# AI Accelerators - GPUs, Custom Silicon & Optics - Week of June 1, 2026: Feldman Says CUDA Lost 70% of Frontier Training

> AI accelerators and custom-silicon newsletter for the week of June 1, 2026 (strict window May 25 to Jun 1, with operator tape from May 18 to 22 flagged inline). Cerebras' Andrew Feldman claims CUDA has lost ~70% of frontier training share, Jensen Huang names a triple supply bottleneck, and Gavin Baker leans into the connectivity picks-and-shovels.


## Custom Silicon vs Nvidia

### Week of May 25 – Jun 1, 2026: Feldman says CUDA lost 70% of frontier training

---

```request-access
variant: ghost
heading: Want this kind of synthesis on your own coverage?
buttonText: Request access
```

**A note on the tape:** the strict 7-day window (May 25 – Jun 1) was thin, three in-window episodes, dominated by post-earnings pundit commentary. The week before (May 18–22) carried the operator weight: Cerebras' Andrew Feldman on *Odd Lots*, Jensen Huang with Michael Dell on Bloomberg Intelligence, Gavin Baker on *All-In*. I've included that material with dates flagged inline, per the 14-day fallback rule.

---

**TL;DR**

- Cerebras CEO Andrew Feldman claims CUDA has lost roughly 70% of frontier training share to TPU and Trainium in twelve months, and discloses a binding AWS term sheet for disaggregated inference (Trainium prefill + Cerebras decode). Treat the share number as a claim from an interested party; the multi-vendor footprint at Anthropic is corroborated three ways.
- Jensen Huang, in his own words after the $81.6B / +85% YoY Q1 print: CoWoS-R, CoWoS-L, CoWoS-S, HBM and 3nm are *all* supply-constrained. The cap on Nvidia's near-term revenue is TSMC and Hynix, not demand.
- Atreides' Gavin Baker has Astera Labs at 7.4% of book, second-largest position behind QQQ puts. The same Baker who is the loudest voice pushing back on the ASIC bull case is leaning hard into the picks-and-shovels.

---

## What's New

**1. Cerebras CEO says CUDA's training moat is half-gone, and an AWS binding term sheet to back it up.** Andrew Feldman, twice in one week, on *Tech Disruptors* and *Odd Lots*: "A year ago, 100% of the large US state-of-the-art frontier models had been trained with a CUDA flow. And they have since lost 70% of the market share." He then disclosed a March 2026 binding term sheet under which AWS will run a disaggregated inference architecture: Trainium for prefill, Cerebras CS3 for decode, served through Bedrock ([Tech Disruptors, 2026-05-28](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOhO2WSBgavWeVS3xQ-2FV4PnAgqxxvt1AOC75O-2BqzhDsRRIjxwDDJaTXyvDQW54ybq1WxQ06FLwByXCovK96Ni5zMupaiwWU6Pond1-2FNKsg2Lxw-3D-3DEvBA_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbWTI6j9lVG131dDZrB8RYRja-2Fw5klvWHO-2FjrggLOMdxGgfqN3OMEawlTRDDh9VlZMVAQSLJXivr9YwxK9mfZG7hWtphyAQXdQONeI1dxXCZCwoBcHsHdFMF7T20E3mM4e8p3tFPRf5hKcQqzeNTrzcYF-2FIyEgWxSfrFZYu-2FWQmVGQ-3D-3D); [Odd Lots, 2026-05-21](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOgp2GSdCl5PXxMimT6MOoMESTBfDMoZ4sc-2FbFurUgaDm4HEepOuV4q0ZVPBF-2BrQqCMQBDiQIFcaEOF02M6-2B-2FNaDeJq97DN1OHlW7VNj4dtSAw-3D-3DmyHA_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbWTI6j9lVG131dDZrB8RYRja-2Fw5klvWHO-2FjrggLOMdxGnln10KnHTTsOXMKSFpbknPe61qmtGD11GnSZ-2Fwftj9B5okspZCgp3rxIjn4fvEX4-2B-2B-2Bw5c6Fmf4ijUcPjifw8tqY7oC5o5OQKURc2-2FVgR-2BRdgcL4n7W0EaorsnzexEqrA-3D-3D)). If true, AWS itself is conceding Trainium can't carry fast-inference decode, a material qualifier for any Trainium share model.

