Newsletter · · Ashutosh Agarwal

AI Accelerators - GPUs, Custom Silicon & Optics - Week of June 22, 2026: Memory Shortage, Broadcom's Backstop, Optics Awaken

AI accelerator newsletter for the week of June 22, 2026. Operators at Intel, HPE, and a neocloud converge on memory, not the GPU, as the binding constraint into 2028, while Broadcom's $35B Anthropic backstop turns ASIC competition into a credit question and the optics narrative stays early on the tape.

AI Accelerators: GPUs, Custom Silicon & Optics

Week of June 22, 2026: Memory Shortage, Broadcom's Backstop, Optics Awaken

Issue 002, Monday, June 22, 2026 | Window: June 15–22

A twice-weekly read of what operators and investors are actually saying on the podcast tape about AI compute hardware. Operator/insider claims and pundit opinion are flagged separately; speaker-asserted numbers are marked unverified.

The tape in one line

The bottleneck conversation has migrated off the GPU. Operators at Intel, HPE, and a neocloud converged on the same point this week: silicon is no longer the binding constraint, memory is, and it stays tight into 2028. Meanwhile the financing plumbing under the buildout is becoming the story itself, with Broadcom's Anthropic backstop the cleanest example of vendor balance sheets being put to work.

1) GPUs, the secondhand market is the tell

The most useful GPU data point this week came not from a hyperscaler but from a leasing operator. [Operator] