Leading Edge Semi Shortages - a deep dive techinvestments.io/p/leading…

Former data scientist at Meta explains why transformers improved their ad revenues so much (Tegus):

How TSMC's pricing exactly works:

Former semi engineer at Google and Groq says Nvidia's moat in training is unbreakable for the foreseeable future, but the market in fine-tuning and inference will be more competitive. AMD doesn't have a chance in training and for inference & fine-tuning, hyperscalers are moving to ASICs. (source: Tegus).

Product manager at Microsoft Azure says their $AMD MI300 fleet isn't being used much, $NVDA dominates completely with a more than 90% market share in the installed base (via Tegus)

Very bullish TSMC conference call on AI: "Even after more than tripling in 2024, we forecast our revenue from AI accelerator to double in 2025 as the strong surge in AI-related demand continues. We now forecast the revenue growth from AI accelerators to approach a mid-40% CAGR for the 5-year period starting off the already higher base of 2024." Definitely staying long $NVDA $AVGO & $MRVL here Also bullish for semicap. As we predicted, TSMC had been overly cautious on its capex guidance and is now going for peak capex in '25. And given the capex-to-revenues ratio then of only 36%, there will be much further upside in the years thereafter. Especially as we head into the Angstrom era, and TSMC has to build both in Taiwan and the US due to the geopolitical climate: "Our plans for second fab and third fab in Arizona are also on track. This is where we'll utilize even more advanced technologies such as our N3, N2 and A16 based on our customers' need." Bullish for $ASML $AMAT $LRCX $KLAC $ASM

Microsoft Director says that $AMD's GPUs are now a good source for inference workloads (Tegus interview):

ASML can easily make gross margins in the 60 - 65% range if they want to, they’re just far too generous..

AI architect at Google mentions the lifetime of datacenter GPUs at current utilization levels to be 3 years at most. This is v bullish for $NVDA end-demand as most analysts were assuming a lifetime of around 5 years.

Cloud customers are sticking with $NVDA's GPUs for inferencing as they like the CUDA stack, so $MSFT is struggling to keep up with demand (Tegus):

Contact at Microsoft says they've now uncancelled three data center projects which they had cancelled earlier in the year. The plan is to build more than 100 data centers each year for the coming four years. (Tegus interview late March)

Manager at Amazon says an $NVDA A100 replacement cycle is coming - will be being replaced by Blackwells soon. (via Tegus - showing parts of my notes below from his comments)

Engineer at Apple says that 30% of their AWS cloud workloads are already over Amazon's Graviton CPU as opposed to $INTC or $AMD. He thinks start-ups will go 100% over Graviton as AWS is heavily promoting its own CPU, telling customers that Graviton is 35-40% more cost effective than Intel CPUs. Going forward, he sees workloads over Graviton increasing. Historically they haven't done so as they like to be able to shift workloads over different clouds e.g. if AWS goes down, start up the workload in GCP. But with Docker/ Kubernetes-based containers, this additional software layer can abstract away from the local hardware you're using. So if you're running apps in multi-architecture containers, you can easily shift apps between clouds while still making use of the local cloud optimized ASICs such as Graviton. So basically with multi-architecture containers, you get the best of both worlds i.e. the software flexibility to switch clouds and the benefits of cheap localized ASICs.

Former Intel manager on why Intel will struggle against TSMC in foundry. Expert available via Tegus: $INTC $TSM

$NVDA scientist on why he isn't worried about competition from inference ASICs, basically you want the flexibility of the GPU to deal with model changes going forward. We're running out of new data to train on, and so innovation will have to come from model changes. Only a flexible GPU can deal with this.

Goldman buy thesis on $ASML: "We believe the pause in Lithography intensity is temporary, and as such see an increase in EUV exposed layers in the subsequent nodes post the 2nm process (given it is materially difficult to execute an architectural transition in tandem with incremental shrinkage). Additionally, we believe that any negative impacts from this temporary pause on the Logic side would be offset in future by accelerating EUV adoption in DRAM as customers accelerate layer additions leveraging the 3800E tool. Further, we see a clear path for EUV penetration to rise further as DRAM roadmaps progress from 1β to 1γ nodes and beyond. Meanwhile, on the Logic side, as GAA matures, customer focus is expected to shift from architectural transition toward continued EUV shrink, supporting a renewed step-up in lithography intensity. As such, we continue to see Lithography as a key strategic technology to manufacture smaller and more powerful chips which would be needed to fulfil the increasingly complex compute requirements going forward. ASML recently stated that it expects a gradual normalisation of its China exposure in coming years, with Chinese sales contributing around 20% of total revenues going forward vs mid 20s today. We note that management’s confidence is underpinned by healthy demand due to an emphasis on domestic production in areas such as Automotive. While management expects China demand to soften in 2026, we view the SMIC capacity additions as a positive driver, albeit SMIC’s ambitious approach on 7nm Logic could to some extent be balanced by a lack of access to EUV (and a long tail of smaller Chinese players has seen spend dampened to some extent by reduced local government activity). Further, on the positive side, incremental investment from Memory players should also impact demand levels positively, though the US redefinition of advanced DRAM has slowed progress on cutting-edge memory ramps. More broadly, we continue to see healthy revenue and GM margin contributions from China as the region continues to develop self-sufficiency in the semis space. We note that recent datapoints from our GS Communacopia conference point towards continued strong demand for leading-edge semis specifically as a result of AI. Furthermore, we view the Samsung and Tesla deal as a positive for ASML given it broadens the Lithography customer base (assuming Samsung is successful) and lowers ASML’s dependence on one particular customer."

$ASML's rally explained - full article in bio

Director at Intel on which stocks in the semi industry he'd invest in (via Tegus):

$NVDA is widening the R&D gap with its closest competitor. Annualizing last quarter's R&D spend, NVDA is investing now almost at twice the level of $AMD. And the latter will have to split R&D between GPUs, CPUs and FPGAs, whereas NVDA can fully focus on building out the AI stack.

This is a credible long term bear case I've come across for $NVDA. The hyperscalers - from the large public clouds to the big internet giants - have an extremely strong position and the means to move to internal silicon over time. This will take time i.e. years and years, but as Google has shown by successfully training LLMs on their own TPUs, there is a way and now the others are following in this path. The key for NVDA will be to leverage their know-how and annual $10+ billion R&D budget to stay ahead and keep a high market share in this market. But long term, I suspect it won't be easy and that gradually the hyperscalers will be able to transition more AI workloads to internally developed silicon. Positive for $AVGO / $MRVL, or least for as long as the hyperscalers need them.. A director at Qualcomm gives further details:

Former Samsung & current Intel manager says $INTC's 18A is 'way too good', he also gives his views on Samsung's problems at 4 and 3nm (on the right)

$MSFT datacenter architect on the software abstraction layer they've built to allocate AI workloads between $NVDA and $AMD GPUs. And how ROCm is narrowing the gap with CUDA. AMD is still largely being used for inference, they have a HBM advantage here, whereas NVDA's GPUs are really optimized for training with a higher compute to memory ratio.

UBS on $NVDA's revenue outlook - the most bullish I've seen: "The company noted on its FQ1:26 earnings call that it has visibility into "tens of gigawatts" of AI infrastructure projects in the "not too distant future". Assuming a "low case" pipeline of 20GW and NVDA's stated range of ~$40-50B per GW, this puts its total revenue opportunity for this pipeline at a minium of ~$1T. While the company did not specify a timeframe for this pipeline, based on our conversations, we believe these projects are likely to be rolled out over a 2-3 year period. Using the average of this timeframe, this suggests the company may effectively have "visibility" to ~$400B/yr in data center revenue, or about 2x our $233B data center revenue estimate for C2026. This is obviously very heady, but we did note in UBS' deep dive on OpenAI's Abilene AI Factory that Crusoe alone has ~20GW in project pipeline and this is just one digital infrastructure project developer."