**2. Anthropic is now multi-vendor: Trainium + TPU + Nvidia. Three independent confirmations in one week.** Feldman: "Anthropic's models, trained on Trainium, no CUDA. Served on TPUs, on Trainium, and on GPUs" ([Odd Lots](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOgp2GSdCl5PXxMimT6MOoMESTBfDMoZ4sc-2FbFurUgaDm4HEepOuV4q0ZVPBF-2BrQqCMQBDiQIFcaEOF02M6-2B-2FNaDeJq97DN1OHlW7VNj4dtSAw-3D-3DLr0t_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbWTI6j9lVG131dDZrB8RYRja-2Fw5klvWHO-2FjrggLOMdxGsuuWcw68QBhiJ45gDq6lgsaHIQ4x9TdvOw54PSZ5W0Mt86sTQ4X72Kz33lkXfqu8KcK51qZG0DoZJFRuTAi-2BYpmw9p9uk6yOe4nwt5S999xda78PAbiERn0CE78im7SoQ-3D-3D)). Jensen on *Squawk on the Street*: "This year, we had the benefit of also winning Anthropic" ([Squawk, 2026-05-21](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOhpnaxL6X4mTNCrT13Dce8PmgH7soNGnLqFQwIggKRmiOQx2-2FIdEddtOfRdbvjtcGJDe8i19TpYTrZPZS-2BQovG64jAcVkfuSUXi-2Bpc7b-2FLcmA-3D-3Dy18-_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbWTI6j9lVG131dDZrB8RYRja-2Fw5klvWHO-2FjrggLOMdxGjPKqXCQOgJKR0IEdCKDHzHZwIGeZh3DIqZorxlmzQEd-2FbR8KsWw65QZXfLrKWkPM2Kx29x-2BBb5M8i2RW7LZk6u8ho79wK-2FMueQWd56BsY7rYbXHnwPnGy8uxClvYgTg0A-3D-3D)). And David Faber, citing the fresh SpaceX S-1: Anthropic is paying $1.25B a month, $15B a year, to xAI to rent Memphis Blackwell capacity. The "Anthropic = Trainium exclusive" narrative is dead. Project Rainier GW figures were not refreshed this week.

**3. Jensen names the triple bottleneck: CoWoS, HBM and 3nm.** On Bloomberg Intelligence with Michael Dell ([2026-05-18](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOjKDyj786UQyEVyVyiYgeJxDfXpLxZHV6HsvGiQl9Uutn0dlzpV3MUwTHkXjnibuSe5Xo4Q8AlUuq46Pl30nwwOMICh-2Ff5lTWMzWypJ7fPhkg-3D-3D8sn6_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbWTI6j9lVG131dDZrB8RYRja-2Fw5klvWHO-2FjrggLOMdxGsdw6GgT79sg27tTPiMsdAcCChljAmul9hmL7dHhGGinBMhw8H0zvNajdft2ugP5dDFojB6q4EVdB6KLcw2IGlmDKlrBTnwyc7MQRw-2B7oMuObpD4lojazDISxcLUPuB8EQ-3D-3D)): "The CoWoS is lined up with the HBM, which is lined up with the Grace Blackwells… CoWoS-R, the CoWoS-L, the CoWoS-S. All of it is all lined up. The silicon photonics is lined up. Everything is all lined up. It's just that the demand is much greater than the overall capacity of the world." Feldman, independently, named the same triple, HBM, CoWoS, TSMC 3nm, as the binding constraint on *every* vendor including the ASIC houses. Two operators from opposite ends of the value chain agreeing on the same three chokepoints is about as good a confirmation as you get outside of an earnings call.

