The agency was the key player in the middle, connecting publishers to brands. This intermediary had the knowledge of how to run advertising campaigns while as they purchased ad inventory in bulk, they also typically could offer brands a better deal.

In the internet era, this value chain started to fragment considerably and a wide variety of new players emerged. First, there are the ad networks, which dominate large swaths of the online advertising market. The two dominant ad networks, Google and Meta, are amongst the largest companies in the world. Basically, on these ad networks, advertisers and brands can buy inventory themselves to run ad campaigns.

Secondly, there is the open internet, where publishers open up their ad inventory for real-time auctions with each impression being sold to the best bidder. Both price and quality of the advertiser can be taken into account here. For example, Netflix will prefer to show an ad from Nike as opposed to one from OnlyFans. To run this auction process, key players in the middle are the supply side platforms (SSPs) and the demand side platforms (DSPs).

So, when a user visits a website or smartphone app, the publisher’s server will detect that an ad impression is available. The server then sends a request to the publisher’s SSP which collects information about the consumer. Data collected can include age, gender, interests, browsing history, browsing behavior, the smartphone device in use, typical engagement metrics, and geolocation. The richer and more precise the dataset on the user, the more valuable the ad impression becomes.

The SSP forwards this impression to the DSPs, which evaluate the impression to place their bids. To more accurately value this impression, the DSP can make use of additional data it has on the end user, which may have been collected by the DSP, the brand itself, or a third-party data supplier. Data collection on internet users is a fairly complicated industry, with tons of continuously evolving regulations.

The Trade Desk (TTD) is the dominant DSP as it has both the widest and best quality ad inventory available, some of the richest data sets on consumers, the best platform capabilities, and as it is an independent and neutral arbiter in the middle. This final point means that there is no conflict of interest between the advertiser and the DSP. Unlike Amazon’s or Google’s DSPs, TTD doesn’t own ad inventory itself, so it won’t try to unfairly promote its own impressions to advertisers. The sole aim of TTD is to identify the best available impressions for its customers.

Typically, advertisers make use of Amazon’s DSP when they want to run a campaign on amazon.com or Prime Video, while Google’s DSP is mainly leveraged to run campaigns on Youtube. However, the aim of large brands is to re-target potential end customers across a variety of platforms and with different ads. So they might show you one ad on TikTok, another on Spotify, the Financial Times, X, LinkedIn, Netflix, etc. To realize such a multi-platform campaign you need TTD.

An additional advantage is that a good DSP keeps track of which ads have already been shown to each end user. Showing a potential customer the same ad for the twentieth time is just wasted money. So, a clever DSP can create substantial cost savings for customers as it can figure out when to change ads to re-target a customer, or on the other hand forego the impression altogether.

This is a former manager at TTD explaining the company’s strong position in the DSP market (via Tegus):

“The Trade Desk’s value proposition is largely based on being independent as a neutral platform in the middle, while having access to large swaths of inventory across pretty much every streaming player within the connected TV (CTV) space, audio streaming, ads on mobile apps, and every publisher within the programmatic space. Then, The Trade Desk has all sorts of connections to make it easy for the client to bring in their own and third-party data, while also additional data from a publisher itself can be integrated directly. For example, Disney might be willing to share some insights about their audience with The Trade Desk which can be tied to the available impressions.

In CTV, advertisers have been migrating their programmatic buying over to The Trade Desk. Historically, Roku was someone that you would have to work with directly and Netflix was in the same camp, but both are now opening up. And you can also buy Hulu inventory over The Trade Desk. There’s a lot of opportunity in CTV.

Audio is another channel that is often forgotten, and players like Spotify and Pandora are integrated with the Trade Desk as well. Audio pales still in comparison with media spend on Connected Television, but it is growing because the audio platforms have authenticated and logged-in users, and having those types of identifiers is helpful in digital advertising.

The average client on The Trade Desk typically has an annual digital media budget upwards of $10 to $12 million, with some that are spending more than several hundred million a year. Probably more than half of these budgets on The Trade Desk are agency driven, both from major holdcos and smaller specialized agencies, but there are an increasing number of clients that are working with The Trade Desk directly. These have their own in-house teams of programmatic trading desks, but this is only for large companies because it’s very costly.

The Trade Desk is now introducing AI to automate the setup of campaigns, effectively taking over that manual work that’s placed on agencies or in-house trading teams. This creates an opportunity for smaller advertisers to work with The Trade Desk directly, as you don’t have to rely anymore on sophisticated programmatic buyers to set up and manage your campaigns. You can now do this by pressing a few buttons and then smaller advertisers can also get into additional channels like connected television. Minimizing the burden of programmatic buying will allow more advertisers to move away from the ad networks of Google and Facebook—where you can activate a campaign with a couple of buttons and a credit card—to a platform like The Trade Desk.”

To further strengthen its strong position in the programmatic value chain, TTD has been building direct connections with the large publishers themselves. The attraction for publishers is that most brands are running campaigns on TTD anyways and so now they can forego having to pay the SSP its take rate, which can be substantial. This new feature is what TTD calls OpenPath, and the attraction for TTD is that they can offer the widest and best quality inventory from the open internet on their platform. For example, one smaller agency mentioned in a Tegus interview that it was impossible for them to obtain ad impressions for the Olympics on NBC, however, these easily became available over TTD.

Traditionally, we had the below value chain for digital ads on the open internet. An advertiser uses a DSP to connect to a SSP and then bid on an impression from a publisher:

However, with OpenPath, TTD is putting itself at the center to connect advertisers and publishers on the open internet:

Over the past 1.5 years, TTD shares sold off heavily due to a variety of factors. We can see that EPS performance actually still has been positive, with negative share price momentum being driven by multiple compression. The shares are now trading on a 10x forward PE:

Is TTD a buying opportunity or is the story structurally broken? We’ll go through a variety of interviews with industry insiders as well as insights from sell side analysts to evaluate..

The Trade Desk Disrupted?

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

The DRAM market has always been extremely cyclical. To make matters worse, this also used to be a highly competitive market resulting into terrible profitability in the 2000s. Despite the boom in gigabytes of DRAM shipped over time, dollar revenues in 2014 were still the same as during the DRAM supercycle of the early 90s:

Moore’s Law was a big factor here—basically the dollar per GB dropped by a factor of 10 every 5 years:

The other factor was a highly competitive market. Too many competitors vying for market share in a commodity-like market resulted in aggressive pricing and low margins. This situation improved dramatically in the early 2010s with the three player oligopoly—that we still have today—finally emerging:

In those days, the memory names were still dealing with a large stigma from investors. Around this time, we actually started arguing that DRAM names had become investment grade due to this three player oligopoly combined with a long term attractive demand outlook, which turned out to be reasonable as this is when Micron’s EPS started taking off. However, the industry remained highly cyclical and so investors had to time when to enter and hold these names:

Improved EPS were driven by more disciplined pricing, which resulted in more attractive margins compared to the 2000s:

As we know that the memory industry has historically been very cyclical, and with share prices having exploded over the last few years, a rational investor might argue that it’s time to lock in profits now, wait for the inevitable multi-year downturn, and then retake a position once we start seeing new green shoots emerge.

The counterargument is that we’re still in the early stages of an AI supercycle. The annual data center GPU market is still measured in the millions of units, whereas one can start thinking about future AI capacity measured in the number of GPUs per knowledge worker. I.e. each knowledge worker will need x agents, and those agents will need their own agents etc. Software engineering is at the forefront of this trend—despite the marginal cost of writing 500-1000 lines of code effectively having dropped to zero, the demand for software engineers continues to increase.

Thus, we’re moving into a scenario where agents are writing out large parts of the code base, and then skilled engineers are needed to instruct the agents how to write the code—i.e. “use this database with this database schema”, “use this type of data structure” etc.—and to then review the code to make sure it’s robust.

There are one billion knowledge workers on the planet, which means that the future AI data center market will very likely become much larger compared to today. It’s well possible we’ll move to an annual data center GPU market measured in the hundreds of millions of GPUs vs the ten million we’ll be doing this year. And then if AI can successfully get inserted at the edge, i.e. robotics and automation, we’re also going to need a lot of GPUs for edge compute, both for training and inference.

Currently, the market is pricing in that we’re at the peak of the investment cycle for key AI names. For example, looking at forward PE ratios—Nvidia is at 22x and Micron is at 10x. Also the sell side is modeling in moderate growth as from ‘27-28:

Thus, both the buy side and sell side are saying “we’ll be peaking soon in terms of AI investment”. This is by looking at market valuations and estimates. However, even the DRAM supercycle in the early ‘90s lasted four years, while it looks very likely that the current AI supercycle will far surpass that:

So, this is why we think that Jensen’s long term forecast of $3-4 trillion of AI data center capex could actually be realistic.

James Schneider

Jensen, you’ve previously outlined the potential to get to $3 trillion to $4 trillion of data center CapEx by 2030, which implies a potential acceleration in growth rates, which you’ve sort of guided to at least this next quarter. Do you still feel good about that $3 trillion to $4 trillion envelope?

Jen-Hsun Huang

Yes. Let’s back that up and just reason through it from a few different ways. So token generation is at the center of almost everything that relates to software in the future and relates to computing. If you look at the way we use computing in the past, the amount of computation demand for software in the past is a tiny fraction of what is necessary in the future. And AI is going to only get better from here. And so if you think about it, well, the world was investing about $300 billion or $400 billion a year in classical computing. And now AI is here and the amount of computation necessary is 1,000 times higher than the way we used to do computing. So the amount of token generation capability that the world needs is a lot more than $700 billion. We’re going to continue to invest in compute capacity from this point out. Just as a computer has a lot more computation capability than a DVD recorder, artificial intelligence needs a lot more computing capability than the way we used to do software in the past. Now from an industrial level, because all of our companies in the final analysis are powered by software, and the cloud companies are powered by software, and if the new software requires tokens to be generated, and the tokens are monetized, then it stands to reason that their data center build-out directly drives their revenues. And so compute drives revenues. And I think people are increasingly starting to understand that as well.

So, this is contrarian call, but we believe that the market is still far too skeptical on long term AI demand and required AI investment. Token consumption is currently exploding, and this will likely continue for the foreseeable future, and possibly well beyond.

At the same time, DRAM and HBM requirements per GPU are increasing dramatically. We can see that HBM capacity for Rubin Ultra will go up more than 5x vs the current Nvidia Blackwell:

On the supply side, due to the three player oligopoly in DRAM and HBM, capacity additions look to be fairly disciplined, reducing the risk of oversupply. For example, Micron has regularly been telegraphing on their conference calls that they do not seek to gain market share in HBM beyond their current market share in the DRAM market. So, we can see that in HBM, all three players are gradually adding capacity, with not any player looking to dramatically upset the balance in the market:

And as the DRAM industry is focused on building out HBM capacity, which remains the key bottleneck when it comes to memory and AI, DRAM capacity expansion will be much more limited:

This is Micron’s CFO at the recent Wolfe conference explaining that node transitions won’t be sufficient to meet demand, and that the trade ratio of HBM vs DRAM in terms of needed wafers will continue to go up:

“Traditional node transition is still the most efficient way for us to bring on supply, but those node transitions are less efficient than it used to be. There’s simply not enough ability to produce enough supply out of a node transition. Adding to the supply challenge is the increase in high-bandwidth memory, which is more silicon intensive. And that trade ratio, 3:1 historically for HBM3, that is increasing over time as we go to HBM4, 4E and 5. So, that trade ratio expands and further pressures supply. With that backdrop, we see supply-demand tightness beyond ‘26. We’re committed to our investment plans to invest in a disciplined way to supply customers what they need over the long term.”

We don’t believe that investors should worry about Chinese capacity for two reasons. One, Western hyperscalers won’t source Chinese DRAM anyways for obvious geopolitical reasons. But, secondly, because China doesn’t have EUV (from ASML), as well as advanced immersion tools (from ASML again), their DRAM won’t be competitive anyways with those of the big three. So, we think that Chinese DRAM suppliers will be purely focused on supplying the local market.

Where China could be a threat is if all the Chinese smartphone makers start sourcing local DRAM, this would basically take a chunk out of the big three DRAM manufacturers’ historical customer base. However, the data center market will be the key driver for Memory stocks, and Chinese competition won’t be able to make a dent here.