These two Tesla engineers should get a big fat bonus

UBS buy thesis on $ASML "Following a year of underperformance with the stock down c20%, we see the declining lithography intensity thesis and overhang from the Chinese market as well understood by the market. Given ASML's long product lead times and high level of integration into customers' long-term roadmaps, the market is likely to look through a relatively well telegraphed weak 26E to 27E, when we see the return of ASML as a quality compounder delivering 20% EPS CAGR 26-30E. We upgrade the stock to Buy. Following 2025-27 EPS revisions of 20-30% in the past year, we believe our Neutral thesis on litho intensity is well understood by the market. We had argued that litho intensity would resume, which we believe should start from 2027. We estimate the key driver will be production ramp of A14 in logic at TSMC, with our bottom-up model suggesting an uptick in EUV exposures from 19-22 to 20-24 for the A14 node. We also see uncertainty around Intel/Samsung fading, either with a revival or TSMC taking share. The A14 production ramp will also deliver the adoption of a small number of High NA layers. This will likely be in production in 2028-29 for A14P (2nd gen) but should come through in backlog in 2026/27 given two-year lead times, before A10 mass volume ramp in 2030-31. Our analysis of Low vs. High NA production ramp timelines and Intel's progress as a first mover suggests meaningful adoption in the next two years is highly likely. We forecast 6/10 High NA shipments in 2027/28, representing c30% of group revenue growth in 2027-28E or 1-3 percentage points of share potential gain of WFE. We expect 2026 guidance at Q3'25 or Q4'25 results to be a clearing event, allowing the market to look through to 2027. Whilst this is a long-term story, we see catalysts over the next year: 1) Clarity on incremental EUV exposures and High NA insertion for logic at events like SPIE Advanced Lithography Feb-26; 2) Commentary on High NA adoption from ASML at Q1-Q4'26 results; 3) Launch of new low NA EUV-F model with higher ASP in H2'26; 4) New customer announcements at Intel/Samsung through 2026."

Some thoughts on $NVDA's CUDA as a moat and what the actual moats for the company are: A lot of analysts have the impression that AI coding is happening in CUDA, which would give Nvidia a strong moat on the software side, but this is unfortunately not the case. Practically all AI code is written in Pytorch, which can run on top of CUDA, CPUs, and now also increasingly on AMD’s ROCm GPU computing platform. If you want to run a Pytorch model on another GPU, you have to use a compiler provided by the manufacturer or alternatively, write your own. Both Pytorch and Hugging Face, i.e. the community to share AI models, are very motivated to extend support for ROCm as well as other GPU computing platforms as naturally they don’t want the future of AI to be controlled by a single manufacturer. So why is everyone then still buying Nvidia GPUs if we can now run our Pytorch model over AMD GPUs as well? Nvidia’s GPUs are simply still way better at running AI workloads. Additionally, as LLMs have to be deployed over multiple GPUs, Nvidia has both the networking hardware and software libraries to easily do this both in training and inferencing. So CUDA won’t be a sufficient moat in the long term to stay ahead of competition. In my view, what Nvidia really needs is a fast pace of innovation in the AI ecosystem, both on the software and the hardware sides. The company is now spending $8 billion annually in R&D to advance the Nvidia ecosystem both on the hardware and software sides, and there will be very few players in the world who will have both this budget as well as the required know-how and expertise to even have a chance of competing with this. If the pace of innovation slows down, it will be easier for competition to narrow the gap. For example, if the evolution in LLM architectures becomes predictable, hyperscalers will be in a strong position to offer custom silicon to address these workloads. Similarly, if Moore’s law and more-than-Moore technologies start proceeding at a slower cadence, competitors and hyperscalers can more easily build out their ecosystem of competing hardware and accompanying software libraries. A final moat for Nvidia is that the company is moving ever deeper into the AI training stage, where the goal of its Omniverse software platform is to provide a physics-enabled virtual environment for AI training. This is particularly useful for reinforcement learning type tasks, Mercedes has been training its autonomous driving system in Omniverse for example. So you can integrate the latest AI transformer technology in your models for example and then improve it further via reinforcement learning inside Omniverse. Both robotaxis and humanoid robotics could become vast markets in a decade or so, and as Nvidia both provides the physics training ground combined with the GPU platform that goes inside of the robotaxi or robot, the company is in an interesting position to play an important role in this value chain.

Barclays buy thesis on $NVDA: "When tracking AI capacity additions over the LTM, AI TAMs don't seem so outlandish anymore and NVDA looks like the most interesting name in our group. When Jensen first forecast a $1T industry by the end of the decade, we admittedly balked. With the wave of announcements that have come over the last 6-9mo, we now estimate over $2T of planned spend at ~40GW of power in total. Within that, we attribute ~65-70% to compute & networking with more deals likely in the pipeline, which starts to make the updated guidance of $3-4T look much more real. Given the variation of data available for each project, we used both the conversion of 1GW=$50-60B of total spend and used the more recent 1M GPU per 2GW associated with the OpenAI deal announced last week. When summed, this equates to $1.5T of compute & networking spend and 19M GPUs, which we acknowledge isn't perfectly pro-forma. We also acknowledge that some of these dollars will go towards custom silicon but in the tracked announcements thus far there is little reference to any specific programs. We introduce an AI capacity tracker that aggregates announced compute deployments, power, and chips that we will update real time and offer to clients. We see this as a positive for all accelerator names (AVGO + AMD) but we see this largely flowing into the NVDA P&L over the next 5+ years, moving numbers materially higher and making this the most attractive name in our space. We move our price target to $240."

Manager at Bloomberg on why they're going for the $NVDA stack and not for any of the competing solutions:

Bloomberg has been trying $AMD and $INTC GPUs but is sticking with $NVDA for 95% of their workloads. Their H100 demand is not fulfilled, they are still trying to get the H100s they need. Expert made available via Tegus:

Divisional head at AWS says the ROI is definitely there for lots of enterprises to invest in GenAI, gives a number of useful examples (interview on Tegus):

Deutsche buy thesis on $META: "We have been here a couple of times before, and each time it ended the same way... good for META's longer-term equity value. This time, Meta is accelerating its investment in AI/infrastructure faster than we anticipated, which will likely result in Capex being much higher in the near-medium term. Now, investors are clearly skeptical again and our FY26/27 free cash flow estimates come down by ~(40)%/(30)%, respectively. That said, this time it's slightly different from prior investment rounds in our view . To start, this time around, the company is building out infrastructure, which should supplement the current advertising model, vs. prior cycles which typically resulted in near-term cannibalization of better monetizing, though aging surfaces/formats. In fact, the FoA business is fundamentally very strong at the moment, with core AI investments driving continued engagement growth, and the ad platform becoming even more performant. As such, as the fastest growing ad platform at scale, we argue Meta is still leaning into its growing Gen AI ambitions from a position of strength. On the other hand, some will argue that META has yet to scale any real tangible scaled new products from this growing investment cycle. We argue that the infrastructure is somewhat fungible and can be leveraged to scale the ad platform even faster in a worst case scenario. In our view, while the scale of these ambitions (and investments) is vast, and requires significant investments that have a longer- dated payback horizon, we contend that these capital investments are also having strong returns in terms of engagement and ad performance. This is translating into Meta's current durable advertising revenue growth. Simultaneously, it is affording the company the ability to strategically lean in to foundational AI model training work, which could unlock vertically integrated opportunities across large TAMs including business messaging, Meta AI (potentially Search Advertising), and scaled wearables. All in, we lean on the company's track record of high ROI investments which have, in the past, resulted in products that drive scaled consumer adoption. Perhaps more importantly, given the reasonable valuation and fundamentally better advertising platform that’s gaining share, we argue it seems prudent to invest with Mark Zuckerberg's growing AI ambitions now."

$NVDA dominates the datacenter

Former director at TSMC discussing the semicap space, especially Tokyo Electron - source: Tegus

One stock I've been thinking regularly about over the past year is $GOOGL. I still have my long position and the bull case I can see is that with better search results and direct answers being driven by AI, there will be a higher user intent to make use of the search engine as the likelihood of getting an useful answer increases. And then the increase of searches that are monetizable will boost the revenue line. While the increased costs of AI remain manageable due to better ML techniques such as pruning etc which have dramatically lowered inferencing costs. The other part is that Youtube and GCP are just great businesses, and both will benefit from AI as well. Youtube will likely be the biggest winner of AI video generation, with new tools allowing talent around the world to quickly translate engaging storylines into video. This could be a headwind for NFLX, as independent creators can often come up with far more creative content than large corporations. GCP is a great cloud platform and having set up apps in all three of AWS, Azure and GCP in the past, I always thought that GCP has some of the best tooling and documentation. Needless to say, with AI, more computation will take place in the cloud, both for training and inference, and GCP will be an obvious beneficiary. I wouldn't buy GOOGL for the TPU business. While its a fine accelerator and Google has a solid software stack on top, most advanced AI engineers are used to customizing CUDA to get the maximum performance benefits. This is an obvious moat for NVDA. When it comes to nascent businesses, the robotaxi business has much bigger potential. Robotaxis can become a massive market and GOOGL with Waymo is obviously in a leading position. Next to Tesla. The company's interesting capabilities in quantum computing are out of my investment horizon. But it can become an interesting business next decade, in the 2030s or so.

