# AI Accelerators - Week of May 31, 2026 (Issue 001): Nvidia Prints, HBM Sets the Ceiling

> Inaugural issue of the AI Accelerators newsletter, covering May 26–29, 2026. Nvidia's blowout print was the headline, but the pods converged on a different story: high-bandwidth memory, not GPU dies, is now the binding constraint on the cycle.


## AI Accelerators: GPUs, Custom Silicon & Optics

### Issue 001, Weekend of May 31, 2026

---

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

Nvidia did the thing again ($81.6B, 75% gross margins, a $91B guide, an $80B buyback) and the stock fell every day after. That alone is a story. But the more interesting one playing out on the pods this week is that the bottleneck on this whole cycle has quietly moved off Nvidia's loading dock and into Boise and Icheon. If you only had time for one operator interview this week, it was Andrew Feldman's, and it was about memory.

---

## TL;DR

- Nvidia's print was the headline, but HBM is the cap. Multiple shows landed on the same line: it isn't GPU dies that gate the cycle anymore, it's high-bandwidth memory, and that's why Micron and SK Hynix both just walked into the $1T market-cap club.
- The 2027 deceleration debate has begun on the buy-side, even as sell-side notes are uniformly positive. The hyperscaler-ASIC and on-prem-inference threads are the two paths the bears are walking.
- The ASIC/custom-silicon and optics names had a quiet week on the tape: no fresh Broadcom or Marvell color, no MTIA/Maia/Trainium specifics, no CPO roadmap. The signal this week was on memory and on Nvidia's own networking line.

---

## What's New

**1. Nvidia broke out the networking line, and it's bigger than people thought.** On The Circuit (EP 166), Jay Goldberg flagged that Nvidia eliminated standalone networking reporting but, in the process, told you the number: networking was roughly $15B in the quarter, about 20% of data center, and "on track toward a $100B business." Goldberg also noted the new data-center cut: "Hyperscalers" vs. "ACIE" (neoclouds/enterprise/sovereign), with gaming, auto and robotics now grouped as "Edge Compute." That's a tell about where management thinks the next leg of growth is being booked.

**2. HBM, not GPUs, is the binding constraint, and the market has noticed.** On Bloomberg Tech (May 27), Ed Ludlow put it bluntly: "It's not because NVIDIA can't get enough of [GPU dies] necessarily. It's the corresponding high bandwidth memory is not there." Bloomberg's Ryan Vlascelica added that Micron's revenue nearly tripled last quarter, the fastest pace since the 1990s. UBS reportedly tripled its Micron target toward a $1.8T market cap, arguing memory deserves Nvidia-adjacent multiples (~15x fwd EPS vs. the historical ~5x), with Micron trading under 10x fwd EPS at episode time. Both Micron and SK Hynix were up roughly 70% in May and 200%+ YTD.

> "After TSMC, which is right after fab space, memory is number 2... Micron producing numbers where they have 80, 85% gross margins... software gross margins on memory."
> Andrew Feldman, Cerebras CEO, on The Twenty Minute VC, May 26

**3. The only true operator interview of the week was Feldman on 20VC, and he gave the captive-ASIC TCO math in one line.** Same 20VC episode: "When Google or Cerebras puts our equipment in our own data centre, we have a significant advantage over a neocloud because neoclouds are buying hardware with gross margins of 70%, 80% for Nvidia." That is the entire TPU/Trainium/MTIA economic argument in a sentence: captive silicon arbitrages Nvidia's gross margin, and the cap on the strategy is internal demand. Feldman also took a swing at Nvidia's neocloud strategy: "They have funded and backstopped and overallocated to the neo-clouds. They have created a dependence which is probably not healthy."

**4. Enterprises are starting to ask about on-prem GPU racks, a new TAM wrinkle.** From The Circuit: Ben Bajarin's color from Dell Tech World was that, for the first time, enterprises are actively asking about bringing GPU racks on-prem, with one telling him "we didn't know why we'd want to do it a couple years ago, but now that we see how much tokens we're spending in the cloud, we kind of realized it might make sense." Anecdotal, but a directional positive for Dell, HPE and Supermicro, and a small drag on the pure-cloud-inference bull case.

**5. The 2027 deceleration debate is now on the buy-side mic.** On 20VC's NVDA episode, Rory O'Driscoll laid out the bull math (roughly $100B AI capex this year, GPUs about half of capex, Nvidia commanding share gets you to today's ~$320B run rate; Jensen's $3–4T 2030 number gets you toward $1T Nvidia revenue at 70% share of 50%). Jason Lemkin took the other side, citing reports of "Uber COO says the incremental ROI isn't there" and "Microsoft allegedly moving off Anthropic, saying it's too expensive." Treat the Uber/Microsoft lines as speaker claims, not verified, but the framing is the question of the year.