While all three players are now constructing new fab shells, should demand disappoint at some stage, filling these shells with semicap equipment can be halted. You don’t have to fully fill a fab if demand isn’t there and you can limit the number of manufacturing lines. This limits risk during a cyclical downturn as capex can be reduced again. Overall, we see the profitability outlook in DRAM and HBM as still being attractive with both markets likely to remain supply constrained for the foreseeable future.

Then, the next question becomes which stocks to pick? While we’ve been invested in both SK hynix and Micron, SK hynix remains our top pick. The reason is simple—SK hynix has been executing extremely well as the leader in HBM and they will retain a dominant share with Nvidia Rubin, from UBS:

“We forecast HBM shipment to reach 19.0bn Gb for 2026, up 51% YoY, and for 2027 26.3bn Gb, +38% YoY. This leaves SK Hynix with a bit market share for HBM of 53% in ‘26 and 49% in ‘27. We continue to forecast SK Hynix to maintain close to 70% share for HBM4 for Nvidia’s Rubin in 2026, and first source status for Google’s TPUs. In general, we see all major AI accelerator vendors trying to move by ‘27 to a 3-supplier status. This explains some of the slight market share erosion we forecast, whilst management will continue to aim for 50% share. On the other hand, it may open market share opportunity at AMD and Meta, where SK Hynix is currently not supplying.”

To be clear, we think all three DRAM/HBM names are fine. As from this year, Samsung Electronics’s profits will become heavily dominated by Memory as well, so the company is becoming less of a conglomerate and more of a DRAM/HBM pure play like the other two.

Subscribe now

However, as the company clearly executed less well in HBM and in their leading edge foundry—where they’ve fallen far behind TSMC in terms of orders and capabilities, with this division actually making losses—we continue to prefer SK hynix. Although also Samsung’s foundry story is improving, Jefferies notes that the business should become profitable as from 2027 driven by AI demand:

“In 2027, SEC’s Taylor plant is expected to begin mass production of AI5 (50% vender share) and AI6 (100% share) chips, which are the “brain” chips in Tesla’s Optimus and FSD. Tesla should account for 3% of SEC’s Foundry revenue in 2027, 20% in 2029 and 30% in 2031. Foundry should see order receipts jump over 100% YoY, while utilization rates improve on increasing orders from AI data centers and HPC firms, and growing AI chip shipments to Tesla. As a result, Foundry should turn around from a KRW7tn loss in 2025 to a profit in 2027.”

There have been substantial concerns in the market that Micron dropped the ball with HBM4, with their design choices potentially resulting into insufficient pin speeds. However, Micron’s CFO dispelled those concerns at the recent Wolfe conference:

“Since our last earnings call, our business outlook has strengthened further. Demand is significantly higher than our ability to supply and the industry’s ability to supply. We continue to expect supply/demand to be tight beyond ‘26. We are doing everything we can to plan and invest appropriately for our customers’ needs. We’re making good progress on multiyear agreements with specific commitments. We are preparing clean room space that will come online over time, and that will give us the ability to grow our bit supply in line with market demand.

And let me address some recent inaccurate reporting by some on our HBM4 position. We have been in high-volume production on HBM4. We’ve commenced customer shipments of HBM4, and we see shipment volumes ramping successfully this calendar Q1. This is a quarter earlier than we mentioned during our December earnings call. Our HBM capacity is ramping well, and we have sold out our calendar year ‘26 HBM supply as we highlighted a few months ago. Our HBM4 yield is on track. Our HBM4 product delivers over 11 gigabits per second speeds, and we’re highly confident in our HBM4 product performance, quality and reliability.”

So, the market is now pricing in a cyclical peak given the low valuations these key AI names are now trading on—22x PE for NVDA and TSMC, and even 6x for SK Hynix:

Actually, these stocks have been climbing a wall of worries over the last three years. Consistent fears of overspending and an AI bubble forming have weighed on valuations for these names for the last three years really. All share price performance basically has been driven by EPS growth:

Given how our own token consumption continues to dramatically move higher, as well as enterprise token consumption in the cloud, we continue to be bullish on the growth in AI applications over the coming 3-5 years.

Tech, Semis, and AI are currently exploding in China. Wouldn’t it be cool to listen to all those tech podcasts by Chinese LLM builders—such as DeepSeek, Kimi and Qwen—and understand the latest innovations?

Our friends at Lang Dojo have built a great course to learn Chinese fast. We’re impressed by how fast we’ve learned to read Chinese characters and understand spoken Chinese after studying the course.

If you like to learn Mandarin Chinese, check out the course on the link below, which includes free lessons and video lessons:

Software Outlook & Stock Picks

We outlined a few weeks ago our bull case for the software industry and recently HSBC published their deep dive covering the intrinsic characteristics of software, which many investors are not familiar with. Their reasoning and arguments are similar to the piece we published, and we highlight the key points of their analysis below:

“The software sector is likely to be a net beneficiary of the AI revolution. Market concerns that AI will replace enterprise software are misplaced in our view. LLMs and vibe-coding agents have been trained on public, not private, data and thus, don’t know the optimized large scale high-fidelity architectures. Thus, the burden of developing a competitive platform would be left to the coders who don’t know the private innerworkings of these complex mega large scale platforms. Even if you have great code, you need to demonstrate years of five 9s+ of uptime; error-free operations, and robustness across a highly diverse set of customer IT environments. Competing effectively in the mission critical enterprise software sector is difficult as one needs an enterprise class sales force, trusted brand, and large referenceable clients to gain trust.

We see AI as the primary source of value creation of the software stack, with the largest share of long-term value accruing to software rather than hardware. This is similar to the evolution of the internet; the early attention was on deploying physical infrastructure, such as PCs, fiber, and servers, but once that foundation was established, the real value shifted to the software sector. In the same way, we see the current AI narrative dominated by hardware build-out, but we expect this phase to be transitory. We believe the value of AI lies in how it is embedded into software applications that reshape workflows, decision making, and productivity. We believe the AI ecosystem will follow a similar trajectory.

We believe agents will be the primary monetization layer for AI consumption. Rather than generic chat interfaces, enterprises will deploy task-specific, workflow-embedded agents designed to provide a controllable and governable framework for AI adoption. Notably, this allows enterprises to ringfence usage and manage risks where hallucination can be handled programmatically in well-defined parameters. We expect inference demand to exceed training demand over time, driving sustained growth in compute requirements. Overall, we believe that as agentic AI becomes pervasive, usage intensity increases, driving token consumption.

Is there a widespread misunderstanding of AI and software? We have been talking about the topics in the software space for over two years, and nothing new in these topics causes us to deviate from our previous conclusions. And these conclusions align with Nvidia’s Jensen Huang’s recent comment that AI replacing software “is illogical”. However, the negative AI versus software headlines continue and, to us, there seems to be a herd mentality behavior developing around the indiscriminate comments regarding the demise of the enterprise software sector. Thus, we wonder if there is somewhat of an echo-chamber forming in the market that is amplifying Mr Huang’s “illogical” scenario? Although we are not certain about the behavior, we do know that there are many media firms and tech companies that make up the market information channel, and each will look for a megaphone to bolster their own worldview. Thus, it is important to keep in mind that many of these pundits put forth a view that places their organization in a favorable light or improves readership and may be self- serving. And their “marketing” may not fully represent an accurate reflection of reality, and a calm assessment of the facts is always recommended.

Will AI write code that will replace legacy software applications? The technology marketplace is an odd thing where the best or cheapest products don’t always win. Software is a business, and there are many success factors that influence how a software business grows, prospers, or fails. And it is important to recognize that enterprise class software is different from simple apps or smaller-scale narrow-focused software programs. Enterprise class software platforms are some of the most complex and have different success factors that determine outcomes within the marketplace. A simple example of a success factor would be first-to-market, as once a large platform application is installed and a customer’s business runs its critical operations on it, switching carries many risks.

One simple, and obvious risk is if there is disruption during a platform replacement, normal operations can stall, loyal customers can leave forever, brands are tarnished, and revenue can hard-stop until the issues are resolved. And a full-stop to a company’s revenue stream would be bad, particularly for publicly traded companies. There are also many other unforeseen switching risks/costs, like lower productivity and numerous process errors as users retrain and become acclimated to a new system. Or the sudden appearance of unwanted feature interactions that only manifest under long-duration at-scale operations in a customer’s unique mega-cluster of other applications (both internally and externally developed). Think of this as a complex system failure where many rare events converge and result in a system shutdown or in unwanted behavior, like errors. These types of failures are sometimes as tricky to understand as they are to resolve

Furthermore, promised incremental value-add is not always realized and less efficient operations after replatforming are a possibility (i.e., theoretical versus actual). Therefore, in our view, the promise of incremental benefits from replatforming are rarely sufficient to warrant a change. Hence, once a vendor platform is installed, persistent vendor lock-in is a common outcome. And looking at the competitive history between SAP and Oracle on the ERP front is a reminder of the difficulty in replacing platforms and how replatforming can fail. A high-profile example that even cheaper/free solutions may not be enough to displace an incumbent vendor is Microsoft’s Office suite. Competitors like Google and Apache OpenOffice have not materially displaced Microsoft’s Office Suite in enterprise, despite offering free office-like suites for decades. There have also been concerted efforts in the past to displace Oracle’s high performance database business by Microsoft, Amazon, SAP, and CRM, with minimal success.

And there are many other technology examples like VHS versus Beta or NTSC versus PAL that have all evolved with inferior technology coming out on top due to numerous other factors. Thus, just because AI can write code does not mean that that code will be competitive or widely adopted, even if it was given away for free. India-based companies have had the ability to create and market enterprise class software for decades at scale. Yet over the decades, even with this insight and a massive low-cost skilled workforce, no regional vendors have successfully emerged to challenge the legacy US vendors. Factors such as enterprise class sales teams, technology cross-licensing agreements, proprietary and patented IP, industry specific domain expertise, aligned workflows with industry practice, first-to-market, brand awareness, scale, or effective go-to-market strategies are just the tip of the iceberg when looking into the key attributes needed to compete effectively in the software sector. Just making code is not enough.

Moreover, wishful AI competitors are using models that were trained with mostly publicly available data on the internet. And although snippets of code are available on the web for AI to learn from, large complex enterprise class system code and proprietary IP are private and are largely not available for training foundation models. Deeply complex code that incorporates business logic, proprietary IP, and private client data are viewed as prized assets by legacy vendors and are typically vigorously protected. And between state-sponsored industrial IP theft, employees moving from competitor to competitor over their careers, and malicious hackers, key aspects of platform designs may be as protected as the formula for Coca Cola.

The actual code generation is only a small part of what it takes to be a successful competitor. In other words, you would need to become an enterprise class software company, with all of its complex business skills and IP. And the key success factors necessary to succeed and compete effectively are grown over time. Meanwhile, there are existing companies with all of the attributes, private information, scale, and vibe coding skills needed to leverage AI’s potential. That is, the existing enterprise class software companies themselves, the same vertical that market conjecture is questioning its staying power. These platform players don’t need time to mature and grow into becoming an enterprise class software company, they are enterprise class software companies now, and they have been aggressively working to leverage and embed AI into their global platforms.

Moreover, we think that most of the legacy platform vendors have spent much of the past two years designing and training intelligent agents. We believe they have been creating these expert agents through a domain-specific fine-tuning process using both private internal and client data. And they have just begun public releases of initial AI platform updates that include embedded agents. Therefore, through this process, we see the enterprise platform vendors as the dominant diffusion mechanism for leveraging AI in the world’s largest enterprises. Yes, some companies will build less complex internal apps and custom software for some tasks, but for the complex infrastructure that creates and runs the world’s commerce, we think that remains captive to the platform players. And we think that many of our Buy-rated software companies will, in fact, thrive to new levels.

The software model and the price vendors charge is highly correlated and dependent on the value proposition of their offerings. This is why a large price can be commanded for software, even with a minimal marginal cost of production. Thus, in dominating AI’s primary diffusion path (i.e., the continuing infrastructure providers of the Global 2,000), previously unthinkable TAM expansion for the enterprise software vertical is happening, in our view. We see this as a rapid and significant expansion in TAM and the software vendors responsible for unlocking this value will have claim to its share. And just as software unlocked most of the value in the internet, enterprise software is set to unlock most of value in the AI ecosystem, in our view.”

Next, we will dive into stocks that we see as AI beneficiaries in software. The space has been engulfed in panic selling over the last months, which usually opens up tremendous buying opportunities for long term investors. While there will be no quick fix for software names, we do believe that software companies that can show revenue acceleration and margin expansion over the coming years, that this will lead to strong share price outperformance given the current cheap valuations.