AI engineer at $GOOGL explaining the fine-tuning market and the opportunity for $NVDA (via Tegus):

LLM engineer explains why renting out $NVDA GPUs is such good business for the public clouds - 3 year contracts when you need a larger, multi-node cluster:

$MSFT datacenter architect on the company's cooling roadmap:

Some interesting analysis from TechInsights on SMIC's N+2 node (processor in Huawei's latest phone): One of our salespeople in Asia was able to procure a Huawei Mate Pro 60 and hand carried it back to the lab in Ottawa. When we analyzed the previous N+1 device we found pitches in line with TSMC’s 10nm process but also more advanced features like single diffusion break and 6 track cells not seen until TSMC 7/7+ processes. The overall density of the dense logic on N+1 was slightly less than TSMC 7nm but close and we called N+1 a 7nm class device. TSMC’s original 7nm process was all done with optical multipatterning, the pitches were all achievable with double patterning except the fin pitch that required quadruple patterning. The N+2 Contacted Poly Pitch (CPP) and Metal 2 Pitches (M2P) are both tighter than N+1 but not as tight as TSMC 7nm, CPP in particular is relaxed from TSMC 7nm. CPP is made up of gate length (Lg), contact width (Wc) and gate to contact spacer thickness (Tsp). Lg is limited by leakage, Wc by parasitic resistance and Tsp by parasite capacitance. This indicates to me that SMIC is still struggling to achieve low leakage and low parasitic resistance and capacitance, M2P is much closer to TSMC 7nm. The overall high density logic transistor density for N+2 is intermediate between TSMC 7nm and 7nm+ making it a solid 7nm process. There is even some room to further shrink the pitches with double patterning to achieve something along the lines of TSMC 6nm densities in a future process (N+3?). N+2 is an incremental improvement over N+1 moving from a borderline 7nm process to a solid 7nm process. This process is still within the limits of what optical double patterning can achieve and even has some room for additional shrinks. Full article: semiwiki.com/semiconductor-s… Comparing the Mate60 chip (CN) to the previous one (TW):

LLM researcher mentions that Amazon AWS has worse performance than CoreWeave for AI training (Tegus). The main reason is that CoreWeave runs everything over Nvidia's ecosystem, i.e. infiniband + Nvidia libraries, whereas AWS uses their own networking interface called EFA.

$MU engineer explains how low the yields in HBM currently are:

JPMorgan - $AMD can make $30-35 billion in annual revenue from its deal with OpenAI alone, 'still headroom here for estimates to move materially higher': "For AMD, a starting point for thinking about how to frame revenue per GW is with its recently announced supply agreement with OpenAI, which the company has characterized as “significant double-digit billions of dollars” per GW (which to us suggests at least $15B/GW). Given that AMD’s deployments for OAI commence in C2H26, we think it is reasonable to back into pricing for MI450/Helios based on NVDA’s Vera Rubin platform, which ramps in the same timeframe and is comparable to Helios on a performance basis (though notably, Helios is set to carry 50% more HBM4 capacity - see Figure 5). Given AMD’s historically aggressive pricing for its DC GPUs relative to NVDA equivalents, we think it is reasonable to assume Helios is priced at a 20% haircut to Vera Rubin. Additionally, given AMD will be providing significantly less in-house networking content for its rack deployments than NVDA (only some Pensando content), we assume only 5% attach for networking (vs. 20% for NVDA). We ultimately land at a ~$20B/GW estimate for Helios (see Figure 2), and could see MI500 generating something in the mid-$20B/GW range assuming gen-on-gen revenue growth similar to what we expect for Rubin Ultra vs. Rubin (in the ballpark of +20%). Specifically on revenue from the OAI deal, as much as we do acknowledge that OAI is technically committing to only the first gigawatt of compute, we frankly see little risk to the announced 6 GW of capacity being deployed over the 4-year timeframe outlined in the agreement. On a run-rate basis, this implies $30-35B of annual revenue for AMD just from its deal with OAI, while Street is forecasting a total of ~31B of DC GPU revenue for AMD in 2027 - in other words, there is still headroom here for estimates to move materially higher."

$AMD is trying to mimic $NVDA's products as close as possible - same amounts & types of required cooling, same number of GPUs per pod.. - from AMD's Head of Data Center at Goldman:

The top 10 moats Which one am I missing and which one should go?

Director at Intel explains why ASML has been struggling due to GAA, and will struggle with the move to CFETs as well (via Tegus). The bright spot in terms of order flow can be high-NA adoption later this decade, or EUV multiple patterning, but clearly order flow will be highly sensitive to which technology transition comes when.

Microsoft architect on the company's data center buildout plans - still very bullish. They want to make sure that every data center they're building is profitable when it lights up, while the projects that they cancelled - because they wouldn't be ready fast enough - were immediately taken over by other hyperscalers such as AWS. (Tegus interview from late March)

Memory dominates in $ASML's bookings with 60% of order value. I'm reading that as HBM demand being very strong. Positive for $NVDA.

$NVDA's Jensen confirms the $CDNS and $SNPS bull case: "But in the area of these tools, Cadence and others, they're going to build their own copilots, and they will rent them out as as engineers. I think they're sitting on a gold mine."

A key question for $NVDA is how long the current AI GPU boom will last. The company is currently shipping datacenter GPUs at a rate of around 3.5 million per annum. This translates into a penetration rate of only 13% of global cloud servers having one GPU. Ultimate demand will depend on revenues these LLMs can generate, but it illustrates that we're nowhere near full penetration, as one modern server can carry 4 to 8 GPUs..

A number of insiders at Microsoft have been mentioning performance issues with $AMD's MI300. Below are the comments from a recent interview which gave quite some detail. Basically there are a number of issues with the MI300, so they had to alter workloads as the chips were underperforming. As a result, they pushed out some orders for the MI300, but that being said, given that inference demand is so high, they might have to pull in those orders again. So net-net, could still be positive for AMD given the massive end demand from external customers. Although AMD might have to offer better pricing given the lower performance of their chips. Performance per dollar spent, the MI300 is still better than the $NVDA H200 for inference workloads, but just less than expected.

How I'm currently thinking about investing in AI. Most investors are focused on the box on the left, the infrastructure buildout, however, long term the biggest value creators will be placed in the box on the right i.e. AI software. Similar to how the long term winners from the rise of the internet weren't Cisco and AOL, but Google and Meta i.e. the software apps leveraging the global network. Some of these AI software markets should become huge, such as autonomous driving and AGI or specialized LLMs. I'm probably missing a fair number of names so feel free to add in the comment section..

$GOOGL is processing a huge amount of tokens, and at a very manageable cost, from Barcap: "We break down Alphabet’s recent disclosure around inference tokens. Key conclusions include: 1) AI Overviews are fueling a huge ramp in inference at GOOGL, 2) GOOGL is processing 5-6x more tokens than MSFT (Azure), as Search is ~6x the size of ChatGPT, which is ~2-4x the size of Gemini, 3) GOOGL is using nearly 10% of its total AI compute capex for inference tokens, the bulk (90%+) is still likely used for training new models and powering AI products like recommender systems, etc., and 4) all these inference tokens only cost GOOGL around $750m in 1Q25 using Gemini 2.5's rate card and a few assumptions, hence the rate of deleverage from infusing AI into Search appears manageable (which may come as a surprise to some investors). AI Overview costs are likely around 1% of Search revenue compared to core costs at around 18% of revenue (i.e., core costs to run organic and ads, excluding TAC costs)."