**4. Atreides has ALAB at 7.4%, second position behind QQQ puts.** Josh Kale unpacking Gavin Baker's latest 13F on *Limitless* ([2026-05-28](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOhko0GND3oA7-2FgrRh-2Bhey50gl9WY4Fci6nqmeA3emgYXlCeRLO9Gk4nM0wBR10-2FL2AtUdBHt4G6vyg8eKcgBvqNbNPtonm3XbmpDnyK-2BYj6TA-3D-3DTrNP_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbWTI6j9lVG131dDZrB8RYRja-2Fw5klvWHO-2FjrggLOMdxGlRefPYWx584jy3CkLFqvu0AkhzEuADscTwnBDKe9-2F88il1Ra71EKb6ZyDHaGepl3E7t3-2BOXzhLOlk9wisBEm0Q4R3HIIopblzbLsXiha4J7BIrmHh8I8JFwDC8bH0QbQw-3D-3D)): "As AI clusters scale to hundreds of thousands of chips, the bottleneck stops becoming GPUs specifically. But it starts becoming that transfer window… the plumbing system that Astera Labs builds." Worth noting the same Baker, on *All-In*, is the most vocal pundit pushing back on the broader ASIC bull case, he wants the connectivity attach regardless of who wins the compute war.

**5. Synopsys CEO discloses a license-plus-royalty pivot on hyperscaler IP.** Sasim Ghazi on *Squawk on the Street* ([2026-05-28](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOhor9PLCS4eBsYELS0Q5vqYplMF-2B-2FQYqKCgfelRC7JPiQS4zi57-2FPCz6CxkMRqV6tHEt-2BcqxfFoUBRJYUF3vxexsvJUJyQlxqiHGQQC-2Ft-2BYtA-3D-3DHni7_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbWTI6j9lVG131dDZrB8RYRja-2Fw5klvWHO-2FjrggLOMdxGv33LmrOnINgslgQKQQ8DpkFvJAXSeQh-2F8dAtZpFqW6RqigErkz1hZFFnN59-2FFGI0T1jKWal1-2B8T3MDsH-2BOOa9Wg16Q-2BFiXmAvQdFnC6UdwfOSDa27Y6PXA-2BW-2FQWICirZA-3D-3D)): "Today they pay us as a license and that's it… We don't get paid on volume and success. So it's going to be a license plus a royalty, which will increase the opportunity of monetization." If hyperscaler XPU programs are sticky enough that Synopsys can finally charge royalties, that says something about the durability of the in-house silicon stack, and adds incremental cost into the ASIC TCO bridge that bulls don't usually model.

---

```request-access
variant: inline
heading: Want this kind of synthesis on your own coverage?
buttonText: Request access
```

## The Debate

**Steel-man for ASICs (Feldman, Boone, Goldberg).** TPU and Trainium are demonstrably shipping at frontier labs. CUDA's training moat has visibly cracked inside twelve months. Citizens' Andrew Boone, on [The Exchange (2026-05-19)](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOiqLVfRAzDfF8GKrf4adZmue9XTwS7sw7quZV8ZnK1BetxryWJcPIcfLC8A-2BQdTyL4MX-2BowH3y3gj3RYiJIotWikeVWjVgP8GR-2By7NPDyFptw-3D-3Dr8ib_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbWTI6j9lVG131dDZrB8RYRja-2Fw5klvWHO-2FjrggLOMdxGj67YRvNlNnh3vSQiRY7BXxHQqNOFYQgY1SDoAYpHk6LGCuEuDEc-2FBtrmU1USl2bpiFlVha-2FNQLQp0ID2-2BJoknVmFiAndwjL1nl0qzJ2vSkV-2FgK1keZ61ugjzc7h95HtPQ-3D-3D), is modelling TPU revenue from "$3 billion… this year, up to $25 billion by 2027," validated by the Google-Blackstone $5B JV. Seaport's Jay Goldberg, the only sell rating on the Street, on the same show: "Anytime NVIDIA doesn't get a deal, it's gone to Google's TPUs."

