SemiAnalysis

GB200 Hardware Architecture - Component Supply Chain & BOM Hyperscale customization, NVLink Backplane, NVL36, NVL72, NVL576, PCIe Retimers, Switches, Optics, DSP, PCB, InfiniBand/Ethernet, Substrate, CCL, CDU, Sidecar, PDU, VRM, Busbar, Railkit, BMC https://lnkd.in/dF6GCBgK

Chevron deference has just been struck down by the US Supreme Court. What is it and why does it impact semiconductor companies ? Under that 40-yr-old legal doctrine, US federal agencies had the power to create their own rules & regulations when a law is ambiguous. In our industry, this is particularly relevant for technology export controls – agencies were in the driver’s seat and didn’t have to worry about being challenged in Court. This is now over, and the power has been handed back to the Court system after the Supreme Court’s ruling in Loper Bridge Enterprises v Raimondo. The US Commerce Department’s Commerce Control List is a set of items subject to export controls and includes a plethora of Semiconductor products and manufacturing equipment. Currently over 500 pages long, items on the list are identified in exacting detail by Commerce Department officials with subject matter expertise without direct input from Congress, and free from any serious worries of court challenges. Updating the list to close loopholes and adapt to rapidly evolving technology is a perpetual game of whack-a-mole. Now that Chevron has been struck down, it’s open season for Semiconductor product and equipment manufacturers to mount challenges to export regulations in the court system – as long as they can point to some ambiguity in the laws authorizing these export controls.  With tens of billions of sales in semiconductor products and manufacturing equipment to restricted destinations and parties at stake – expect companies to at least look into challenging aspects of the current rules – opening up the possibility of some rules being set aside or future regulations being snarled in injunctions and other legal challenges. In the most chaotic scenario – attempts at implementing controls could be made ineffective as new rules become snarled in time consuming legal battles, unable to keep pace with rapid technological developments in the Semiconductor industry. However, Congress, if it possesses sufficient political willpower, can still act to modify relevant statutes related to export controls and make such rules explicit and defined by statute, thus resolving any ambiguity and reducing the surface area for legal challenges. Indeed, export controls and the ongoing industrial competition with China seems to be one of the only areas where Congress is likely to find sufficient consensus.

100,000 H100 Clusters: Power, Network Topology, Ethernet vs InfiniBand, Reliability, Failures, Checkpointing Frontier Model Scaling Challenges and Requirements, Fault Recovery through Memory Reconstruction, Rack Layouts https://lnkd.in/dvBtcvvB

Bigger isn’t always better? Rapidus, Japan’s gamble to start a brand new 2nm domestic foundry in 2027 has had many puzzled, but recently they revealed how. They are touting the advantages of small batch processing as helping yield, cycle time, and more. Couter-intuitively, they claim small or even single-wafer batches are faster than large ones – up to 60% by their numbers! Traditional flows use batch processing to increase speed and reduce cost, typically in chemical processes such as oxidation furnaces, CVD chambers, etc. Entire 25-wafer lots or more can be run at once, and a single high-capacity tool is cheaper than many smaller ones. Rapidus claims that large isn’t always faster: the process is only as fast as the slowest wafer in a batch. And large chambers, for example in deposition tools, can take a while to heat or cool between batches. These benefits are even stronger in R&D, where rapid iteration is needed. For a single experiment they show a 10x or more reduction in time. WFE suppliers will love this since it means more tools needed. The downsides for chipmakers are higher cost and that metrology is much more difficult – wafers processed individually will all have unique signatures… we’ll see if Rapidus is right within a few years as they aim to launch their in-house 2nm process by 2027. The companies that stand the most to benefit from this are ASM International and Lam Research as they offer more comprehensive suits of single wafer ALD and single wafer etch product. Interestingly, Kokusai Electric, one of Japan’s largest equipment companies stands to lose out, as the leader in batch deposition. Their stock has more than doubled since we wrote about them less than half a year ago. See the technical and financial breakdown and their primary niches here : https://lnkd.in/gx8zZuD6