$DDOG is losing its largest customer, OpenAI. From Guggenheim: We are downgrading shares of DDOG from Neutral to Sell and introducing a PT of $105 ahead of potential significant optimization by OpenAI, which we believe is Datadog’s largest customer. Our tactical positioning is based on our belief that OpenAI’s observability software roadmap is shifting towards more cost efficient, in-house managed technologies, and our model setup that indicates 2H risk. We don’t believe this is representative of broader enterprise trends; rather, OpenAI is an outlier in terms of its hypergrowth and associated infrastructure costs that become financially impractical via third-party solutions at scale. OpenAI may have already started to move off Datadog for log management onto its internally built solution, followed by planned deprecation of other Datadog functionalities (e.g., metrics and traces). We estimate this could result in 2H revenue risk, especially in 4Q where we’re modeling 17% growth in our plausible scenario.

LLM engineer explains where they use $NVDA CUDA and where Pytorch:

Former manager at Intel on Gelsinger & 18A - $INTC

Jensen on $NVDA's ability to change fabs if needed - e.g. a geopolitical event happening around Taiwan.

How Google's TPU chip works (source: Google): The core engine is really the matrix multiplication unit, as linear algebra is the backbone of all deep learning models. Bandwidth is often a key bottleneck in how fast the chip can operate, as all necessary data has to be transported from the HBM into the chip itself to run the operations.

$CDNS and $SNPS, the next winners from AI 👇 techfund.one/p/cadence-and-s…

$NVDA's CFO was interviewed by JP Morgan, very bullish outlook, highlights: - PCs are now going to be a big part of that AI era in front of us. We are seeing a need to enable the developers that are working on AI. And NVIDIA RTX already is ready for AI for PCs. We have more than 100 million units shipped of RTX GPUs, that provides a huge installed base for us. - Automotive, you are seeing the use of Orin to be such an important driver of the AI computer inside these cars. That will be followed by Thor. We're seeing new EV OEM wins from Great Wall Motor and Xiaomi, who both will be on Orin. But we also have an additional expansion of new wins with Li Auto, who will be using Thor, and then also ZEEKR. - We're seeing a growing adoption right now of Omniverse for automotive factory digital twins, not only with Mercedes-Benz, BMW, but many other companies. We are also seeing them use it for synthetic data generation. - Early days, folks are changing their infrastructure to deal with generative AI. Our plan is to continue growing supply all the way through the current calendar year that we're in right now. We're very pleased with what we've seen so far in our ability to ramp supply. So our largest challenge had been with CoWoS, a very important part of our process and our Hopper architecture. But we have continued to ramp additional suppliers for that and they have been able to be brought online. We still have the ability to expand that. - We have done a phenomenal job of also incorporating our networking when we are selling our GPU solutions. Sometimes we can be a little bit short on some of the cabling and network we use building out their data centers, some of the first parts that they do even before the GPUs come in. We've worked again in terms of expanding that. - We're still on an allocation model right now, where we are working as fast as possible to meet the expectations when our customers want to receive their computing. - Calendar year '25, that's going to be where we will see some of our new products coming to market. So as Jensen responded to that question, "yes, we can grow" as we move into calendar '25. The market's adoption of AI is just in the beginning stages. This is likely something for decades to come as we continue to grow out and the inclusion of AI in so many different solutions. The accelerated computing opportunity alone is looking at an installed base of about $1 trillion worth of CPUs, or you can break that down to be about a $250 billion per year. What you see going forward though is that size of data center spending will need to address new productivity solutions and will need to incorporate AI and accelerated computing. So that mix of what is just CPU-based and what will be GPU-based / acceleration-based is going to change over time. Even as early as one year ago, we were maybe in the low single digits as a percentage of that market. And now we're seeing a shift as folks are looking at the productivity improvements, the monetization opportunities, the efficiencies of moving to accelerated computing. If you are looking to procure CPUs, you're not seeing that much of a performance increase for the money and the capital that you would have to spend. You're going to see a continued shift, folks shifting to accelerated computing and then also shifting to AI solutions. - Generative AI is a massive TAM expansion on top of both hardware and software. If the CSP invests, let's say $1, they can likely generate $4 to $5 of return just because they have that expertise of setting up that cloud. - We talked about the productivity tools such as Microsoft Office. But you also have how they've infused AI just into search. Search that's been with us for a couple of decades has transformed over and over again with the use of AI and our solutions to help them. - Putting language models in front of all databases will be a very large market. You see SAP, Snowflake, Dropbox, Databricks, all of them are really focusing. - The PC market for many years has been focused on x86. But there is an opportunity for more ARM-powered products. We are focusing right now on ARM in terms of the data center, which is one of the most energy-efficient types of operating systems. - We believe we're really close to that annualized revenue run rate of $1 billion of software. NVIDIA AI Enterprise, is a very important piece of the work that we are doing there. For example, if you purchase our DGX Cloud solutions, it is coming with that full stack. If you are purchasing our solutions with many of our OEM providers, whether that be the Dells or the HPs, again, you have the opportunity to buy that NVIDIA AIE solution from them. - We still have infront of us DRIVE software for our automotive business. Mercedes-Benz as well as JLR will be important folks to watch as we both install that hardware inside of the cars and we work in terms of that software solution that will be available within all of their fleet.

Interesting numbers on 2nm fabrication: $28 billion to build a 50k WPM fab, resulting in a price of $30k per wafer (Nikkei) - $TSM:

$INTC's downfall in the server CPU market & $AMD's rise. The biggest factor is that AMD backed the right horse - TSMC - while Intel struggled in manufacturing.

This is a cloud architect at $MSFT discussing the deal they did with CoreWeave (image below) - it's clearly a legit company, but will obviously be a high-beta play on AI demand. So if AI demand continues to be in excess of GPU capacity, CoreWeave can generate peak margins. However, the drawback is that once the three big clouds - Azure, AWS and GCP - have sufficient GPU capacity, a lot of clients will prefer to run their AI workloads alongside their other apps. It's just much easier from a system architecture / software engineering standpoint. This will cause CoreWeave having to discount its prices to attract workloads, which will impact margins. The safest plays on long term AI demand are clearly $MSFT and $AMZN. They have the scale to build their own ASICs, software stacks, and then mix these with $NVDA GPUs in their fleet. So whichever way the market share battle between ASIC accelerators and NVDA GPUs plays out in the end - e.g. 20-80 or 40-60 - these businesses will generate attractive and growing cash-flows. I'll do some more work on CoreWeave, but my thoughts are to simply keep a large position in MSFT and AMZN in my portfolio. I see CoreWeave more as a speculative play on AI demand continuing to outstrip GPU capacity, which I think will be very unlikely over the long term. Supply tends meet up with demand in the end and then usually overshoots it, especially in semis. The image is from an article, link is in my bio.

$APP - Crocs is reducing ad spend on Applovin. As they kept increasing spend, they weren't get the same returns i.e. ROAS (Tegus interview)

This is the exact reason why Intel has been such an utter disaster. They missed every opportunity to provide the silicon for the new killer apps of the last two decades - iPhone, AI and ASICs. In the meanwhile, they bet everything on maintaining their leadership in manufacturing, a race which they lost to TSMC starting from around 8 years ago. Now, they're faced with huge market share losses in the datacenter CPU market, where especially the cloud part is commoditizing as Amazon is just moving its customers over to Graviton. And in the meanwhile, things aren't looking much better in client CPUs with also now Qualcomm with its ARM-based CPUs entering. A current cash burn of $16 billion per annum is the result, with hopes of breaking even maybe in 2-3 years time. Perhaps the Trump administration can help, by leaning on all the US semi designers to start allocating some product to Intel's Foundry, it's the only real bull case I can see at this stage. What they should have done 10-20 years ago, when they still had the money, is to allocate maybe $600 to $700 million per annum on designing silicon to support all these emerging applications, rather than obsessing about next quarter's margin. Jensen invested in CUDA, Intel obsessed about margins, and it's clear again that tech companies that are focused on R&D are the big winners. A former Intel engineer explains some of the big businesses they passed on:

Head of Cloud Strategy at AWS discusses $ANET

$ASML CEO: "The semiconductor industry is currently working through the bottom of the cycle and our customers expect the inflection point to be visible by the end of this year. Customers continue to be uncertain about the shape of the demand recovery in the industry. We therefore expect 2024 to be a transition year. Based on our current perspective, we take a more conservative view and expect a revenue number similar to 2023. But we also look at 2024 as an important year to prepare for significant growth that we expect for 2025." On the last call, the CEO was still looking for revenue growth in 2024 combined with increased orders coming in the next quarters. So obviously the low bookings number this quarter is a disappointing result (orange bar on the chart) and as a result he's been getting more cautious on '24. Both orders coming from Logic and Memory saw a drop off. The main driver for the next leg of order growth will come from the 2 and 3 nanometer fabs which are coming online in '25, which I suspect he will talk about on the conference call this afternoon.

$AMD has successfully taken a share of the AI accelerator market for supercomputers (data from the Top500 list). With the launch of the MI300 in the second half of this year, this puts the company in a good position to take further share in the coming years. One example here is HP’s ‘El Capitan’, the world’s fastest supercomputer, which should become operational later this year and will be accelerating on MI300s. AMD’s head of datacenter explained their strategy at a recent conference: "In GPUs, we again took a phased approach to attacking the market. And we thought that the most accessible portion of the market was going to be in the exascale, at the very highest end. The software to have the systems optimized was more tractable than in the broader AI market."

Former Director at Synopsys explains the impact of gen AI on the EDA (chip design) software industry - $SNPS $CDNS:

A key reason why $MSFT is winning market share vs $GOOGL and $AMZN in the cloud - it's just cheaper. For example, to rent a Nvidia A100 40GB server in the cloud:

The big upside for $ASML and semicap in general is that TSMC's current capex-to-revenues ratio is still bottoming. In a big investment upcycle (like we will have in the coming years, and possibly till the end of this decade), the market has historically underestimated TSMC's capex plans. And then the same will happen in Memory due to HBM being very capex intensive. Plenty of analysts where bearish on ASML as well in '15-'18, and the situation today looks very similar today. I still like semicap in general as capex intensity will continue to increase at next nodes – in both logic and memory (especially HBM) – while also reshoring has become a big topic with all geopolitical powers looking to secure their onshore fab capacity. And then we have a huge wave of capacity demand coming from AI.

Goldman buy thesis on SK Hynix: "We expect one of the strongest memory upcycles to pan out throughout 2026, as the incrementally higher AI spending mainly from the hyperscalers and their intentions to continue to do so is driving our view that memory demand from servers (server DRAM, SOCAMM, HBM, and eSSD) will signi fi cantly outpace supply. While consensus (BBG and VA) earnings expectations have been rising rapidly since last month, we believe substantial upside remains, with our expectations of 1) new demand drivers in SOCAMM reaching 5%/9% of global DRAM demand in 2026E/2027E, 2) Hynix HBM volume growing close to 50% yoy and meaningfully outgrowing consensus expectations, and 3) almost the entire DRAM supply growth (54bn Gb, GSe) next year to be absorbed by the incremental server-related DRAM demand (53bn Gb, GSe). As such, we upgrade SK Hynix (Hynix) to Buy with our new TP of W700,000 implying 25% potential upside."

Both $NVDA and $AVGO held their respective AI events last week, presenting two opposing viewpoints for the future of the AI datacenter: techfund.one/p/the-ai-datace…

There's a lot in this - who I'd back in the HBM market, my thoughts on why AI clouds should be great investments, AMD's large opportunity no one is talking about (hint: it's not the MI300 or MI350), and KLA's coming semicap ramp! techfund.one/p/the-battles-f… $MU $AMD $NVDA $KLAC $AMZN $MSFT

$ASML's CEO on why high-NA is ahead of EUV at a similar stage of its introduction. For those that remember, EUV was a big mess 10 years ago with continuous market fears that the technology would 'never make it'. EUV productivity was low due to low source power and low tool availability - both of these aren't a problem with high-NA:

My current portfolio. Historically, the more hate a certain position gets from colleagues, the better it usually performs afterwards. The best ideas tend to be the contrarian ones where you have high conviction. 20% cash will be allocated based on research over the coming months or if there is market sell-off, to increase positions.. $NVDA $AMD $AVGO $MRVL $AWE (UK) $ASML $AMAT $8035 (JP) $BESI (NL) $VLN $WOLF $SNOW $ESTC $ZS $PANW $GOOGL $UBER

The bull case for HBM is fairly obvious..

Former manager at ASML sees three challenges for the company: 1. Customer consolidation 2. Emergence of new semi technologies that are not litho intensive 3. 2nm and 1nm will be hugely expensive nodes, limiting wafer demand

Barclays' bottom-up $ASML EUV model - muted outlook: "Though we model large leading-edge capacity additions, we do not see any increase in EUV exposures at N2 vs N3, and expect a 25% increase at A14 from N2 (only 20% without backside power), but that could be optimistic given efforts to reduce EUV usage. TSMC has reportedly stated that it aims to maintain a similar level of complexity in terms of processing steps from 2nm to A14, without the need for High-NA ( EE News , 29 April 2025), while Tokyo Electron forecasts only a modest gain of EUV layers from N2 to A14. We expect no high NA usage at A14. TSMC has reportedly not found a compelling reason to adopt high NA at A14 ( Reuters , 27 May 2025) and we do not expect Intel to adopt either. Intel clarified at its foundry day that it has the option to use either a low-NA or high-NA solution for 14A, each with comparable design rules and no impact to customers whichever path it chooses. Intel sees one high-NA EUV exposure with a single-digit number of process steps able to replace three low-NA exposures with an additional 40 process steps. Looking at DRAM, we see customers tempering EUV exposure increases from here onwards and lower our EUV layer increases. The transition from 1A to 1B DRAM over the last two years saw a big step up in EUV usage from SK Hynix, though we expect modest exposure increases in the next transitions. We may see Micron stepping up exposures given it has only started using EUV at its upcoming 1 gamma (1C) DRAM node, though we think any material gains could be some time away. We model the transition to 1C capacity to start picking up in 2026, with Samsung set to use 1C for HBM4, while SK Hynix and Micron are sticking with 1B. We assume the next DRAM node, 1D, will continue to use the current architecture (6f2) and so see EUV increase but the following node will see the architecture evolve to 4f2. At this point we think there could only be no increase in EUV so as to minimise the complexity of adopting a new architecture. We expect EUV intensity to increase in following nodes. We see light demand for high-NA tools before a high volume ramp of 1nm begins in 2029. We see DRAM adopting layers at 0B and beyond, representing the majority of demand initially. The risk is that DRAM moves to alternative structures, potentially negating the need for high NA at all, but it is too early to know which path memory players will take. In logic/foundry, we expect TSMC to remain the vast majority of the market and thus their apprehension in adopting high NA likely means we may see limited need for tools before the end of the decade."

William Blair on $ENVX CFO tour: On Friday, August 15, we hosted Enovix CFO Ryan Benton for a series of investor meetings. Our time with Benton was encouraging; his experience strengthens the management team, and we expect clients to have a similar impression. We got an update on CEO Dr. Raj Talluri’s customer tour in China, the warrant transaction, and the key catalyst of the first smartphone purchase order. We reiterate our Outperform rating, and advise adding exposure during any volatility ahead of an inflection in the Enovix story later this year. Dr. Talluri was in China visiting smartphone customers and posted images on LinkedIn at Xiaomi and Vivo, two probable candidates for the lead customer, in our opinion. Six months ago, the team focused all resources on getting the lead customer over the goal line. Benton explained that the purpose of the China tour was to let customers know Enovix is open for business and now has bandwidth to reengage with more customers. The first qualification is the most arduous, but after Enovix debuts its first commercial success, we expect timelines to truncate as rival smartphone OEMs vie for allocation. In our conversations with equity and convert holders, investors have been enthusiastic about the deal structure. Thus far, $50 million of the total $253.8 million has been exercised early. The stock has traded above $10.50 for 18 of the last 19 days, a strong indication it will trigger the early exercise feature of 20 out of 30 days, and the company will be able to close on the transaction well ahead of the October 1, 2026, expiration date. Management believes this growth capital will complete all four lines in Malaysia, providing a path to over $500 million in revenues by our estimates. Benton was very confident that the deep two-year engagement with the lead customer has little risk of last-minute design tweaks and in securing a PO by end of year. We expect an order for roughly 100,000 units to get Enovix batteries in the field in a diverse set of regions, a common rollout technique in smartphones. The smallest of China’s OEMs produce 100 million units a year, so a 20% share will fill two lines in Malaysia. After the field test, we anticipate a large follow-on order and competition to set in among other customers for allocation. Two Common Questions on Competition. First, how does David (Enovix) beat Goliath (ATL)? What is not understood is that even in China, smartphone OEMs desperately want a second source supplier for supply chain resiliency and leverage. On top of that, Enovix’s AI-1 is now the best in the industry at 900 Wh/L, besting ATL’s premier battery at 815 Wh/L by 10%, incredibly valuable to customers. Second, won’t advanced silicon materials render Enovix architecture unnecessary? They do provide higher energy density, but swelling still limits loading below 10%. Enovix’s design enables 100% loading of advanced silicon materials, a true “one plus one equals three” situation. We expect to see this in AI-2. Valuation and Risks. Enovix trades at 19x our 2026 EV/sales estimate. We expect volatility to remain as investors digest the impacts of the warrant distribution; however, underneath, management is executing against its key milestone of launching a smartphone product in the fourth quarter. We see any volatility ahead of the PO and ramp-up toward the end of the year as buying opportunities. Investor risks include technology risk with Enovix’s unique cell architecture, scaling risks and timeline shifts at Fab2 in Malaysia, customer concentration risks, and competition risks from incumbent and other next-generation battery technologies.

Microsoft architect on the four phases of liquid cooling, mentions a reduction of 80% in GPU failure rates with liquid cooling as from phase 2:

Interesting Deutsche expert call on the outlook in the HBM market with former SVP at Samsung who's now running a consulting firm => market bullishness on Samsung share gains and concerns around Micron's competitiveness due its HBM4 base die are overblown: "We recently caught up briefly with Dr Harold Lee, founder and head consultant at Itz Lee M Co, a consulting firm. Prior to this, Dr Lee was SVP at Samsung, where he was in charge of HBM, hybrid bonding, and 3D DRAM for future product development. This update follows two HBM expert calls in Dec 2024 and June 2024. There has been much discussion of late around HBM4 at Nvidia with claims in the Korean press ( here ) that the Samsung comeback was "on" regarding major share recovery at HBM4, potentially rising to "up to 30%" in H2-26. Based on what Dr Lee has picked up from various sources, he notes that SK Hynix delivered its HBM4 samples to Nvidia in March, with completion of qualification likely next month. By contrast, Samsung delivered only a small amount of HBM4 samples in July, with major samples likely delivered in September. This puts Samsung six months behind in terms of HBM4 ramp timing versus the leader, SK Hynix, and three months behind Micron. While Samsung has had execution and yield issues some of which will likely sustain into the HBM4/HBM4E era, Dr Lee notes that what favours Samsung's long-term share recovery is the amount of potential clean room space it can offer up to purchasers. In the near term, Samsung offers capacity growth potential via the Pyongtaek campus (P4 ramp ongoing, P5 in construction, P6) with logic conversions on top, while in the longer term, there is a new site in Yongin (6 megafabs, from 2030). By contrast, SK Hynix has its new campus in Yongin (4 megafabs), operational from May 2027, while Micron has its new fab in Idaho (output to begin in 2027), followed by a new megafab in New York (2029+). Dr Lee believes that the discussion around Nvidia wanting >10GBps pin speed (vs 8-10GBps in the JEDEC standard) to counter AMD competition is over-blown. Samsung and Micron have said they have achieved 11Gbps, while SK Hynix has said it is already above 10GBps. In HBM4, the big change is that the number of I/Os is doubled, which drives bandwidth, while the overall bandwidth per pin is determined by two factors: circuit design in DDR5 and circuit design for the base logic die. On the base die, SK Hynix is working with TSMC (likely at 3nm finFET), while Micron said it will maintain a conventional base die at HBM4, which was perceived to be a concern. Samsung meanwhile will use its own foundry process (4nm logic). However, using a 2xnm process for the base die is 'good enough' at HBM4 in the expert's view, so this is not likely be a major driver of share shifts in his view, even though Micron is likely to move to TSMC 3nm at HBM4E. A more important question remains progression at Samsung on 1cnm yield (one report suggests it is nearly 80% while another points to 50%, but with variability). While Micron has delivered improvements in the maturity of its packaging yield to 70% (from 30% a year ago) using its packaging method (TC-NCF), Dr Lee notes that SK Hynix likely has an equivalent yield of over 90% by contrast. As a result, even if Micron's HBM gross margin has topped 60% using TC-NCF in recent quarters, SK Hynix, in his view, is likely standing much higher using MR-MUF. In addition, SK Hynix's thermal dissipation remains an advantage. Furthermore, going into higher stacks, the gap between MR-MUF and TC-NCF is likely to widen again, while Samsung still likely has lower packaging yields than Micron. With die stacks trending to 16-hi at HBM4E for Rubin Ultra, higher stacks remain then a key challenge for Samsung and Micron, despite TC-NCF yields improving markedly. Dr Lee notes then that ASMPT and K&S then have an opportunity to drive fluxless TCB introduction to improve the TC-NCF process at both players. This would likely mean some new toolsets being required (not upgrades), while cleaning the surface of the micro-bumps is very different. SK Hynix by contrast has more room to go with MR-MUF to scale up to 16-hi and potentially to 20-hi. As a reminder, SK Hynix only uses TCB for temporary bonding, so its usage is much lower. As for hybrid bonding, Dr Lee remains relatively sceptical on hybrid bonding insertion for HBM until 20-hi at the earliest (HBM5). He notes using hybrid bonding is still a delicate process because of several problems that have yet to be solved, while there is a significant increase in cost of production at this point, which could lead to big challenges for the likes of Micron or Samsung if they introduce early at 16-hi (overall HBM packaging cost is ~2x in his view vs TCB) Lastly, we discussed HBM4 share and pricing. In Dr Lee's view, IF Samsung can improve its yield to that of Micron, its HBM4 share should reach 10% by Q4-26, which is lower than the "up to 30%" claim reported above. However, on HBM3E, Dr Lee expects the impact of a 30% discount from Samsung post its recent qualification to drive HBM3 pricing down while this discount should allow Samsung to gain 10-20% share at Nvidia on HBM3E, primarily on low-to-medium- end SKUs. As for HBM4 pricing, Dr Lee notes that HBM4's premium over HBM3E should be significant given a much higher production cost. He hears that SK Hynix is pricing HBM4 at 70% higher than HBM3E, at $23-24 per GB, compared with the current price of HBM3E of below USD14 per GB (original HBM3E price was USD17 per GB). The 40% increase on HBM4 vs HBM3E at start date reflects a 30% higher die penalty (~2x conventional DRAM, going to 2.6x) with yields (higher stacks, higher I/O pads) takes up the remaining delta. We continue to view Micron as a strong beneficiary of growth trends in HBM, even despite the multitude of bearish reports suggesting downside to market share and pricing. On HBM3E, we remain confident that Micron can maintain share in the 20- 25% range through 2026-2027, with Samsung’s incremental share likely to come at the expense of SK Hynix rather than Micron. While Samsung’s qualification on HBM3E may have some drag on HBM3E 12-hi pricing (we model a 5% decline for Micron for CY-26), we do not believe the impact to Micron will be as bad as feared, or as bad as the impact on SK Hynix. On HBM4, we are encouraged by the expert's commentary on initial pricing, with his initial +70% estimate gen-on-gen growth well above our assumptions (we model only +10% growth gen-on-gen, while some buy-siders assume pricing could be down gen/gen pending market share/supply dynamics). For Micron, we believe earning higher market share in HBM4 will remain a challenge especially with many of Samsung’s technical issues likely in the rear-view mirror, but we believe recent fears over pin speeds are overblown, with Micron’s base die likely to be sufficient until HBM4E. In addition, on HBM4E, we also believe that the number of qualified suppliers could decline given the higher degree of customization of the base die, a potential opportunity for Micron. While Samsung should relatively easily convert its existing front-end 1anm capacity over to 1cnm (as well as packaging lines, C1 and C2, in Cheonan) to support its HBM4 ramp, looking further out, we continue to believe that greenfield HBM spending more broadly should be a key driver of strength in WFE spending in 2027-30, which provides a strong support of WFE growth. Our Buy-rated names on the front-end that are levered to this theme include ASML (EUR 870.10 ) , Lam ($145.04 ) and ASMI(EUR 534). As for the back-end, the commentary on the timing of hybrid bonding insertion (primarily 20-hi) is still in line with what we assume for Besi (Buy; EUR 138.50), but the potential for MR-MUF extension at SK Hynix into 20-hi remains a risk, as is slow progress on improving hybrid bonding cost and yield. On the test side, we continue to like Technoprobe (Buy; EUR 8.83) on its potential to enter the HBM4 supply chain with its vertical MEMS probe cards for singulated stack tests, and potentially later on for DRAM wafer-level tests at HBM4E/5 with the introduction of high-speed (ie known good die) tests & hybrid bonding."