---

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

## The Debate

The bull side got the louder microphone this week. Nvidia just guided to $91B in a quarter, networking is becoming a $100B business inside the company, HBM is sold out for years (Feldman: shortage "at least the next several years"), and a paraphrased Gavin Baker quote on Limitless (May 28) put it about as bluntly as it gets: "NVIDIA could sell $2 to $3 trillion of GPUs this year and next year if only TSMC could supply them." If demand is uncapped and supply is gated, share-loss anxiety is a 2027+ problem, not a 2026 one.

The bear side this week was structural, not specific. Bajarin on The Circuit framed the buy-side thesis explicitly: Nvidia data-center share "is not going to be 80%... it's going to be something lower," analogizing to the Apple/Android moment circa 2009–2011 with hyperscaler XPUs as the Android. The CUDA-plus-networking-plus-CPU vertical-integration argument is the bull counter, and so far it's winning the print, if not the multiple. The marginal-dollar-ROI question (Uber, Microsoft) is the macro overhang to monitor, but the operator voices that would corroborate it were not on the pods this week.

The custom-silicon and on-device-inference legs of the bear case (Broadcom design wins, MTIA/Maia roadmaps, NPU TOPS, Copilot+ traction) were essentially silent on the tape this week. Worth saying plainly rather than fabricating sources.

---

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

## Read-throughs

- **Memory:** the cleanest derivative trade in the week's tape. Micron and SK Hynix already re-rated; Samsung is the laggard. Feldman's 80–85% HBM gross margin call (20VC, May 26) is the number to anchor on.
- **TSMC / CoWoS:** Feldman: "TSMC, which is right after fab space," i.e., advanced packaging is the #1 constraint, memory #2. Reinforces the CoWoS-leveraged names (and complicates the bear case on near-term Nvidia volume).
- **Astera Labs (ALAB):** Limitless (May 28) flagged Atreides' 13F shows ALAB at ~7.4% of fund (#2 position), framed as the connectivity/plumbing layer: interconnect becomes the binding bottleneck once clusters scale past hundreds of thousands of chips. Lumentum and Coherent also held.
- **Marvell (MRVL):** Bloomberg Tech's Ed Ludlow called the MRVL custom-silicon model "exactly the same idea [as Broadcom]" and noted Nvidia's $2B investment in Marvell (March 2026) and MRVL's photonics/optics/networking focus, with the stock roughly doubled YTD. No fresh design-win specifics this week.
- **ByteDance / hyperscaler capex:** Bloomberg flagged ByteDance considering up to $70B 2026 capex, potentially $100B in 2027. Demand signal across the entire GPU/ASIC/HBM/optics chain.
- **China DRAM (CXMT):** The Circuit noted CXMT (~7% market share, +150–160% YoY) filed its IPO with no mention of entity-list status or HBM ambitions, though CXMT DDR5 is already in U.S. gaming cards. A slow-burning policy risk for HBM pricing power, not a 2026 event.

A couple of louder claims came up that I'm not citing as fact: a reported $4B Nvidia investment split between Coherent and Lumentum, a $6B Meta-Corning optical deal, and a $6B Anthropic lease of xAI's Colossus, all from The ACID Capitalist (May 29). Hugh Hendry framed them with conviction, but the figures didn't trace through to primary disclosures in the tape. Flagging in case any of them firm up.

---

## Sources

- [The Circuit, EP 166: NVIDIA Earnings, Dell Tech World and Storage Shock, Tech IPOs (May 26, 2026)](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOj3n6mQNSgFnlLNywBbbAUiAmXIwmvIrWI0rH93IZDUGaAM35kArWec6rxp1SGu4QbTZRVI3a2eVuCaP-2F1Pr492IF2JPh2R9btR2Qk0I6zeKg-3D-3DB22q_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbW6ZLBWGUklZCzb6hzgjw97HnA8M-2FutCMTFYl76BeuZ4o6xfZXbuu8xudl-2BIbEssuNq4cLc-2F6cVIOdkr0Eu-2BEBJVbKQULvzALIGjXr8AYYxyyOr-2FtA2dWpncIRiJ2S7aorwUy5K-2F12HdHiPjeybytYiT-2B1qJ4SAh1VVDSj-2Fnw-2B0AQ-3D-3D), Ben Bajarin, Jay Goldberg
- [The Twenty Minute VC: Andrew Feldman (Cerebras CEO) on Data Centres, Token Costs and Memory (May 26, 2026)](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOjgqgR7syAQqCDqVO-2BIbFzeEIm4HioRqM8lpUD688I5JdJIOEfFOgN2xrxRxjP0iW8ubf8yHXFABfgJb57csBsmrf5qycpc5h-2B9Y8jBb0n4Vw-3D-3DVw5V_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbW6ZLBWGUklZCzb6hzgjw97HnA8M-2FutCMTFYl76BeuZ4oSxJfMHONWCAtdIn74aqNR9gxM978jEJE9oTNIivIIMK0pOwQXgG7v-2BlgN-2FFFgpnxAUxZ6D-2FOqputO-2FlSDAd5pmZQBPc9kMTzw4tZwCFF7jrVV9D5a-2BHrD709HO2Zy-2F-2Bw-3D-3D), Andrew Feldman (operator)
- [Bloomberg Tech: SK Hynix, Micron Join $1 Trillion Market Cap Club (May 27, 2026)](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOjVzYQp2jzDrDO5Kslw72Ge11RpvwYxGlgeHOon-2BmdBRxTQNbO1-2F6YtFpiU92N0qntXpdY9YVXYeGqZhPjuIhePSl2wSF320egVS13lHF7Iqw-3D-3DG6aT_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbW6ZLBWGUklZCzb6hzgjw97HnA8M-2FutCMTFYl76BeuZ4n3LG2oDYfBNTnF7KRsyuBattBaTbEolY1d7-2BRFEwzOhWySd3Yeee7vTRGFRDMkbMwHvIv91amr7Wy-2BcsCHojaQWGeKgsi-2FU0Kb3PZrdlK-2Boxk7dkkNNa41x2oxyjcslBw-3D-3D), Ed Ludlow, Caroline Hyde, Ryan Vlascelica, Ian King
- [The Twenty Minute VC: NVIDIA's $81BN Revenue Quarter (May 28, 2026)](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOiHmw0Toa-2B3Z9Zr9Px-2FMqLMxC-2F6wgxhcehYTO2ExZsmeDoTBsis0Yycb3Gdh4cLrrb30Rr9psHBYCeiMD6Q5mbT-2FgbXxbgNpNCA477JPn0-2FNA-3D-3DjKcW_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbW6ZLBWGUklZCzb6hzgjw97HnA8M-2FutCMTFYl76BeuZ4krtG0a9tXuQM9zSm8kTj5Xa-2B89LCuT-2FPwa-2BfcQxv4XrfBJoQfn2Af1MZ11vVrrfThIL4Duu6mgRBoaHv21AfKkYlCUxprekBUXHSkU375VZEeUzb5yrj43howeVr85Z5g-3D-3D), Harry Stebbings, Rory O'Driscoll, Jason Lemkin
- [Limitless: An AI Podcast: How Gavin Baker Invests in AI (May 28, 2026)](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOhko0GND3oA7-2FgrRh-2Bhey50gl9WY4Fci6nqmeA3emgYXlCeRLO9Gk4nM0wBR10-2FL2AtUdBHt4G6vyg8eKcgBvqNbNPtonm3XbmpDnyK-2BYj6TA-3D-3DuHuT_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbW6ZLBWGUklZCzb6hzgjw97HnA8M-2FutCMTFYl76BeuZ4nBdR5ouFwi18DGpc2a3ZotiXiz-2B-2Fv71wfuBdzp7FMG-2FSBuMmRv6deBy-2B0K0Qm8VzBg4ZKr8dGK0k5htTbORFbibiHzV8pqQU4fp9L7YcFGCXgUfCwURd5f7wFJwc8le3Q-3D-3D), Josh Kale, Ejaz (paraphrasing Gavin Baker, Atreides)
- [The ACID Capitalist Podcast: Compute Is Revenue (May 29, 2026)](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOgu9YWOr-2F50I7SI3jeXIqB2HsXVz2ChwXjgfKD6PUEUyM-2BlFWe2YYHZ1i6TYp1lLhVu-2B2szst9uR7kDy4C1AYp6TKhWMYmWkFz6u-2FdFrLeNOQ-3D-3DBSiL_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbW6ZLBWGUklZCzb6hzgjw97HnA8M-2FutCMTFYl76BeuZ4jSrpbtKhPtIF328OgetkizeqXWYiF8YeQ-2B-2FNY5QVdZ5N0-2FPO7nvtGTW2nWl80NlNoLN6L32wRbo5iQ6Jp6extSsKZ0YoIvbCCNbmn266fnd9FcEKMLYfsdFWUm3s7qoNw-3D-3D), Hugh Hendry, "Trader Mike"

```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
```