Basically, software names that can show that they’re AI winners by embedding AI agents and automation inside their platform, and translate this token consumption into revenue monetization, the market should reward those names with higher multiples. This is also what happened with Google. After the successful launch of the latest Gemini model, combined with their progress on the TPU side, this led the market to rethink its view that Google will be an AI loser, and the market then decided to reward Google with a much higher multiple. Combined with growing EPS, this led to strong outperformance with the shares doubling in less than a year.

Read more

The three major clouds—Amazon AWS, Microsoft Azure, and Google Cloud Platform (GCP)—have been some of the most straightforward investment cases over the last decade. The breadth of web services and apps available on their platforms, their vast network of data centers around the world, and enterprise IT workloads being effectively locked in on these platforms due to high switching costs have resulted in some of the most formidable moats we’ve ever come across. As enterprises shifted workloads from capex-intensive on-premise data centers to on-demand consumption in the public cloud, it was easy for the big three clouds to capture that enormous growth:

With the explosion in token consumption—as AI finds ever more applications in the real economy—cloud growth rates are accelerating again. The major clouds are the logical place for enterprises to deploy their AI workloads, as this is where they run their apps and store their data. Moving data around is costly (data egress fees) and also time intensive. Due to the plethora of data centers the major clouds have available in each region, enterprises can create systems with ample backup compute and data. Basically, if you design your workflows well, you can have 100% uptime on the major clouds. Simply create a global load balancer with behind it VMs in multiple regions and then also a multi-regional database. All this can be easily created on each of these major clouds.

This is Satya explaining how AI consumption drives increased demand for all of Azure’s services:

“The number of customers spending $1 million plus per quarter on Foundry grew nearly 80%, driven by strong growth in every industry, and over 250 customers are on track to process over 1 trillion tokens on Foundry this year. And of course, Foundry remains a powerful on-ramp for the entire cloud. The vast majority of Foundry customers use additional Azure solutions like developer services, app services, and databases as they scale.”

Satya illustrated well how easy it is for the big clouds to launch massive new apps and cross-sell these to their customer base. Basically, Microsoft built a Databricks-clone that is now doing close to $2 billion in revenues already. Obviously, this business scaled much more quickly than either Databricks, Snowflake, or Palantir in their initial years, it’s not even close:

“Two years since it became broadly available, Fabric’s annual revenue run rate is now over $2 billion with over 31,000 customers, and it continues to be the fastest-growing analytics platform on the market with revenue up 60% year-over-year.”

On the other hand, neoclouds are mainly focused on pure GPU workloads with much less customer lock-in characteristics. They typically have a few data centers while being heavily geared with loads of debt to build these. So, we think that the quality of the business is much lower, combined with a highly risky financial profile due to a large current debt load already (e.g. CoreWeave) or a coming high debt load in the coming years as GPU capacity gets built out (e.g. Nebius). Neoclouds are also heavily reliant on megadeals with the big clouds, or other hyperscalers, which may or may not get renewed 4 years from now, further increasing the risk profile.

The financial profile is compelling in the major clouds on the other hand. Major clouds can finance their growth from highly attractive operating cash-flows. This makes these three assets—AWS, Azure and GCP—some of the best quality plays on the long term growth in AI consumption. We see Amazon as the best play on this growth story, due to the high contribution from AWS to Amazon group profits (already close to 60%), and this ratio will continue to increase as AWS will be the main growth driver in profits going forward.

A potential risk we see for AWS is that the company is heavily focused on its custom, Trainium-based AI stack whereas Nvidia is obviously innovating extremely rapidly in this space. So, if we were running AWS, we’d probably rely more heavily on Nvidia to drive AI revenues from AWS’ wide enterprise customer base. However, from the recent call, it seems that the Trainium-based model is working well and that AWS has all the necessary intermediary software layers ready. This means that LLMs can easily be deployed on their custom hardware and then those can be made available to customers over API.

Being cloud customers ourselves (heavily GCP focused), we don’t really care which underlying silicon is being used. As long as the price per million tokens is attractive enough and the speed with which those tokens arrive is sufficient, we don’t really care about anything else for inference as long as the model can do the job. We suspect that most enterprises will similarly look for these two criteria.

A weakness is that Amazon doesn’t have the most advanced models of ChatGPT or Gemini available, and so the best models of those families being exclusively available on Azure and GCP respectively will be an attractive feature for those clouds. However, Anthropic’s Claude currently leads in coding and is a very capable model overall, so Claude will likely be a key AI demand driver on AWS. Similarly, AWS also has the open-source models available such as DeepSeek, Qwen, Llama, etc. Overall, we do think that GCP is best positioned here with the exclusivity of Gemini being highly attractive, with at the same time also having Claude available, but GCP is only a small part of Google’s conglomerate (11% of EBIT last year). If we could access GCP as a pure play, we would probably have this one as our top pick.

That said, we do think that AWS has a sufficient breadth of capable models available and due to their massive share in the public cloud market, Amazon is well positioned to capture large growth from AI demand in the coming 3-5 years. Andy Jassy made a pretty compelling case why AWS is executing well on the recent call:

“AWS growth continued to accelerate to 24%, the fastest we’ve seen in 13 quarters. AWS is now a $142 billion annualized run rate business and our chips business, inclusive of Graviton and Trainium, is now over $10 billion in annual revenue run rate and growing triple-digit percentages year-over-year. Our backlog is $244 billion, that’s up 40% year-over-year and 22% quarter-over-quarter. And we have a lot of deals that are in the pipeline.

We’re continuing to see strong growth in core non-AI workloads as enterprises return to focusing on moving infrastructure from on-premises to the cloud. We’re adding significant EC2 core computing capacity each day, and the majority of that new compute is using our custom CPU, Graviton. Graviton is up to 40% more price performant than leading x86 processors and is used extensively by over 90% of AWS’ top 1,000 customers.

We consistently see customers wanting to run their AI workloads where the rest of their applications and data are. We’re also seeing that as customers run large AI workloads on AWS, they’re adding to their core AWS footprint as well. But the biggest reason that AWS continues to gain AI share is our uniquely broad top to bottom AI stack. In AI, we’re doing what we’ve always done in AWS, solving customer challenges. Customers are realizing as they get further into AI that they need choice as different models are better on different dimensions. Amazon Bedrock makes it easy to use these models to run inference securely, scalably, and performantly. Bedrock is now a multibillion-dollar annualized run rate business and customer spend grew 60% quarter-over-quarter. The second challenge is how to tune the model for your application—it takes a lot of work to post train and fine-tune a model for your application. Our SageMaker AI service, along with fine-tuning tools in Bedrock, make this much easier for customers.

Another challenge is cost. If we want AI to be used as expansively as companies want, we have to make the cost of inference lower. A significant impediment today is the cost of AI chips. Customers are starving for better price performance. It’s why we’ve built our own custom silicon Trainium, and it’s really taken off. We’ve landed over 1.4 million Trainium2 chips, our fastest ramping chip launch ever. Trainium2 is 30% to 40% more price performant than comparable GPUs and is a multibillion-dollar annualized revenue run rate business with 100,000-plus companies using it. Trainium is the majority of Bedrock usage today. We recently launched Trainium3, which is up to 40% more price performant than Trainium2, and we’re seeing very strong demand for Trainium3. We expect nearly all of our Trainium3 supply of chips to be committed by mid-2026. Although we’re still building Trainium4, we’re seeing very strong interest already.

Looking ahead, the primary way companies will get value from AI is with agents. Enterprises are apprehensive about deploying to production because these agents need to securely connect to compute, data, tools, and other elements. This is a new and hard problem where a solution has not existed until we launched Bedrock AgentCore. Agents can be fully autonomous, run persistently for hours or days, scale out quickly, and remember context. At this past AWS re:Invent, we launched Frontier Agents to do that, Kiro autonomous agents for coding tasks, AWS DevOps agents for detecting and resolving operational issues, and AWS security agents for proactively securing applications throughout the development life cycle.”

So, we think that with the arrival of AI agents our thesis above becomes even more compelling. Cloud growth rates will continue to accelerate as AI agents are XPU/GPU intensive, and as agents need access to data, apps and also CPUs, the major clouds are the obvious place to run these. This should also accelerate growth for names such as Snowflake and Databricks as these are hubs for data analysis where agents can be deployed.

Satya notes that agents can also lever Azure security roles, making it safer for enterprises to operate them as you can define what an agent can and can’t access in the cloud:

“We are also addressing agent building with Copilot Studio and Agent Builder. Over 80% of the Fortune 500 have active agents built using these low-code/no-code tools. As agents proliferate, every customer will need new ways to deploy, manage and protect them. We believe this creates a major new category and significant growth opportunity for us. This quarter, we introduced Agent 365, which makes it easy for organizations to extend their existing governance, identity, security and management to agents. That means the same controls they already use across Microsoft 365 and Azure, now extend to agents they build and deploy on our cloud or any other cloud. Partners like Adobe, Databricks, NVIDIA, SAP, ServiceNow, and Workday are already integrating Agent 365. We are the first provider to offer this type of agent control plane across clouds.”

While the demand story in the coming years for the big clouds looks compelling, their share prices underperformed over the last year. What drove the recent sell off in Amazon’s shares was the capex outlook—with a heavy investment cycle, the company will go into negative FCF for the coming two years or so. However, the attraction is that most of this capex can get financed with operating cash-flow, limiting the cash burn:

So, the market is worried by a heavy investment cycle in Amazon for the foreseeable future, but we see the risk as limited due to the limited cash burn. If growth comes in less stronger than expected, capex can be dialed down again to unlock strong free cash flow. Vice versa, if growth continues to come in stronger than expected, rising capex can be financed with the growth in operating cash flow. So, negative FCF should already come down again in ‘27 on JP Morgan’s estimates:

As Amazon now goes into a new investment cycle, we think it’s best to value the shares on a gross profitability measure, i.e. pre capex or depreciation, such as operating cash-flow or EBITDA. We see strong operating cash flow growth in the coming years (27% CAGR, table above) and similarly strong EBITDA growth (23.5% CAGR on consensus estimates in the coming three years). As Amazon is now trading on a trough EV/EBITDA valuation, this positions the shares to start capturing that profitability growth as little is priced in here:

Similarly, also on a PE basis, Amazon is now at its cheapest valuation ever despite, in our view, a compelling growth profile:

Whereas data center capex plays have been the biggest winners over the last few years, at some stage growth rates in annual data center capex will start to moderate. We see this as being the big catalyst which will drive outperformance for the major clouds, as the expansion in operating cash flow relative to more flattish capex will unlock steep growth in FCF.

So, AI capex plays need ever more capex every year to show growth, but this is not the case for the major clouds. Clouds can monetize their entire installed data center base via recurring revenues. As we discussed, pricing power is strong due to customer lock-in, while expanding the cloud platform with more apps and services creates a flywheel of growth. While major clouds haven’t performed well over the last year, we continue to see them as some of the most attractive plays on the long term growth in AI demand.

Deutsche had some entertaining comments following Amazon’s sell off and we agree with the big picture here:

“Amazon is not becoming more capital intensive, rather, we see the capex ramp as a pull forward of capital that would have been deployed in the cloud over many years to drive the digital transformation of the economy which is now happening much more quickly and in a narrower time frame. Cloud was always set to transform the % of GDP that flows through digital services and it’s why we believe the opportunity in the space measures in the trillions as the COGS of this transformation. Probabilistic thinking machines (AI) are simply accelerating that transformation and making it more tangible in the medium term. Amazon has spent the better part of the last 20 years watching AWS demand signals and converting that into capacity plans, there is no company with more data and experience to make this capacity growth decision in ‘26 and beyond. The risk of underinvesting is certainly higher than overinvesting as the latter would mean that the company would simply grow into the capacity over time, while wringing out efficiencies within the infrastructure. Akin to what we have seen the company achieve in the retail business since exiting the capacity boom initiated during the pandemic.”

A credible pushback on the thesis above is that a high proportion of cloud bookings is currently coming from AI labs such as OpenAI, which may or may not have the means to pay for these 3-4 years from now. However, we think this misses the big picture that AI is finding ever more applications to raise productivity. And so even if OpenAI blows up, it’s highly likely that OpenAI’s booked capacity will simply get taken up by Anthropic, Grok, or overall enterprise AI demand. We agree with Deutsche above that should there be an overbuild of AI capacity at some stage, capex can dramatically be lowered for a number of years so that demand can grow into the capacity available and drive high data center utilization rates.

We think that the market is currently making a number of illogical moves. For example, some have pointed out that Amazon has been underperforming Walmart. However, Walmart’s outperformance has been driven by pure multiple re-rating, whereas Amazon’s EPS performance has been stellar but its multiples have gotten tremendously cheaper.

We think that the market has this completely wrong—Amazon should be trading on 45x and Walmart on 20x or so. Investors are now asked to pay an astronomic valuation for a company like Walmart with mediocre growth. Basically, the market is scrambling for safety in physical assets. Usually, when shares get absurdly expensive like this, subsequent returns are lackluster at best. Note that Walmart also traded on a value-multiple over the 2005-2015 period as the company ‘was going to get disrupted due to e-commerce’. Somewhat similar to the fears in many software names today.

Next, we will dive into semi stocks that should benefit from rising capex at the major clouds, and a number of other names where AI can drive meaningful upside in the coming years.

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

We’re taking a contrarian view here and actually see AI as bringing in more opportunity than risk for the quality software names—quality being defined as software that either has high switching costs and/or large amounts of required user training. Jensen sees a similar scenario playing out, his comments from this week:

“In our case, chip design and system engineering, we partnered with Synopsys, Cadence, Siemens and Dassault, so that we could insert our technology and whatever they need, so that we could revolutionize the tools which we use to design what we do. We use Synopsys everywhere, Cadence, Siemens and Dassault, and I will make sure that they have 100% or whatever they want so that I have the tools necessary to create the next generation.”

So, Jensen’s key aim is for AI to boost productivity of his engineers with as much as possible. By taking over more workloads, AI in combination with EDA software can boost productivity multifold. The bull case for the EDA names, such as Synopsys and Cadence, is to monetize part of this new economic value added. We see Jensen’s key goal here being as not to save 1 point of margin by trying to rip out and replace complicated software systems, with questionable odds of success, but instead to multiply the productivity of his workforce by as much as possible.

We think this is true for a lot of good software tools. For example, in Excel, we see the most likely scenario being that analysts will use spreadsheets with AI built-in. So, Claude will first build a complicated spreadsheet and then analysts tweak parameters or certain formulas to get the final output that they want. This is also how we’re currently writing code for our software projects. Our usage of IDE software tools hasn’t decreased, it has increased. Why? Because with the help of AI it has become much easier to write large amounts of code, and we can modify it inside an IDE to finally deploy it. Thus, we’re writing more code than ever before and our IDE usage has increased as a result.

This also gives us good insights into the quality of the code Claude is actually writing. And we think this is where a lot of confusion is coming from. For people who don’t understand code, it will seem magical. I give Claude a prompt, it writes lines of code, I test it on my computer, and it works. “You can go to localhost right now to check out my new app”, we’ve seen written on Reddit. So, we can see why some are jumping to the conclusion that “SaaS is toast as I can easily write my own SaaS now.”

Well, there’s a big difference between code that’s good enough to run on your local machine, and code that you can actually ship to run crucial enterprise workloads on. Production quality code has to fulfill all three of the following criteria: it has to be secure, it has to be scalable, and it has to be cost efficient. Currently, the code that AI is writing fulfills none of those three. Zero. In simple terms, it writes a lot of useful code but it also does a lot of stupid stuff. And we’re not the only ones saying this, Google has similar results. This is from the recent call:

“About 50% of our code is written by agents, which are then reviewed by our own engineers. But certainly, it helps our engineers do more and move faster with the current footprint.”

In other words, heavy human oversight is needed for production grade code. Also Andrey Karpathy, who coined ‘vibe coding’, points out that oversight and scrutiny remain crucial:

“Today (1 year later), programming via LLM agents is increasingly becoming a default workflow for professionals, except with more oversight and scrutiny. The goal is to claim the leverage from the use of agents but without any compromise on the quality of the software.”

Yes, AI massively boosts coding productivity. But, at the same time it does a lot of stupid stuff. So spinning up tons of AI agents to write a huge code base that an enterprise can safely deploy for a crucial workload, e.g. to run my warehouse management on, we don’t see this happening any time soon. Remember, an ERP system has 10 to 100 million lines of code, no enterprise is going to deploy a system that hasn’t been fully vetted and battle-tested. If your ERP suddenly crashes, this is an existential risk even for a giant like Walmart. It can take the whole company down.

The other problem is that this AI agent approach remains hugely expensive on large code bases. We have the most expensive Claude plan (£100 per month or so), and when we give Claude a large code base, we burn through our credits in a few prompts. So, Anthropic then notifies us that we have to take a timeout and can start coding again in 5 hours. Obviously, this isn’t great. For a large code base, we find it much easier to give Claude a narrow set of code, e.g. a few example functions and some context, and then explain the code it needs to write next. This way, Claude can write huge amounts of code without burning through all our credits.

So, are enterprises replacing best-in-class software with internally developed tools? While there will be some instances where this happens, we think that this will be fairly limited. For example, let’s say I want to replace my supply chain management software with an internal, more minimalistic tool. I would still have to hire a team of engineers to build and maintain it. Those engineers aren’t cheap, plus if they leave, it might be hard to find replacements who can understand and update that code base. So, we think it still makes much more sense to purchase a battle-tested and secure software off the shelf, that corporations know they can scale and that they can trust to run their operations on.

And, it’s not like cheap software hasn’t been available already. Free open source software has been available for decades, yet, they haven’t been able to disrupt the growth story for the best-in-class software names. A key reason is that even if you deploy an open source software, you need to hire an engineering team to manage it. For example, in databases, there are plenty of free database tools available. However, if you want to deploy it, you need skilled engineers who know how to spin up a cluster in AWS for example, deploy the open source database on this cluster, manage the backups, software updates, patches etc. And you have the risk that if these engineers leave, it could be hard to find replacements. Spinning up and managing an open source Cassandra cluster for example is fairly complicated, and we say that with a good amount of cloud engineering experience.

So, this is why a lot of enterprises choose a managed database. You simply connect to an API in the cloud and then MongoDB, or Google Cloud Platform etc. will manage the database for you. By purchasing a software, you not only get access to the software, but also to the entire support network of the software provider. So, if my system suddenly crashes, you actually can call people who know how to fix it. If you don’t have this, you could get wiped out. We saw this a few years ago when CrowdStrike crashed, loads of businesses went down. Imagine an internal core system that suddenly crashes and the key engineer who understands it is on holidays, it could take your business down completely.

So, is MongoDB going to get disrupted here? We’re really not seeing this, as free databases such as Cassandra, Postgres, and the open-source version of MongoDB itself have been available for decades. Enterprises go for the leading software products because they know these have been battle tested, i.e. they can scale to large volumes, they know that these are secure, and that there is a wide support network available to fix your systems.

At the same time, trying to replace one of these systems is actually a big risk for enterprises as well. The last decades have seen plenty of quarterly calls with sudden earnings warnings due to struggles in a new software rollout. And there have been instances where companies simply gave up on the rollout of their ERP, supply chain management, or database system. Remember, these are often 2-4 year projects. So, the assumption the market is making here that enterprises are going to try to move away from their ERP, supply chain, databasing, data analytics, finance, CAD, etc. tools is rather naive. Ripping out one of these is a 2-4 year project with not guaranteed success, so mathematically, there’s not going to be much disruption for the foreseeable future.

Looking at the history of software, software is extremely resilient. Take mainframes for example—a 1960s tech which got disrupted in terms of new demand a long time ago—it’s actually still alive and growing:

Similarly, Excel and Windows later on had to compete with much more modern systems and even free open source tools. Both remain in the core stack of many enterprises today. Despite the rise of much better and more modern programming languages in the last few decades, tons of code today is still being written in 1970s C. Python, developed in the early ‘90s, remains the dominant programming language today even in AI coding. Python has a lot of flaws—it’s not even a static programming language and it’s extremely slow. However, once a software tool really becomes dominant in a certain niche, history suggests it’s extremely hard to replace.

The ‘Lindy effect’ suggests that the future life expectancy of non-perishable items—like ideas, technologies, or books—is proportional to their current age as their resilience and robustness increases with time. So—C, Windows, and mainframes are actually much more likely to still being used 20 years from now than a hot and new app such as Claude Code. It’s even possible that two years from now no one will be using Claude Code as Gemini Antigravity or Grok might become absolutely dominant.

At the same time, it’s also not the case that enterprises have an abundance of deep engineering talent that know how to build AI workflows themselves. As most companies want to unlock the productivity gains from AI, the only way to do this in our view is to integrate AI into their core tools that run the backbone of the business. And this means revenue acceleration for those software names as those AI workflows can be monetized. Accounting software provider Xero illustrated this on their call:

“First, we’re using AI agents to automate actions and workflows across accounting, payments and payroll, so we can help our customers and give them time back. Think about automating bank reconciliations, or automatically closing books. And we’re also building agents that provide actionable insights, which helps our customers run their business better. Imagine being able to run detailed scenarios for your cash flow and test out different assumptions without ever having to use Excel. Additionally, SaaS apps no longer need to be unintelligent. They can intelligently surface to you what you need to pay the most attention to. Finally, accounting and payments are not areas that you can have errors in. People need to be able to trust their financial management systems to reconcile transactions, close books and calculate your tax. This is crucial information which LLMs do not have, and it gives us a strong advantage.

Through FY ‘26, AI has become ubiquitous across our platform because of the value it can deliver for our customers. We are prioritizing using AI to automate the key jobs that 90% of our users use and that are the most time consuming, like automated bank reconciliation, AI invoicing and document ingestion. This means that our AI creates maximum value for them. We designed our AI for a future where agents complete tasks end-to-end in the background and SaaS applications become the space where humans oversee their agents and get the insights that matter. Previously, our homepage was a to-do list of tasks for humans. Now it’s a control room for AI agents and insights, showing users what AI has already done.”

We think that the amount of SMBs that are going to turn off their core accounting software is going to be rather limited. Are you really going to invest time in a difficult accounting software transition, or are you simply going to deploy AI in your existing accounting software to automate your workflows? We suspect that 90+% will do the latter, which means substantial upside for the accounting software names. So, we think that the best-in-class accounting software names are heavily oversold here. Intuit’s EPS are still trending up—and these will very likely continue to trend up, we think it will actually accelerate—and at some point in the coming years the market will start figuring out that this is actually a great business and then we’ll see a rerating:

Vitec made similar comments on their call—yes, customers want the benefits of AI, but they also want their software not to change too much:

“They really want to hear that we’re working on it and we have it in the pipeline, but a lot of them are super conservative and say, yes, we want all of that, but please don’t change anything. They are really reluctant to do any big bang changes because, and this is one of the moats around vertical market software, is that it is so embedded into the customers’ processes. Yes, they want to benefit, but they don’t want to change anything at the same time. So, you will have to do this very gradually. The important thing is that they feel that they can benefit from these improvements over time.”

In our personal workflows, at this stage we can’t think of any tool that we’re going to shut down. We still need to tweak spreadsheets in Excel. Typing in precise prompts is also time consuming, it’s often much faster to tweak the spreadsheet just yourself. So, we use AI when we need to write 100-800 lines of code and then it makes sense to invest the time to write a precise prompt. However, you don’t want to have to state a prompt for every tweak you need to do in your spreadsheet, as doing it manually is faster, and much more fun

While AI is superb at generating images, we’re even keeping our Canva subscription. Writing prompts and waiting for a good image to be generated is slow. It’s much faster to just quickly get a great image already available in Canva and do some quick editing inside Canva if needed. A business like Canva is one of the most at risk from AI, however, even here you can see a case why it will likely still retain a large amount of users. Similarly, for Adobe, a free and cool video editing software from ByteDance called CapCut has been available for free for quite some time now, however, all the gifted video editors we know swear by Adobe Premiere Pro. We’re not saying that Adobe isn’t at risk, however, we do think they will retain a large amount of users.

Overall, we’re making more use of our tools than ever before as now our AIs are using them as well. For example, for our financial data terminal, we’re sending ever more requests to the API as AI is now analyzing loads of data to come up with investment ideas. So, Grok actually initially told us to take a look at this deep tech disruptor, and we subsequently reviewed it and then invested in it:

Gartner makes a similar conclusion, with a likely shift towards agent monetization based on usage versus pure SaaS subscriptions:

“Look, Claude Code can write code, software is going to zero” is a pretty extreme conclusion in our view. It doesn’t address the fact that free open-source software has been available for decades, which illustrates that it’s not purely about price why enterprises implement a certain software. Reliability, security and networking effects around software are three crucial factors. At the same time, ripping out a single core software system is often a 2-4 year project involving considerable business risk.