3D DRAM coming early 2030s - more deposition, more etch, less litho. Thus, the initial roadmap for $ASML is strong in DRAM with EUV intensity increasing. However, the transition to 3D DRAM will require less advanced litho steps which can be done with older equipment. (chart via Yole)

Former consultant at TSMC details the company's CoWoS expansions and he mentions that as from '26, the bottleneck for datacenter GPUs is likely to shift from CoWoS to HBM. He mentions that DRAM manufacturers are still cautious when it comes to capacity additions. Currently $NVDA is taking up 54% of TSMC's CoWoS capacity, while $GOOGL - $AVGO take 21%, followed by $AMD and $AMZN - $MRVL. (source: Tegus)

The $SNPS + $ANSS combination will be one of the most attractive assets to own in the entire semi value chain (together with CDNS): In foundry, we'll be looking at a battle between TSMC, INTC and Samsung at 2 and 3nm, and while it is likely that TSMC maintains a fairly dominant and large share, it's not great that US generals are saying they will blow up your fabs should China take control of the island.. Similarly, semicap is consolidated but highly cyclical. And a large portion of orders is driven by technological advances i.e. Moore's law + moving vertical. A further slowdown in Moore's law will weigh on annual order flows, as orders per node get spread out over more years.. NVDA will remain dominant in GPUs in the years to come, but long term, it is likely that the hyperscalers will build up custom silicon-based solutions like they've been doing in the CPU market. What NVDA really needs is a fast pace of innovation to continue i.e. in Moore's law, advanced packaging, and AI algorithms. They will remain the GPU player with the best scale and know-how to move swift and take advantage of these changes. This brings us back to EDA (SNPS & CDNS). This software industry is highly consolidated and engineers have been trained on these tools for years, making the business extremely sticky with very limited competition. Additionally, there is a huge shortage of semi engineers and now EDA tools can increasingly automate tasks thanks to AI in the design workflow, giving a clear path to grow into semi R&D budgets. Finally, this industry clearly benefits from the rise in custom silicon, as in-house teams have to adopt the tools as well to design their semis.

$MSFT Azure continues to take share vs $AMZN AWS, and the outlook in enterprise cloud spending continues to improve, from UBS: "The consensus view from our checks was that Microsoft Azure again took share. All 3 partners we spoke to said that their Azure practices accelerated in 3Q25 or were expected to in 4Q25 (in contrast, they generally described their AWS practices as landing slightly below expectations in 3Q25 with forecasts to be steady/flat in 4Q25). We still believe that Microsoft is winning a disproportionate share of remaining on-premise to cloud migrations, workloads such as large SAP and VMware-based migrations. We also believe that Azure growth is getting a unique boost from motivating on-prem Server Product customers to move to Azure, explaining the Azure upside and the Server Product declines. The progression in tone about overall cloud infra spending is notable. Ahead of the 1Q25 cloud infra prints, the tone from customers and partners sounded tough (many tapping on the spending brakes in reaction to the Feb-May tariff/macro noise). Ahead of the 2Q25 prints, there was a clear improvement, with far fewer citing spending softness (indeed, 2Q25 turned out to be one of the best quarters ever in the cloud infra market). This latest round of checks was even better. While several of the F500 enterprises spoke to flattish IT budgets in 2025, none – zero – were planning on incremental spending cuts or delays and in fact some partners said that they began seeing a loosening of budgets in 3Q, an “unfreezing” with elongated sales cycles starting to revert back to normal. We pressed on what spending categories were the biggest beneficiaries of this loosening of budget constraints and “cloud infra” was at or near the top of that list. This should also be a tailwind for rivals AWS, Google Cloud and Oracle OCI. For about a year now, Microsoft has consistently messaged that enterprise spending on its data software stack is strong and getting pulled- along as customers scale on its AI stack. On a recent pod, Azure head Scott Guthrie repeated this message, saying that “a lot of” Azure’s growth is coming not just from direct AI workloads but “also from things like databases that comes along with AI”. Indeed, several customers and partners in this latest round of checks cited an AI-induced pull-through of data software demand (helping the hyperscalers but also supporting a positive view of pure-plays such as Snowflake)."

Contact at Bloomberg says they are looking to move away from $CRWD, considering to consolidate all their endpoint security vendors onto one platform, $S is a candidate here that they'll evaluate. Interview date: Sept 2

Market shares in the datacenter - $INTC vs $NVDA vs $AMD

$ASML is seeing a very strong cycle for memory orders, on a rolling two quarter basis this is easily illustrated, EUR 6.5 billion in orders over the last two quarters alone, highest ever basically in two quarters:

Both $SNOW and Databricks are winning share from traditional datawarehousing, but it's also clear a battle is brewing between the two. Databricks' advantage is that it is really built around data science and machine learning workloads, whereas SNOW is designed for traditional business analytics, although on a novel cloud architecture where storage and compute have unlimited and independent scaling. They're both trying to move into each other's fields now, and SNOW's weak guide might be related to that they're expecting some customers to move data and workloads over to Databricks' data science platform. SNOW's CFO: "We do expect a number of our large customers are going to adopt Iceberg formats and move their data out of Snowflake where we lose that storage revenue and also the associated compute revenue. We do expect though, there'll be more workloads that will move to us. But until we see that incremental revenue on workloads, we're not going to forecast that." My initial impression is that the guide looks very conservative and I expect both platforms to keep winning share, resulting in strong top line growth. But with current enterprise budgets heavily allocated towards AI, Databricks' platform will be better positioned in the short to medium term. SNOW's new CEO has a machine learning background, so his job will really be to build out their data science platform and compete effectively with Databricks.

Goldman on $ASML Scandinavia roadshow => sounds like major revenue upgrades are coming once newly announced AI capacity gets translated into TSMC orders and then ASML orders.. "ASML continues to see the shrink to 1.4nm as beneficial for EUV layer count growth in 2027 and 2028 once the industry moves beyond the Gate-All-Around (GAA) transition in logic. Further, ASML expects the shrink to continue with 4F2 design in memory, and we see introduction timing around 2028, or potentially beyond 2030. As such, we believe that this transition could help EUV layer growth in 2028 (and thus alleviate recent investor concerns associated with the transition and Litho intensity in our view). We note that the industry is currently tracking around 4-5 EUV layers on Low NA on DRAM for 2026 and the company sees potential for 7-10 EUV layers in 2030. Finally, we note there are several factors that could slow down the transition to 3D DRAM and therefore any impact on EUV layer growth. Notably the industry would need to master multiple challenging transitions such as 1) horizontal deposition, 2) vertical Epitaxy, and 3) horizontal and vertical etching; all of which are very hard to execute, implying that the potential transition to 3D DRAM could take many years. Further, we believe that if recently announced AI targets by the ecosystem participants are accurate then there would be a need to build more capacity versus the current infrastructure. Furthermore, we highlight that these AI targets imply production capacity significantly larger versus what ASML’s customers have told them (as part of multi-year forward-looking plans). ASML stated that it has not built or planned this capacity because it has not received orders yet. More broadly, for ASML to build such capacity, it would be necessary for orders to be placed at the leading Foundry(ies) from their customers, thus triggering a need for these semis makers to order equipment. That said, we believe that given the lead time for ASML’s EUV equipment is 12-18M, and the time required to build new factories would be similar, should such revenues materialise they would benefit time periods beyond 2026 (following on from the AI related revenues we already expect next year). We see Samsung’s 3Q EUV orders for both Memory and Logic as positive, and believe that AI demand will not be satisfied by only one Foundry. Additionally, the company reiterated its view that Samsung’s partnership with Tesla on leading-edge Logic ramp at its Taylor fab in the US represents a clear positive, enabling Samsung to accelerate its process learning curve."