**Steel-man for the merchant.** Gavin Baker on [All-In (2026-05-22)](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOjhJiHfIkL7X9PFZLdafPBichcET5rnOjqc7CV94ZBEfJ-2F4EOjRtwXtuMpGQfC9UDqmNP0xbRjrvVGQbDXhVe81QPct3HYGwoMl8loIOBiCdw-3D-3Du53u_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbWTI6j9lVG131dDZrB8RYRja-2Fw5klvWHO-2FjrggLOMdxGmAedDeZgHntIH9GdyqeKbmU8opVkWVFA70P5H19QXfBCznPPDFkZ740z873Wox84N5-2BDcEBDJS-2F9O35E5eOvfhnFocljwbYOtA65FwgSpluUMl5Ths4YvVmuW-2Fm72Ut6Q-3D-3D): Broadcom guided +143% YoY for AI semis, but Nvidia's Western (ex-China) AI revenue is *still* growing faster on an apples basis. None of the hyperscaler ASICs have submitted to MLPerf, until there is a clean GB300 vs TPU v7 vs Trainium 3 benchmark, the perf-per-dollar claims are unverified. Bloomberg Intelligence's Kunjan Sabani on [Bloomberg Tech (2026-05-21)](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOiKlSHhGJ421X4Dy1JI-2FiKsrwJe7lQOfetalY-2BbA6Mbb-2Ft7EMrbmQaIzUOdPvDzdK6NGpeSlkgQma54a0qmN-2Fdx0kRr1K8SJUnHVqgbCtPyAg-3D-3D44O__7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbWTI6j9lVG131dDZrB8RYRja-2Fw5klvWHO-2FjrggLOMdxGtPPP0yfxWvVZ6vanFQRFO0StfQ1BA1pEqW4d0hEtvknFdIhjuszfQqoY4Ia78CG2y0ZjaocR3KcojkKpNvLz4QxbQoDPNHjsK6GJNWOQ-2FeHi2oPhJaI9REjXkXYVZcNWg-3D-3D): the ~50% of NVDA revenue not flowing to the four hyperscalers, sovereigns, neoclouds, enterprises, AI labs, faces zero ASIC competition. That's the monopolistic margin pool.

**Where I come out this week:** the ASIC *footprint* at frontier labs is now real and broad, and that's largely already in the AVGO and GOOGL multiples. What is *not* yet in Nvidia's number is the upside from Jensen explicitly saying he's supply-constrained across three independent variables, that's a quarter or two of revisions if TSMC and Hynix capacity lands as planned. Long NVDA, long ALAB, watch the ASIC names off any benchmark catalyst.