Enterprises spend single digit percentages of revenues on software versus 5x that amount on personnel. It makes much more sense to optimize and boost the productivity of your workforce rather than doing a set of highly risky software transitions with the aim of saving a few points on margin. It makes much more sense currently for an enterprise to invest its time into inserting AI into the business and automate as much workflows as possible. This is where you can really get the cost savings, or alternatively, to grow your top line on the same cost base. This is clearly also the strategy that Jensen is going for.

We suspect that current market action is being driven by a lot of trading flow aimed at ‘cutting losses’ and also by generalists, or quant funds, who have limited knowledge compared to the software specialists. JP Morgan makes a similar conclusion:

“In our view, generalist money flows responding to the rapid AI product rate-of-change dynamic are overwhelming the deeper-thinking fundamental software sector specialists who are slightly more grounded in the principles that create stickiness for Enterprise software businesses. These generalist money flows are invoking more knee-jerk selling, which is being exacerbated by index arbitrage basket selling, programmatic de-grossing, cross-correlation factor contagion, and a passive flow liquidity vacuum.

Two to three years ago, our investor call series with Boston Consulting Group (BCG) highlighted the evolution of the software stack, with the emergence of a data layer (Databricks, Snowflake, etc.) creating a foundation for AI Agents to handle some portion of what had previously been handled by core systems of record. It feels like an illogical leap to extrapolate Claude Cowork Plugins, or any similar personal productivity tools, to an expectation that every company will hereby write and maintain a bespoke product to replace every layer of mission-critical enterprise software they have every deployed.

The current state of investor psychology has started to feel like the inverse of the 2021 software euphoria period, in which the same institutions that are currently selling software stocks at 25-30 year valuation lows were buying Software as fast as they could when the group was at a 20-year valuation high, fully convinced that Cloud was the world’s greatest business model and interest rates would remain at zero forever (and kindly recall our differentiated negative stance during this period). 2020-2021 may have been an equally surprising period of time in terms of mass investor psychology shifting radically into a very unrealistic mindset.”

However, we do see one bear case scenario. The most credible bear case in our view is if AI is able to take humans out of the loop entirely. Overall, we would define such a scenario as being ‘AGI’. This would be a scenario where Jensen can fire all his engineers and a bunch of AIs are running Nvidia and reporting to Jensen or to an uber AI. Part of the reasoning above is based on the view that you need humans and AI agents working alongside, and so both need a common set of tools. If at some stage, the AI becomes so smart that it can manage to take humans out of the loop completely, we think that the AI will indeed simply write its own tooling.

Personally, we’re skeptical that this level of intelligence can be achieved with the current transformer architecture. Although we’re not excluding that with further advancements in AI this scenario could be possible 10 years from now or so. Therefore, we continue to view portfolio diversification as important. This is a key reason why we cover a wide variety of tech topics in this newsletter ranging from semis, software, biotech tools, space companies, industrials, robotics, online apps, fintech, adtech, etc. That being said, if AGI arrives and AI is running the entire economy, or large parts of it, there is no reason anymore for a capitalistic economy anyways. A socialist system will be the easiest way for everyone to share fairly in the abundance that AI creates. So, potentially, portfolio diversification won’t matter in that scenario anyways as everything will get nationalized.

Coming back to the current reality, we see the risk-reward as still being attractive in software, and so we’re comfortable to run a 15-20% portfolio weighting in this space as we think that revenue and EPS growth will accelerate from here and that the market is massively underestimating just how resilient various types of software have been over the last 60 years. Software is more like geology, where one layer is built on top of the next. And we think that the next logical step is to build AI on top of proven and reliable systems that also humans still know how to operate.

Next, we will dive deeper into AI and semis in particular where Credo got hammered due to the copper vs optics debate. We’ll give our top semi picks in the AI data center space and our views on ASML after a strong and deserved rally, and what we’re doing with our shares now. We’ll also overview our buy-rated list and our top picks as we’ve been going through most earnings calls so far and we’re making a few trades here, including taking profits in some winners.

Credo, Copper vs Optics

Read more

After years of share losses in the CPU market, Intel is finally seeing stronger demand for its products. However, this time, the company didn’t have the necessary supply available. This is Needham on the mishap:

“Intel is currently facing supply constraints, most acute on Intel 10/7 where the bulk of its volume resides. As a result, 1Q revenue guidance of $12.2BN came in below Street ($12.5BN), though management indicated revenue would have tracked well above typical seasonality if sufficient supply were available. A more pronounced revenue decline in CCG relative to DCAI is expected in 1Q, as internal supply continues to be prioritized toward server demand. While under-shipping is expected to persist beyond 1Q, the quarter should be the trough for supply constraints. Supply is expected to improve gradually through the year, with revenue tracking better than seasonality in the remaining quarters, led by DCAI.”

JP Morgan points out that data center CPU demand will be strong, while consumer markets will see weaker demand due to surging memory prices:

“Wafer supply should begin to improve in Q2 as Intel transitions CCG wafers to DCAI and continues to drive yield improvements, but this will still likely prove insufficient given the level of demand for traditional server CPU compute that is materializing on the back of rapidly rising volumes of AI inference workloads (traditional CPU compute is increasingly being leveraged for a variety of tasks/processes incl. workload orchestration, pre-/post-processing, data retrieval etc.). Intel will also no longer be able to lean on inventory drawdown as a means to support revenue growth in Q1, exacerbating the impact of the capacity shortage.

Looking beyond near-term supply-demand dynamics though, Intel is addressing the shifting demand landscape by streamlining its server CPU roadmap, focusing on high-end Diamond Rapids and Coral Rapids SKUs, and pushing to accelerate go-to-market timing as it looks to stem share loss to AMD (and potentially recapture share at key hyperscale and enterprise customers). The outlook for Client/CCG is less “rosy” though, with memory/storage supply constraints and elevated pricing likely to crimp growth for the PC market this year (as we highlighted in our 4Q preview, our global team is now forecasting a 9% decline in PC shipments in 2026).

On net, we still think there is enough of a demand tailwind in DCAI to more than offset a decline in CCG this year, driving LSD-MSD % Y/Y overall growth for INTC. Gross margin should improve incrementally through 2026 as volumes ramp, 18A yields improve, and higher pricing in server CPU begins to take effect, though we still anticipate GM exiting 2026 barely at the 40% level.”

The big hope for Intel is that the company can start building out a credible foundry business towards the end of this decade. While we think it is likely that the company will start seeing smaller orders for future products from the US semi designers as there is strong support for Intel in Washington, including from the White House, it’s unlikely that these customers will commit to large orders for the foreseeable future as Intel will have to prove itself.

For example, Jensen might be willing to give Intel orders for their gaming GPUs, and potentially Broadcom/AMD/Marvell/Qualcomm could give them one or two smaller products as well. These could also for example be less advanced products such as IO dies in a chiplet architecture. This would still provide a credible bull case as should Intel execute well on those, it could expand its foundry business over time. While TSMC is obviously awesome, no one wants to be beholden to a single supplier in tech. For example, Jensen also gave gaming GPU orders to Samsung not too long ago.

So, besides political support, we do think there would be organic interest as well from the advanced semi designers to be able to source from an additional credible foundry, and with US-based fabs. Naturally, this is if Intel can do it. However, Samsung is building out a fab hub in Texas as well; so obviously Intel is facing a strong competitor to be a second supplier to the semi industry. Also UBS is constructive on Intel’s Foundry story:

“We remain constructive on the story in foundry and with the release of the 14A production design kit (PDK 1.0) late this year, we see INTC possibly signing deals with one or more of NVDA, AAPL (M-Series), and maybe Amazon and a high-end consumer product. This will likely be a positive catalyst and we are tweaking our PT up just slightly in recognition, but we still struggle with the earnings power here - especially as INTC is giving up about half of the incremental profits to Apollo and Brookfield due to the financing agreements signed several years back.

Moving to new products on its 18A process will help foundry margins, but hurt product margins, so we think the net effect is only modestly positive. Ultimately, we quite like the story but the stock has gone up so much and the bull case to support further upside from here is 3-4 years away, in our view. So we land in about the same place as before with a slight PT increase from $49 to $52 based on SOTP and a little more confidence that external foundry deals are coming late this year.

For an indicative bull case foundry analysis, we consider 15% of AAPL, NVDA, and AVGO TSMC wafers moving over to Intel 14A in 2030. At a $32.5k wafer price this works out to $6.9B of external foundry revenue for front end. We add $5B of advanced packaging and Intel’s own internal wafer consumption works out to $42.7B at the same transfer price, suggesting $54.7B of foundry revenue in this scenario. At a 25% op margin and 10.5x EV/EBIT multiple, this works out to $143B of EV or $118B after subtracting the SCIPs, substantial upside (to our base case foundry valuation). Still, the nearest catalyst is a customer announcement at the end of this year and even then the ultimate movement in Intel’s financials will be remote, with volume production expected only in 2029.”

While Intel has been a large share donor in data center (and consumer) CPUs:

Evercore points out that Intel’s market share on their latest data center CPUs has been solid vs AMD:

Overall, the outlook for Intel is looking a lot better now than in previous years. It looks like 18A is going to be a solid node with yields that will get to sufficiently high levels. At the same time, AI workloads in the data center provide a new growth opportunity for the company’s otherwise rather dull CPU business.

We do think that consensus estimates look reasonable here and that the company can start growing its top line again. In the last two quarters, FCF has turned positive again—the first series of two consecutive quarters of positive FCF since 2021—with Intel actually generating $1 billion in cash. However, even if the company can work itself back to $3-5 billion of annual FCF over time, investors are still only getting a 1-2% FCF yield on the current EV of $250 billion. This illustrates how the market is now pricing in quite a rosy scenario already.

We don’t dislike Intel but execution has to be fairly flawless to justify current valuations. We can see the shares doing well as we think the company will manage to get orders for some of the smaller future products from the leading US semi players. Both Lip-Bu and Jensen are good buddies with the President, with the latter two also being Intel shareholders. Washington wants leading edge semi manufacturing in the US and no doubt that Lip-Bu will have told them that he needs customers at 14A in order to do so, like he has been saying on the earnings calls. However, needless to say, Intel remains a speculative story at this stage and it’s not the sort of stock where we would hold a large position in. We would rate it as a ‘soft/speculative buy’ at best.

Next, we’ll dive into further developments in leading edge semis with insights into TSMC, SK Hynix, and the overall DRAM and NAND industries. Finally, we’ll also review the sell off in Unity and the current panic in software stocks.

Read more

Planet Labs is a vertically integrated leader in Earth intelligence. Basically what this means is that the company builds satellites, gets them up into orbit via SpaceX, and then does a daily scan of Earth with its satellites to get insights into what is going on. Key customers are the US government and other allied governments around the world as thanks to AI, Planet Labs can detect all military infrastructure being built by adversaries.

We like this company as barriers to compete are fairly high. You need to get a satellite network up into space, develop the systems to manage these satellites and analyze all imagery, and then create a portal so that customers can use it for insights. So, it’s a bit like an AWS or Bloomberg terminal for space-based data. This is a business with recurring revenues and pricing power. Gross margins are starting to get attractive at around 60% while the business has also turned FCF positive, and with a net cash position on the balance sheet. As valuation last year looked reasonable for the opportunity set, this is an investment idea we reviewed at the time and took a position in.

However, one of the reasons that we’re currently more cautious than the market—following a very strong share price rally over the last months—is that growth in commercial is actually down. So, growth is mainly coming from military intelligence for government customers. From the last call:

“Shifting finally to the commercial sector, where revenue was moderately down both year-over-year and quarter-over-quarter. While this trend is expected given our increased focus on large government customers, we remain confident in the commercial sector as a significant market opportunity for Planet, especially as we continue to advance our solution capabilities. We believe that AI-enabled solutions we’re developing for our government customers will enable us to deliver insights that can serve applications across a broad range of industries and use cases from supply chain, security, and optimization, to insurance, finance, energy and agriculture; where we have had a number of marquee customers today. We expect these solutions will help unlock growth in the commercial sector, bridging the gap from data to insights for those customers.