$SNOW vs Databricks - Databricks grew faster percentage-wise last year, but SNOW still added more revenues dollar-wise:

Why I expect $AMD to be able to take share over time in AI training vs $NVDA's current market share of >75%👇 PyTorch has become the dominant library to code AI models. This library basically provides Python objects and functions so that you can write all your needed AI code in user-friendly Python. Underneath, Python is running on less user-friendly C which is extremely fast and in communication with the GPU's computing platform. Now, Nvidia's CUDA was the first computing platform which allowed a general purpose programming language, i.e. C, to run on a GPU. As all of Nvidia's GPUs support the CUDA computing platform, it allowed Nvidia to have a dominant position in AI training as code written in Python can easily be run on any of its GPUs. However, $AMD is gradually building out their GPU computing platform as well, named 'ROCm'. And the screenshot below highlights how this platform is already supported by Pytorch on Linux, which is the main operating system in the cloud. Last week, Pytorch's founder also presented at AMD's datacenter event highlighting the investments both sides are making to increase collaboration. Now that AMD has consolidated all of its AI activities under Victor Peng, they should be in a better position to further build out their software ecosystem.

ASML CEO bullish on litho intensity in DRAM roadmap with more EUV layers at next 3-5 nodes. This is v positive long term as AI is highly DRAM intensive with HBM.

$META's engineering team disclosed details on their latest two AI training clusters, both make use of close to 25k $NVDA H100s. It's clear that there is a lot of both hardcore software and network engineering involved here, and that's not simply buying a bunch of H100s and connecting them together. They mentioned they will keep expanding their NVDA fleet: "By the end of 2024, we’re aiming to continue to grow our infrastructure build-out that will include 350,000 NVIDIA H100s as part of a portfolio that will feature compute power equivalent to nearly 600,000 H100s." Highlights: "Today, we’re sharing details on two versions of our 24,576-GPU data center scale cluster at Meta. These clusters support our current and next generation AI models, including Llama 3, the successor to Llama 2, our publicly released LLM, as well as AI research and development across GenAI and other areas. We built one cluster with a remote direct memory access (RDMA) over converged Ethernet (RoCE) network fabric solution based on the Arista 7800 with Wedge400 and Minipack2 OCP rack switches. The other cluster features an NVIDIA Quantum2 InfiniBand fabric. Both of these solutions interconnect 400 Gbps endpoints. With these two, we are able to assess the suitability and scalability of these different types of interconnect for large-scale training, giving us more insights that will help inform how we design and build even larger, scaled-up clusters in the future. Both clusters are built using Grand Teton, our in-house-designed, open GPU hardware platform that we’ve contributed to the Open Compute Project (OCP). Grand Teton builds on the many generations of AI systems that integrate power, control, compute, and fabric interfaces into a single chassis for better overall performance, signal integrity, and thermal performance. Our storage deployment addresses the data and checkpointing needs of the AI clusters via a home-grown Linux Filesystem in Userspace (FUSE) API backed by a version of Meta’s ‘Tectonic’ distributed storage solution optimized for Flash media. This solution enables thousands of GPUs to save and load checkpoints in a synchronized fashion (a challenge for any storage solution) while also providing a flexible and high-throughput exabyte scale storage required for data loading. We have also partnered with Hammerspace to co-develop and land a parallel network file system (NFS) deployment to meet the developer experience requirements for this AI cluster. Among other benefits, Hammerspace enables engineers to perform interactive debugging for jobs using thousands of GPUs as code changes are immediately accessible to all nodes within the environment. When paired together, the combination of our Tectonic distributed storage solution and Hammerspace enable fast iteration velocity without compromising on scale. Our out-of-box performance for large clusters was initially poor and inconsistent, compared to optimized small cluster performance. To address this we made several changes to how our internal job scheduler schedules jobs with network topology awareness – this resulted in latency benefits and minimized the amount of traffic going to upper layers of the network. We also optimized our network routing strategy in combination with NVIDIA Collective Communications Library (NCCL) changes to achieve optimal network utilization. This helped push our large clusters to achieve great and expected performance just as our small clusters. In addition to software changes targeting our internal infrastructure, we worked closely with teams authoring training frameworks and models to adapt to our evolving infrastructure. For example, NVIDIA H100 GPUs open the possibility of leveraging new data types such as 8-bit floating point (FP8) for training. Fully utilizing larger clusters required investments in additional parallelization techniques and new storage solutions provided opportunities to highly optimize checkpointing across thousands of ranks to run in hundreds of milliseconds. We also recognize debuggability as one of the major challenges in large-scale training. Identifying a problematic GPU that is stalling an entire training job becomes very difficult at a large scale. We’re building tools such as desync debug, or a distributed collective flight recorder, to expose the details of distributed training, and help identify issues in a much faster and easier way Finally, we’re continuing to evolve PyTorch, the foundational AI framework powering our AI workloads, to make it ready for tens, or even hundreds, of thousands of GPU training. We have identified multiple bottlenecks for process group initialization, and reduced the startup time from sometimes hours down to minutes."

Director at Micron on why Samsung has been struggling in HBM with $NVDA

Record quarterly bookings for $ASML driven by record memory bookings, logic also bouncing back, shares up 7%. ASML CEO: "The semiconductor industry continues to work through the bottom of the cycle. Although our customers are still not certain about the shape of the semiconductor market recovery this year, there are some positive signs. Industry end-market inventory levels continue to improve and litho tool utilization levels are beginning to show improvement. Our strong order intake in the fourth quarter clearly supports future demand."

An overview of the outlook for leading edge semis - $NVDA $ASML $TSM $CDNS and more! techfund.one/p/outlook-in-le…

$LRCX has developed a tool that can disrupt Tokyo Electron's coating business and also lowers the need for EUV. CEO at Bernstein: "Both in the foundry logic space and the DRAM space, I think many people are aware that the lithography is changing to EUV to be able to print smaller features. And up to this point, that's worked pretty well. But as you continue to try to print smaller and smaller features, you need to change the chemistry involved in the resist itself. And also, we believe you need to change the methodology in which the resist films are deposited. So today, they're conventionally deposit using wet films that are spun on the wafer. And Lam has developed a means by which we can deposit those same types of materials dry, using equipment that looks very much like our etch and deposition equipment. And what we're finding by doing that processing, both the resist deposition as well as the development post exposure dry, is you're finding much better control and the ability to create higher fidelity patterns with less defectivity. The real benefit of the class of materials that we're depositing now also requires less EUV dose to expose the film, which means that the productivity of the EUV tool goes up. We estimate somewhere between 20% to 30% less dose."

Phase 3 cooling systems will change the entire design of the data center => vertical data centers become possible to optimize land use. This also enables data center skyscrapers in metropolitan areas. Microsoft data center architect via Tegus:

Fewer and fewer companies have been able to compete at the leading edge in semis over time - currently there are only two-three left: TSMC, Samsung and Intel (sort of). This consolidated landscape has resulted in tremendous margins for the dominant player i.e. $TSM. Will Japanese Rapidus be able to disrupt the party? And will $INTC be able to make a true Rocky-style comeback? The economics of this industry, i.e. the increasing capital intensity driven by technological advances, suggests that only a very select number of competitors will be able to play at the leading edge.. So I wouldn't count on this segment returning to a true four-player-style market. One or two of the mentioned names will likely be doomed to fail and join the rest of the has-beens. (chart by @sgblank )

TSMC price increases (5% for N3, 10-20% for CoWoS) are great for the entire semi value chain: Nvidia will pass these on to consumers - capacity is fully booked till '26 so the demand outlook is strong - while TSMC gets more gross margin oxygen to embark on N2 and CoWoS investment plans, which will feed through to Semicap. So positive for all of these: $NVDA $TSM $ASML $AMAT $LRCX $KLAC

While both $CRWD and $PANW are the strong players in cloud security, CRWD has a competitive advantage as they are running everything from their single Falcon platform, whereas PANW has been stitching together its solutions in three platforms really. CRWD's CEO on the recent call: “A global financial services giant replaced their Palo Alto Prisma Cloud products in a large 7-figure deal. The Palo Alto Cloud Security products required separate management consoles and separate agents because cloud security is on a separate Palo Alto platform altogether. CrowdStrike was able to deliver an expected 70% time reduction in management as well as more than $5 million in annual staffing cost savings. The patchwork of multi-product, multi-agent and multi-console separate platform technologies resulted in visibility gaps, asynchronous alerts and overall fatigue managing cloud security. Falcon's single platform with its integrated cloud security component was a win for the customer.”