OpenAI Chip Team Is Now Serious Poaching Aggressively https://lnkd.in/dipzbZP8

How Dell Is Beating Supermicro Enterprises, Sovereign AI, CoreWeave, Tesla AI Neocloud economics and orders https://lnkd.in/d2qNUpMt

Apple’s AI Strategy Apple Datacenters, On-device, Cloud, And More When to run on-device, in Apple datacenters, or in the cloud with OpenAI Personal Assistant Deal economics https://lnkd.in/dGN2b5mn

EDA company Silvaco, $SILV, is pricing their IPO this Wednesday. Here’s our quick analysis.   The EDA market is dominated by Cadence, Synopsys and Siemens Mentor with ~75% of the market. The remainder of the market is taken up smaller vendors such as SILV with revenue of only $54mn. This is less than 1% of Synopsys’s scale!   SILV is focused on TCAD for the Power, Display, and Memory markets. TCAD tools simulate device behaviour to validate performance before manufacturing. The main growth driver of EDA spend is the rising complexity of chip designs and this is especially seen in leading edge logic. However, in Silvaco’s key markets of power and display, it is less clear that R&D intensity is rising as much. Cadence and Synopsys are behemoths, but they still outgrew Silvaco! Cadence and Synopsys have more torque to the structural tailwinds of rising complexity than Silvaco does.   SILV is using their simulation capability to expand into a new area: creating fab “digital twins”. This help wafer fab customers simulate and improve production processes to drive down costs and increase yields. Micron is the initial customer and they’ve also invested $5mn into SILV in April. This is a new opportunity for SILV but this space has already been attacked by others. TSMC has their own internal solution. Samsung is using startup minds.ai. The challenging part of this is the key input is the fab’s own data. This is highly sensitive so there needs to be a lot of trust involved and to build the software SILV will need to have a custom solution for each customer. argins won’t be as attractive compared to selling off the shelf software modules.   SILV is sub-scale and it is not clear they should be a standalone company. SILV was loss-making in 2021 and 2022, and barely become EBIT profitable in 2023 with a 2% margin. Gross profit is spent on R&D and SG&A. More dollars are directed to SG&A than R&D where investment leads to more direct business outcomes. Since 2021, SILV’s R&D spend is actually down (!) with S&M and G&A expenses up. This is due to IPO costs and R&D was cut to finance this. R&D investment is key for an EDA business to stay competitive and grow. Both CDNS and SNPS continue to grow their R&D investment while also spending more than SILV does as a % of sales. This shows just how hard it is for the small EDA players to try to close the gap with the giants. How does SILV’s valuation compare to SNPS and CDNS? SNPS and CDNS both have better growth outlooks, stronger industry positions, and are more profitable. However, SILV is launching at a more expensive valuation. SILV’s margins are still low relative to management’s target of 25% non-GAAP operating margin so it may be too harsh to judge SILV on earnings multiples presently.  SILV is cheaper on sales multiples, so the upside investors are playing for is if they can get close to SNPS and CDNS levels of profitability. Of course, that’s a big if and CDNS and SNPS are also far superior in terms of quality.

OpenAI Is Doomed? - Et tu, Microsoft? Meta, Google, Anthropic, DeepSeek, Inflection Phi Wizard, Distribution/Integration vs Capital/Compute? https://lnkd.in/df25b9By

Rest in peace to Robert Dennard, one the GOATs of the modern era. Very few know and respect how much his work propelled our civilization. While the memory he invented requires constant refreshs, our memory of him will be persistent. He invented modern DRAM in 1967 at IBM, the backbone of most computing systems, and still the most prevalent memory technology. Of course there have been major advanced in manufacturing and refresh, the basic structure remains the same. Another famous contribution was Dennard Scaling. In 1974, he wrote "Design of ion-implanted MOSFETs with very small physical dimensions" Basically it modeled transistor performance as transistors shrink. While I knew what it meant, I admittedly hadn't read the paper till just now. - https://lnkd.in/djuy3hdZ It laid the groundwork for "Moore's Law" which was predicted just a year later. Dennard scaling did not break down until the mid 2000s https://lnkd.in/dgJf4V7z