To share a recent commercial highlight, we signed a new operational contract with AXA, one of the world’s leading insurance groups following a successful proof of concept. AXA will integrate data from Planet Basemaps, our medium resolution monitoring satellite and high-resolution tasking fleets, to enhance claim processing efficiency and accuracy for property management. We’ve also signed a strategic marketplace agreement, which will add Planet’s products to AXA’s DCP platform, making them commercially available to AXA’s vast client network of insurance partners. This partnership marks a major step forward in proactive data-led resilience in the face of complex disaster and crisis management needs.”

We don’t disagree that Commercial can be a big opportunity long term, but the current valuation warrants strong execution. Currently, we see that traction is fairly limited and that revenues are actually down. The big positive though is that incremental margins on contracts such as the one with AXA are 90%. As all necessary costs have already been made, the margins on new business are huge.

What really catalyzed the shares to move higher was the following announcement on the recent call:

“Secondly, we recently announced a funded R&D initiative with Google called Project Suncatcher. Suncatcher aims to enable scaled AI computing in space by putting Google’s tensor processing units or TPUs on purpose-designed satellites where they can leverage the energy of the sun and shed excess heat into the natural coolness of space. This is a competitive win for Planet and our strong track record of building, launching and operating over 600 satellites to date. Suncatcher aligns well with our technology development roadmap for Owl, leveraging the same satellite bus and is therefore highly synergistic. As previously announced, we’re planning to deploy two prototype satellites in early 2027. We’re excited to be working with our long-term partner, Google, to develop this promising new technology.”

From the Google paper below—it looks like TPUs can handle the radiation in low Earth orbit (LEO) and that with the introduction of SpaceX’s Starship, launch costs will fall sufficiently far to make this economical:

“Trillium TPUs are radiation tested. They survive a total ionizing dose equivalent to a 5 year mission life without permanent failures, and are characterized for bit-flip errors. Launch costs are a critical part of overall system cost; a learning curve analysis suggests launch to low-Earth orbit (LEO) may reach ≲$200/kg by the mid-2030s.

SpaceX launch pricing data and mass launched from Falcon 1 to Falcon Heavy yields a ∼20% learning rate, meaning the price per kg falls by ∼20% for every doubling of cumulative mass launched (over all vehicle classes). If the learning rate is sustained—which would require ∼180 Starship launches/year—launch prices could fall to <$200/kg by ∼2035. While this would be a substantial achievement for SpaceX (particularly given the technological discontinuity inherent in switching to Starship), it is still far below stated launch targets for Starship. Even if this launch rate is reduced by ∼ 70%, prices could drop to $300/kg in the same timeframe, which would still have a substantial impact on feasibility of largescale constellations. Further, there is precedent for a sustained ∼20% learning rate over multiple decades in other advanced industries leveraging mass-production (notably, solar panels). Given the long lead times required to reach scale for this type of ambitious project, it’s strategically beneficial to commence work on early milestones in anticipation of projected price declines.

Alternatively, our analysis of Starship 4 public specifications and data suggests that SpaceX launch costs to LEO may drop to ≲ $60/kg (10× component reuse); if SpaceX’s 100× component reuse target were achieved, costs could reach ≲ $15/kg. Assuming 10× reuse and even the highest estimates of current SpaceX margins (up to 75%), launch price to customers would drop to <$250/kg (although margins are currently supported by SpaceX’s near-monopoly and hence likely to decrease with the anticipated entry of competitors such as Blue Origin). Realizing these projected launch costs is of course dependent on SpaceX and other vendors achieving high rates of reuse with large, cost-effective launch vehicles such as Starship.”

If SpaceX hits the conservative end of their Starship targets, we already get to $200 per kg. Assuming a future single Starship launch cost of $20 million—covering fuel, pad maintenance, and amortized hardware—and with a payload of 100 tons (100,000 kg), we get to $200 per kg. Both these numbers are the most conservative estimates from SpaceX for Starship. Elon has been mentioning that $20 per kg is the actual target.

Long term—we’re talking 2030s here—this could become an attractive business for Planet Labs. Assuming the company puts 3,000 AI satellites into space on an annual basis at $5 million in contract value per satellite, we get to a $15 billion annual revenue opportunity. Part of this revenue opportunity is that Planet would charge a maintenance fee for the entire Google satellite fleet in operation—managing the positioning of the satellites, monitoring the fleet, software updates, end of life management etc.—and these would come in on an annual basis post the launch, so part of the $15 billion annual revenue opportunity isn’t immediate. However, these would be extremely high margin revenues. The reason is that Planet has already built out the command and control center to manage its satellite fleet, and so additional satellites can now be added at little marginal cost. As these maintenance revenues would be recurring, they would attract a high multiple from investors.

Why not more satellites for Planet? While AI accelerators are able to withstand radiation in low earth orbit, if we move them higher, they will get fried. So room in space is limited actually, it’s not like we can build unlimited data centers in orbit around earth. At the same time, there is a collision risk. So, you don’t want to pack too many satellites close to each other. One collision could create a cloud of debris that triggers a chain reaction called Kessler Syndrome, which could turn a specific altitude into a no-fly zone for centuries. Remember, fragments travel at 17,000+ mph and so they can punch through a satellite like a bullet.

The reason why this remains speculative is that we’re very early in this concept—tests will be done by 2027 and then maybe in the early 2030s SpaceX’s launch costs will have fallen sufficiently far to make this whole idea economical. If SpaceX’s progress is slower, this is a story that’s really a long time away, e.g. 2035 or beyond.

Some of the necessary innovations SpaceX has to make include firstly the catch. I.e. SpaceX must perfect to catch Starship with the tower arms structure called ‘Mechazilla’. Currently, they’re only catching the booster. Until Starship is reusable, you are throwing away the Raptor engines and a massive steel structure every flight, which is too costly. Secondly, the heat shield has to work. This is an extreme materials engineering problem. You need Starship to land and fly again within 24-48 hours, so without humans inspecting and replacing the heat tiles manually. Thirdly, volumes. SpaceX needs to be flying Starship 100+ times a year to amortize the costs of the launch pads and ground support.

So, while exciting, the partnership with Google and this bringing in sizeable revenues for Planet Labs by the early 2030s still looks somewhat of a moonshot. Timelines of 2035 and beyond are probable, which is really far out. Given a high potential value into a distant future, we suspect this is a share price which will be fairly volatile in the coming years, with sell offs and rallies. Although we do think that due to Planet Labs’ strong position in space infrastructure and its positive free cash flow, a high current multiple is justified. New investors looking to get in could wait for a pullback to exploit volatility.

Deutsche Bank has done some further modeling on the economics of data centers in space:

“After further refining our assumptions, we estimate the near-term cost of deploying a 1 GW space data center is at least 7x higher than terrestrial. This gap can narrow to 4x later in the decade and then eventually reach close parity in the 2030s. This reduction is primarily driven by decreasing launch costs and further optimization of satellite design and power efficiency, which leads to a lower amount of mass required for orbital deployment. We assume the current industry price of a launch to low-Earth orbit (LEO) is ~$70m or $4,000/kg. Over time, we assume the cost can decline materially to as low as $10m or below $70/kg with full rocket reuse and operational scale. We estimate the cost of a data center satellite to be $2.1m in the relative near-term or >$40k per kW.

For background, a satellite is typically made up of two main parts: bus and payload. The bus is the main structural framework and support systems of the satellite. It acts as the “vehicle” that enables the satellite to operate in space, maintain its orbit, and survive the harsh environment of vacuum, radiation, and extreme temperatures. Key components include the physical frame, solar panels, thermal management system, propulsion, and optical terminals. The payload is the specialized equipment or instruments that carry out the satellite’s primary mission. For a data center satellite (we refer to as a DCS), the payload is the GPU or TPU.

Why put these in space?

Energy – at the right orbit such as in dawn-dusk SSO (sun-synchronous orbit), solar panels can harness the sun’s free power 24/7 and importantly, garner 40% higher intensity than on the Earth’s surface due to a lack of an atmosphere filtering/scattering sunlight. As such, operators can generate up to 6-8x more energy in orbit (leveraging continuous availability + greater irradiance) and avoid dealing with expensive/complex power grids terrestrially and also eliminate battery backups.

Cooling – represents a large burden, accounting for an estimated 40% of energy consumption, having to use large amounts of water/piping. In fact, Nvidia’s Jensen Huang recently commented for a 2-ton GPU rack, 1.95 ton is cooling mass. In space, cooling will require attaching a passive radiator on the dark side of the satellite (i.e., part not facing the sun) which can dump waste into the vacuum of space.

Latency - optical laser links traveling through vacuum are faster than fiber optic cables on the ground, potentially by >40%. This is due to refractive index of glass and indirect path of travel for cables. In relation, satellites are increasingly generating data in space (e.g., imagery, weather data, climate monitoring). Currently, the data is downlinked to Earth for processing, which is slow and bandwidth-heavy. By having the data centers close to the satellites (“edge compute”), the processing can happen much faster in orbit with only the results/insights being sent back down.

From a TAM perspective, we think DCSs represent an entirely incremental opportunity for both launch and satellite manufacturing companies. The hyperscalers are very well capitalized and can easily fund future deployments if desired which was a concern for some previously proposed LEO mega broadband constellations (e.g. Rivada). Initially, we expect deployments to be small in order to prove out the engineering and economics, beginning in 2027-28. Should these be successful, we expect constellations to reach hundreds and then thousands of satellites in the 2030s.”

Deutsche has a lower cost per satellite than we modeled above, but we included in our calculation the gross margin Planet Labs will make plus the value of the servicing contract to manage the satellite in space. So, Deutsche’s number is basically a manufacturing cost. Therefore, $5 million in revenue opportunity per satellite seems reasonable to us. The big question mark will be volumes in the 2030-2035 period.

There is actually upside to this number as Deutsche notes that the cost per satellite will increase due to rising power requirements:

“Current satellites generate up to ~30kW of power (e.g. ViaSat-3) and we suspect this may need to increase significantly over time to >100kW. With much more power to run GPUs or TPUs, the satellite will need to utilize larger or more radiators and/or develop more efficient ways to transfer the heat from the chips to the radiator.”

While other costs in satellite manufacturing will decline due to economies of scale, the overall cost of the satellite will continue to increase due to increasing power requirements.

Overall, whereas we were initially inclined to take profits in Planet, after doing this writeup and properly reading through the Google paper and Deutsche’s modeling, we’ve decided to maintain our position here. Planet Labs is a rare pure play on the long term growth in space and that is actually making money with its positive FCF. Other space-related companies are typically burning loads of cash, with the exception of SpaceX. Planet Labs’ valuable Earth intelligence platform provides an attractive growth asset, while its capabilities in satellite manufacturing and management provide valuable long term optionality.

Optically, 24x forward sales for 20-30% annual top line growth in the coming years looks expensive. However, we think there will be more opportunities down the line—e.g. the Google deal being one of them—bringing in substantial upside in the long term. Typically, these types of well positioned companies in innovative industries find new revenue streams over time as openings emerge. Examples include SpaceX with Starlink, Amazon with AWS, Apple’s App Store, Nvidia with data center GPUs, Broadcom in ASICs etc. The recent Planet Labs-Google deal is just the latest example of well positioned tech companies finding new revenue streams which the market wasn’t even considering.

Also Deutsche hints at further opportunities down the line in space:

“Looking farther beyond, we think deploying large amounts of compute into orbit is important to cultivating efforts on the Moon and eventually Mars. DCSs orbiting the Moon can enable edge processing for AI autonomy, enabling real-time decision-making in habitats during communication delays. For instance, they might handle data from rovers or life support systems locally, fostering self-sufficiency for a permanent Moon base settlement.

Beyond that, Musk pointed to building satellite factories on the Moon (presumably leveraging Tesla Optimus humanoids) and then using a electromagnetic railgun (also referred to as a “Mass Driver”) to send satellites to lunar escape velocity without the need for rockets. This is because the Moon has just one-sixth the gravity of Earth and no atmosphere (no drag), thereby, requiring less effort to deploy satellites from the Moon. Musk believes this will ultimately allow for scaling up to >100 TW of AI.”