---

## Stocks in Play

- **NVDA**: *Bull:* $81.6B / +85% YoY, 75% GM, $91B Q2 guide ±2%, $80B buyback, dividend +25x (per Baker on All-In). Jensen says supply is the cap. *Bear:* Culper Research alleging >20% of FY26 revenue routed to China via Megaspeed, with firsthand Taiwan/Singapore sourcing on rack-level diversion (per [Better Offline, 2026-05-20](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOi-2BShdqP05MsmT1iijCJjiqmRs-2BKrD5atzUVqmIMNEsHlCBVJlj2saGZv1j75pPLt67C0IW4N-2FrNz5I64zZ3s0ZjGpSefOAfH0kG9FimFobHQ-3D-3DHSxv_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbWTI6j9lVG131dDZrB8RYRja-2Fw5klvWHO-2FjrggLOMdxGhO4wkJajVh7VWGo-2FQeKhwX6Ja2Ggnsiu8-2FY8T3EZ6L7GNL3CQ3lvCrF4aqBhNZNXqukjO8c-2BH9a-2FsESjjXfi6zScuKzBxz0GQ80HZx4vrmUbysMtfMwQsylPIj2EkGlPQ-3D-3D)). Vera Rubin Ultra "Kyber" 1MW racks may strand existing data center designs. *Watch:* MLPerf GB300 submissions; Megaspeed regulatory follow-through.
- **AVGO**: *Bull:* +143% AI semi guide for the coming quarter. *Bear:* still slower than NVDA Western. *Watch:* XPU customer-list disclosures, OpenAI Stargate confirmation.
- **GOOGL**: *Bull:* Boone's $3B to $25B TPU revenue curve; Anthropic serving on TPU confirmed. *Bear:* no external customer beyond Anthropic disclosed this week. *Watch:* TPU v7 / Ironwood deployment metrics; Apple and SSI rumours.
- **AMZN**: *Bull:* Trainium v3 is in market; Anthropic still a Trainium training customer. *Bear:* the Cerebras prefill/decode split implies Trainium decode weakness AWS itself acknowledges. *Watch:* Project Rainier GW disclosure; re:Invent 2026 messaging.
- **MSFT**: *Bull:* The Information reportedly has Anthropic in talks for Maia (per Cramer). *Bear:* zero volume or yield colour this week. *Watch:* Maia 2 silicon timing.
- **META**: *Bull:* still in the custom-silicon design cycle via Synopsys IP. *Bear:* zero MTIA v2 volume disclosure. *Watch:* MTIA v2 inference deployment scale.
- **ARM**: *Bull:* Vera CPU tag-along narrative. *Bear:* no new Neoverse-in-XPU or v9 royalty mix data this week. *Watch:* XPU royalty disclosures next print.
- **ALAB**: *Bull:* Atreides 7.4% conviction position; structural connectivity attach as clusters scale. *Bear:* no new design-win color. *Watch:* TPU/Trainium/Maia attach rates; Scorpio fabric ramp.

---

```request-access
variant: inline
heading: Want this kind of synthesis on your own coverage?
buttonText: Request access
```

## Read-throughs

- **TSMC**: Jensen and Feldman both name 3nm and CoWoS as the binding constraint. Capex cycle still tight.
- **SK Hynix / Micron / Samsung**: per Kale on [Limitless](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOhko0GND3oA7-2FgrRh-2Bhey50gl9WY4Fci6nqmeA3emgYXlCeRLO9Gk4nM0wBR10-2FL2AtUdBHt4G6vyg8eKcgBvqNbNPtonm3XbmpDnyK-2BYj6TA-3D-3D59Yf_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbWTI6j9lVG131dDZrB8RYRja-2Fw5klvWHO-2FjrggLOMdxGvWwF-2BbuwYfCSGh4E-2B8ZAMSJ0rxiaFrctCyQchGqUQ7t8OgIW0HdTuv7ahDGvVXiFlC-2FwJXZkgtV4MJutGiB9lvtu9V3on7HJ1nqKYPKV5e0qBz2uen2R9hdahVdPn7flw-3D-3D), Hynix allegedly fielding $50–$100B HBM forward-supply offers from Google and Microsoft, ~70% operating margins (a figure the podcast left unsourced).
- **SNPS**: license-plus-royalty pivot is a structural margin lift if it sticks.
- **Alchip / GUC**: radio silence this week. Material information gap for any ASIC-services thesis.
- **Cerebras**: strong operator color, but with a recent IPO and a Trainium piggyback deal. Discount the share claims accordingly.

---

## What Changed vs Last Week

This is the inaugural issue, so there is no prior to diff against. From next issue we'll track: TPU revenue ramp vs Boone's $3B to $25B curve, Anthropic vendor-mix shifts, any MLPerf submissions from the ASIC camp, and CoWoS / HBM capacity prints out of TSMC and Hynix.

```request-access
variant: banner
heading: Run your own corpus.
description: matterfact is deployed with select institutional partners. Request access to run it on your own coverage.
buttonText: Request access
```