Planet Labs’ financial model from the investor day a few months ago is below, this is the model from before the Google deal and thus the optionality of AI data centers in space in the 2030s:

Next, we’ll discuss a name which is now in a similar position as Planet Labs was last year, i.e. an under-covered stock and with likely an even larger opportunity set. Revenue growth is already exploding, gross margins are top notch, and the company is already generating attractive GAAP operating margins. A rarity for a hypergrowth tech name. Its excellent financial metrics reflect the company’s strong positioning with a disruptive tech.

Given these characteristics, we see valuation as attractive here. It’s far cheaper than Planet Labs. The stock is also uncorrelated to the AI semis cycle and the current panic surrounding software stocks. So, it’s a nice way to diversify a tech portfolio with a more unique name and which should be looking at a long runway of growth in a massive TAM.

As a bonus, we’ll review one more name at the end where we’re currently doubling our position in. Again, this is a name that few investors will be aware of and with similarly a large long term opportunity.

A Deep Tech Disruptor

Read more

Generally, exposure to pharma R&D is an interesting thematic for investors:

Especially when you can start selecting niche picks & shovels providers with high exposure to the high growth therapies:

However, post the covid lifesciences boom, biotech R&D went into a cyclical downturn with growth rates coming to a halt (chart on the right):

With new clinical trial starts going into a downturn:

Since then, biotech funding (raised capital) started recovering in recent years:

And this has already started to translate into improved order flow and book-to-bill ratios for the bioprocessing providers:

Sartorius Stedim (DIM.FR) is a good quality player in this space, and the downturn post covid is well illustrated in its forecasted EPS chart below. Forecasted EPS have now started to gradually recover and with improved order flow commentary. So, it looks like we’re past the trough now and can start moving into a new upcycle in the coming years:

Nvidia mentioned at the JP Morgan conference that they’re seeing high AI spend in the health care sector overall, with AI usage growing at three times the rate of the overall economy. Again, this is a positive indicator for future innovation potential in this space.

Lonza is somewhat similar to TSMC in the semiconductor industry, as the company is a leading manufacturer of biologics for pharma and biotech customers. In the biotech supply chain, these types of fab players are called the CDMOs (Contract Development and Manufacturing Organizations). So, Lonza typically has some of the best insights when it comes to the outlook for the overall industry. This is the CEO at the JP Morgan conference this week:

“First of all, the underlying pharma market itself is an attractive one, not very cyclical, growing at 6% to 7% a year in terms of value. Then there is the increase of outsourced manufacturing of pharmaceutical companies to partners like Lonza, which adds incremental growth of 1% to 2%. Lonza has active market selection where to invest, into which technologies and subsegments, adding another 1% to 2% of growth increment. This leads to a growth of 8% to 10% on average over time to the CDMO market addressable by Lonza. Then there’s the Lonza Engine, which adds to our capability and actually enables us to outgrow by another 2% to 3%. So, there is a growth potential of 10% to 13% for us available.

The clinical pipeline of more than 7,000 molecules is growing at 9%. We cover services from early phase discovery through clinical phases and to commercial large-scale manufacturing. We are able to address above 90% of the innovative pipeline in pharma. We see that the number of innovative molecules being contributed by small biotechs and midsized companies in this pipeline is actually at 75%. Why is that relevant? Because biotechs actually don’t invest in their own manufacturing capabilities, and every dollar they spend is in terms of innovation. And we also see a relevant trend in terms of pharmaceutical companies outsourcing manufacturing to capable CDMOs. There will likely be a preference of the pharmaceutical companies to create additional capacities in the US. Lonza is operating a well-diversified global manufacturing network. So, we see ourselves in a good position to help our customers regionalize their supply chains.

70% of our business is commercial business although clinical and early-stage biotechs work is important for us to have this pull-through and learn about what’s going to be needed in the future. It’s a business that we want to do and nurture, but it’s not a big growth driver for us upwards or downwards depending on how the biotech industry goes. Usually, when you see an uptake in funding, this takes 2 to 3 quarters to actually materialize. Why? Because companies need to review their development plans, issue RFPs, and then CDMOs have to come back. So this takes time. And the earlier the phase, the smaller these amounts are. There’s a lot of redevelopment work and batch manufacturing that is involved in this. But, it’s usually a healthy sign for the industry.”

Also Goldman sees attractive growth in the coming years:

“In EU Life Science, our analysis suggests that end markets in Biologics continue to look attractive with new therapies and new classes driving growth. In addition, recent news flow suggests US policy risk could have lessened, which, despite the EU names not having significant exposure here as a percentage of revenues, is a positive for sentiment. We continue to believe the underlying Bioprocessing market is in a strong position to deliver growth in the mid-term.”

So, the question now becomes where to position? We never invest in the drug manufacturers or biotech names themselves. This is a hugely complicated and competitive landscape where only highly specialized analysts that read medical publications on a daily basis are well positioned to make stock picks on which drugs they’re going to invest in.

While we’ve seen some investors argue that patents provide a ‘formidable moat’, we note that the effective time that a pharma company can exclusively sell a particular drug molecule is only 10-15 years. So, pharma and biotech companies continuously have to face patent cliffs with generic competitors coming in to sell a discount version of the same molecule. In our view, pharma companies have to continuously reinvent the wheel in order to find new revenue drivers as their products go off patent. At the same time, drug development costs keep exploding and it becomes ever harder to actually successfully bring a drug molecule to market. At the same time, competition is innovating as well and can bring to market new drug molecules with better efficacy or side effects profiles, posing a continuous risk of disruption for even gold standard treatments. This is something that Novo Nordisk investors found out last year.

So, we prefer to invest in the picks-and-shovels providers supplying the plethora of competition in the biotech and pharma space. These are industries which are much more similar to other tech industries such as semiconductors for example, where you regularly have a limited number of leaders that dominate a particular niche. At the same time, you get exposure to rising drug development R&D which provides an automatic tailwind.

Post the JP Morgan conference, we came across a handful of highly-specialized niche leaders and with attractive market positions that we haven’t reviewed before. In this article, we will dive into the first name. This is a smaller company, but with high margins and with high long-term growth potential, including some opportunities that the market is missing. Finally, we’ll also discuss our current biotech picks & shovels portfolio.

A Niche Leader in the Biotech Supply Chain

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

“Every time we bring capacity online, we sell all of it. With the new generation of NVIDIA Blackwell coming online, more customers are interested in purchasing capacity in advance and securing it for a longer period of time. Today, we are very pleased to announce that we signed another major deal, this time with Meta for approximately $3 billion over the next 5 years. In fact, the demand for this capacity was overwhelming and the size of the contract was limited to the amount of capacity that we had available, which means that if we had more, we could have sold more. This deal comes on top of the Microsoft deal we announced early in September with a contract value between $17.4 billion and $19.4 billion. As we said before, we expect to sign more of these large long-term deals, and we are delivering that promise.

As busy as we are with these mega deals, our main focus is still to build our own core AI cloud business. We made great progress here with AI native start-ups like Cursor, Black Forest Labs and others. In order to meet the growing demand, we have accelerated our plans to secure more capacity, and this is actually our main focus for now. Capacity today is the main bottleneck to revenue growth. And we are now working to remove this bottleneck. As we look to 2026, we expect our contracted power to grow to 2.5 gigawatt contracted. We plan to have power connected to our data centers, which means fully built of approximately 800 megawatts to 1 gigawatt by the end of 2026, by the end of next year.

We are also investing in our main product, our AI cloud. To extend our addressable market opportunity to large enterprise customers, we released our new enterprise-ready cloud platform version 3.0 called Aether and our new inference platform called Nebius Token Factory. We believe Aether gives organizations the trust, control and simplicity they need to run their most critical AI workloads. Nebius Token Factory is a production scale inference platform that enables organizations to run open source models with reliability, visibility and control. And we have a large pipeline of new software and services that we are continuing to build, which will differentiate us from other cloud companies. Based on the strength in demand that we see and our accelerated capacity growth plan, we believe we can achieve annualized run rate revenue ARR of $7 billion to $9 billion by the end of 2026.

When we plan for data center CapEx, there are actually 3 stages. The first stage is securing the land and power. The second stage is building the data centers themselves including shell, electrical and cooling equipment, batteries and so on. And the third part is finally deploying the GPUs themselves. And it breaks into 3 spending blocks. First stage, securing land and power, it’s pretty cheap. It’s around it’s around 1% of total CapEx. The second stage, building the data centers and building the connected power, is something around 18% to 20%. And the remaining 80% main part, is for deploying the actual GPUs.

So, what should we do? First, we should secure as much capacity as we can because the cost is immaterial at this scale. Second, we should build as much as our capital allow us. And third, we will fill GPUs in line with contracted or clearly visible demand. This massive 80% spend will come only when we see real demand. That’s why we say that in 2026, we will be securing 2.5 gigawatt total contracted capacity. And we are planning to physically build 800 to 1 gigawatt of connected data centers. This will be done by the end of next year.”

Assuming that the 0.9GW they’re targeting by the end of ‘26 includes having the GPUs online, this would mean that the $8 billion ARR roughly translates into $9-10 billion of revenue per GW of GPU capacity. This number is in line with CoreWeave, but below what the major clouds are doing as neoclouds have to discount in order to attract customers. The bull case to increase this number over time would be: 1) increase the number of web services available on neoclouds over time, e.g. object storage, vector databases, potentially nosql databases, etc. to turn their data centers more into a real cloud where enterprises can run an entire app on. 2) Diversify the customer base to move away from large contracts with the hyperscalers and build up a real client base of enterprises and successful AI startups/ winners themselves.

We’ve written extensively about the strengths of the big three clouds, i.e. Amazon AWS, Microsoft Azure, and Google Cloud Platform, and we’ll do a brief recap here for new readers. The attraction is that enterprises leverage all sorts of software tooling on these clouds in order to deploy their workloads in the cloud. As they replace their old, on-premise data centers – due to the large amount of software engineering involved to manage a cloud – they become effectively locked in on these major clouds.

Enterprises try to reduce this risk by pursuing a multi-cloud strategy, however, due to the high amount of software engineering involved, they typically spread their workloads over two clouds. For example, 70% on AWS and then 30% on Azure. This also makes sense from an economic standpoint, as a cloud provider will give you discounts for a large, multi-year contract. So, an enterprise customer wants to have some leverage by deploying 30% on Azure for example, and then also get a large discount from AWS by deploying most workloads there.

Major clouds have the entire software stack available you need to deploy apps. I.e. CPU-based VMs, databases (SQL, Firestore, DynamoDB etc.), data pipelines (Dataflow, Kafka, etc.), data analytics (Fabric, Snowflake, Databricks etc.), security, access management, load balancers, networking, caching databases, object storage, edge network caching, etc. You can see below some of our favorite products we have pinned in our Google Cloud Platform console, which makes it easy to deploy and manage all your cloud workloads:

Very little of this is available in Neoclouds, these are mostly platforms where you can rent GPUs. CoreWeave has built some neat scheduling software so that users can deploy their AI workloads on a fleet of GPUs with near bare metal performance. However, all the larger clouds are extremely good at writing these sorts of abstraction layers. CoreWeave’s Slurm adaptation (SUNK) for example is built on Kubernetes, which was open-sourced by Google and based on Google’s internal orchestration system called Borg. Nebius has built a very similar system as SUNK called Soperator (Slurm operator) and open sourced it in late ‘24. This means that basically any cloud now can adopt a Slurm-on-Kubernetes system like CoreWeave’s. This system basically makes it easy to manage your fleet of GPUs and even switch between training and inference workloads, so that the GPUs never sit idle.

When some investors look at the type of analysis like below, they conclude that – ‘The big three clouds are behind; CoreWeave, Nebius, Crusoe, etc. have the advantage, and as these neoclouds are growing revenues much faster and winning big contracts, the risk-reward is very attractive in these names’. More extreme viewpoints include that the big clouds will get disrupted by these neoclouds. However, SemiAnalysis is purely comparing GPU cluster performance here, and then a lot of investors mistake this as an assessment of overall cloud capabilities.

The reality is that the two types of business are very different. As both types are being labeled as ‘cloud’, this causes confusion as a lot of non-engineers don’t properly understand what’s under the data center hood. Even if you want to run an AI-centric app such as ChatGPT or DeepSeek in the cloud, the reality is that you heavily need one of the major clouds as an AI cloud won’t do. You need access to databases, analytics platforms, load balancers, caching databases, elastic VM services, observability, security, networking, IAM, etc. What you can do on neoclouds is send an API request to the GPU, so that a GPU can analyze your prompt, run this through various linear algebra algorithms on its CUDA cores, and then send back an answer.

Most enterprises take up as much capacity as they can in the major clouds as it allows them to run their GPU VMs alongside where their data is stored. It’s very expensive and time-consuming to move large amounts of data out of these clouds due to high egress fees and bandwidth constraints. This is why winning contracts with the big clouds presents so much upside to the neoclouds, as neoclouds are only gradually winning their own customers. So, most of neocloud revenues are currently coming from hyperscalers where end customers are not even aware that their GPU workloads are being offloaded onto neoclouds. For example, customers connect to Azure over API and these requests are then sent to CoreWeave data centers as Microsoft is short on GPU capacity. This piece of code below shows what an API call looks like:

The only exception here we’ve seen amongst the neoclouds is Nebius. Nebius was born out of the Yandex cloud, and has a more comprehensive software stack available which normally would take a company years to develop (overview below).

Nebius’ new cloud console will look familiar to AWS, Azure and GCP users, although still much more limited:

However, it will be hard for Nebius to compete for these types of workloads with the big three. One of the main problems for Nebius is that the big three have data centers all over the world and even multiples in every region. Nebius has its flagship data center in Finland. So, in Azure, AWS, or GCP; I could run VMs and duplicate my data in Virginia, Ohio and Oregon, thereby offering 100% uptime and low latencies to users across the US. If one of those cloud regions goes down, e.g. Ohio, you still have the others up and running. And as my international traffic travels over Microsoft’s and Google’s premium global backbones, I get low latencies to users all over the world really. You could even host your VMs and relevant services in Asia and Europe as well if you wish. It’s obvious that this is an extremely premium infrastructure both from a software and hardware standpoint, and a massive competitive advantage if you deploy your apps here vs let’s say in Nebius.

CoreWeave is well aware that these types of software services can increase customer lock-in and bring in high-margin revenues for the company. From the UBS conference:

“We are continuing to innovate right? We are developing new products and services that fit within our AI cloud and what customers will increasingly need in the future. Take things like AI object storage, that’s a product that fits into our broader storage business and that we announced in the third quarter, and has grown to north of $100 million of ARR. It’s growing like a weed. We are deeply entrenched with our customers on a technical level and we understand what their pain points are. We used to live in a world where you were a single cloud customer, an AWS customer. I think you are now increasingly an AWS customer, and an Azure customer, and a CoreWeave customer. Storage was not built for that world. You had data lock in, you had high latency, and if you want to move it to another cloud, you had high egress fees and transaction fees. We built a product that’s low latency with no egress or transaction fees, and customer adoption and attach has been very attractive. We are going to keep building those types of products and services.”

In terms of these types of service offerings, CoreWeave looks to be quite a bit behind Nebius. Object storage is nice but it’s really the most basic of all the web services. It’s one of the very first services that AWS launched 20 years ago. However, a bull case would be that CoreWeave, and also Nebius, can keep building out those types of software services to get access to high-margin revenues besides pure GPU rental.

For premium subscribers, we will dive further into the economics of neoclouds such as Nebius and CoreWeave, including with some detailed financial modeling. We’ll give our thoughts on whether these are attractive investments, and how we are positioning.

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

In Gave’s view, economists mistook China’s pain during this period for a Japanese-style stagnation. During this time, China built up impressive industrial capabilities in areas such as EVs, batteries, solar & wind power generation, nuclear reactors, robotics & automation, industrial equipment, rail, rare earths, etc. However, as from 2025, the Chinese government started stimulating the demand side of the economy. China is running 10% budget deficits of GDP, pushing credit to consumers and real estate again, and actively supporting equity markets. “China is shifting from being the world’s deflationary shock absorber to a reflationary impulse for the first time in decades.” As a result of its industrial capabilities, China now runs a $700 billion annual current-account surplus.

China’s other strength is that the country has an army of engineers and scientists: “I went to China to university in the mid-90s and back then China was graduating 350,000 university students a year. Today, China graduates 12 million university students a year. And by the way, roughly half of these guys are engineers and science graduates.” Due to China’s strong engineering workforce, there will continue to be strong opportunities in the coming decades in the country’s tech, industrial and life sciences sectors as a few examples. To a certain extent, we’re already seeing this play out with stellar stock performances over the last 5-10 years in certain semiconductor, semicap, software, biotech, EV and battery names.

The most credible pushback to investing in China is the geopolitical risk. When it comes to a war around Taiwan, Gave’s view was that “The closer you live to Taiwan, the less anyone worries. People worry about it in Frankfurt, Dallas or Toronto.” Gave mentions that Beijing has every incentive to wait. It’s likely that the KMT wins the next election, which will open the possibility to negotiate a reunification with a 100-year glide path. Although they would need two-thirds of parliament to close such a deal.

For example, an agreement can be made that Taiwan joins the Mainland 50 years from now under a “one country, two systems” framework like there is in place with Hong Kong. This means that in 50 years, Taiwan will officially join the Mainland, however, it will continue to have its own government, police force, legal system, etc. Then, again 50 years later – so 100 years from now – reunification will be fully completed.

Gave noted that the CCP could have already taken Hong Kong much sooner if they really wanted, for example, as Hong Kong’s water and electricity are supplied by the Mainland. In the decades after World War II, the power of the British Empire started to wane and the empire gradually dissolved. However, China chose to wait and negotiate, until finally an agreement for reunification with a long glide path was made. He sees a similar scenario playing out with Taiwan. The people who will sign off on this deal won’t be alive anymore when it gets implemented, which minimizes risk of personal blowback. What happens 100 years from now is out of everyone’s horizon, and so there is generally little pushback to resolve the situation this way. The British recently gave away a highly strategic island chain in the Indian Ocean, which Trump was fine with, as the US got a 100 year lease on its Diego Garcia base (with a large runway for military aircraft). Note that the British originally also only negotiated a 100 year lease for the New Territories with China as it seemed “forever”. And so, China gets what they want in this scenario, an agreed reunification with Taiwan.

In the 2010s up to 2021, we regularly made investments in China as we came across good businesses with high growth rates and at attractive valuations. However, we sold our positions in ‘21 as valuations became high during the covid tech boom, and while it also was becoming clear that Xi was going to crack down on the names we liked, in order to constrain their power:

Since then, as Chinese equities went into a prolonged bear market, Gave’s view is that “everyone who was going to sell has sold now already”. So, the current shares are now held by steadier hands.

Over the last year, we’ve gradually taken positions in a number of Chinese names again, simply as we thought valuations had become extremely cheap while there are attractive companies in this market. Generally speaking, our typical investment criteria in any market are:

1. Companies with a dominant or high market share in their niche, or which have a compelling competitive advantage such as strong innovation or engineering capabilities.

2. An ability to generate high growth rates over the long term.

3. A valuation that gives investors an attractive risk-reward to the upside.

We’ve found this to be a solid framework to catch multi-baggers over the last few decades. We don’t invest in commoditized businesses or those that are operating in highly competitive environments. While value investors are often drawn to these types of businesses due to low multiples, they frequently underestimate the downside on EPS due to competitive pressures. We’d rather pay a premium on top of deep value multiples in order to get earnings which are much better protected. These are the types of businesses you can buy and hold for long periods of time. Good examples are Google, Amazon, Meta, etc. Same goes for semi names, although it’s best to take profits here when the cycle starts getting stretched.

In this post, we’ll go through a number of Chinese businesses that we like and one Japanese name that we studied this week. Some of these companies are already great, and others can become great five-ten years from now. We are seeing similar opportunities in the market like Nvidia ten years ago, where we first invested in the name in 2015 as we thought that AI – or machine learning, as AI was called back then – could become really big one day. Remember, this was a time when models were still being trained on gaming GPUs, or CPUs even. A PhD training a model on five gaming GPUs in his college dorm was state-of-the-art in those days. A time before Nvidia had sold its first data center GPU..

Finally, we’ll also discuss our thoughts on the outlook in AI and future market share shifts between GPUs and ASICs.

Read more

And subsequently Gemini-3’s performance on an advanced physics test was also a clear outlier:

The main strength of LLMs so far has been memorized knowledge. So, they do very well on benchmarks that test what the AI already knows, such as SAT or bar exams, or writing smaller pieces of code (e.g. 100 to 500 lines). With Gemini-3’s more advanced reasoning capabilities, it should perform especially stronger at more complex tasks such as understanding larger code bases, data sets and libraries of scientific work, with an ability to then generate novel and innovative hypotheses.

The other important part to focus on in the first ARC-AGI-2 chart is the x-axis, which illustrates that Gemini-3-Pro can do this more complex reasoning at a similar $ cost than Claude Sonnet or ChatGPT-5. Gemini-3-Deep-Think has even more advanced capabilities, however, its $ cost is currently also extremely high. So, for the moment, it will only be applied for more selective workloads that really need and can afford these capabilities. However, Gemini-3-Pro shows that pretty advanced compute is already widely accessible.

Google also released its new IDE (coding editor), Antigravity, to integrate Gemini-3 directly into coding projects. We spent several days this week coding a native mobile app in Antigravity and it definitely has a lot of innovative features. However, it’s also still fairly problematic. Overall, the codebase it wrote is extremely well organized, the code is high quality and secure, and our MVP app is getting close to ready.

What’s really good is that Antigravity creates a coding plan at each step, which the developer can then correct or validate. This way, it avoids dumb setups being introduced into the codebase which can cause a lot of problems later. However, the main problem is that Gemini-3 still regularly crashed and then we had to restart it to try again. Occasionally, it still made a few dumb mistakes but these could easily be corrected via the coding plan. However, all the feedback we gave to the coding agent could likely be automated by another smart coding agent. So, it’s fairly easy to see how in the coming months agents can start taking off to work independently for hours on coding projects and produce useful work.

Overall, having been using both Antigravity and OpenAI Codex in the past few weeks, our long term outlook on AI agents and the amount of useful work they can produce is definitely bullish.

We’ve always been fans of Google. We first invested in the company in 2010 and so this has been a 20x bagger or so. However, we’ve been reducing our position this year as 60%-plus of its profits stem from traditional Google Search, and it looks very likely to us that gradually this business will get disrupted in the coming years. The company’s bread-and-butter monopoly in retrieving 10-blue-links will increasingly become a legacy business, while in next-gen AI interfaces the company now has to compete with a number of strong competitors such as Meta Llama, Grok, ChatGPT, Claude, Perplexity, Qwen, DeepSeek, Mistral, etc.

Thus, as the stock was getting pricy again at close to 28x forward PE – despite the company facing disruption risks in its key business – we recently halved our position. However, this week, we got a real showcasing of the bull case in Google. Obviously, the company has a deep bench of strong AI researching talent, combined with some of the widest access to AI training compute. In addition, it’s the only company which has successfully built a full AI stack including its custom accelerator, the TPU. While estimates vary, Google should have a strong cost advantage thanks to their ability to train models on their custom TPU clusters. Firstly, Google doesn’t have to pay Nvidia’s 75% gross margin, and secondly, TPUs consume about 30% less power for equivalent workloads.

Currently, it definitely looks like OpenAI is falling behind. ChatGPT-5 is a fine model but wasn’t really much of an improvement vs GPT-4. They basically added a simple LLM in front of it which then decides whether the query should be routed to the ‘fast’ or ‘thinking’ model. While some tried to hype it as a major breakthrough, frankly, it’s something we could have easily coded ourselves. At this moment, it doesn’t look like OpenAI has made much progress in the last year in terms of model improvements. And that might be fine, the real money will be made by the players who can successfully insert these LLMs into consumer apps or to automate workloads in the real economy. Being the best at solving science problems will likely be more of a niche application.

Anthropic is a good example here of how to turn an LLM into business success via a simple API. While the capabilities of its Claude model are not hugely impressive on benchmarks, it remains by far the most widely used model by enterprises to generate code. We use it as well, and its great attraction in our opinion is that it only implements code what the developer asks for. Gemini-2.5 and ChatGPT can’t really be trusted alone with a codebase as they might start changing all sorts of things that can introduce problems. Similarly, Grok regularly crashes or starts reasoning for 10 minutes on end. Claude delivers reliable and good code quickly. So, we would rate Claude as the most ‘stable’ model in coding and as a result, its annual revenue run rate is approaching $7 billion now (vs $1 billion at the start of the year, a 7x increase) and at healthy gross margins.

Next, we’ll dive into the current market turmoil and our outlook for AI, Nvidia, and stocks we’d be taking profits in here.

Read more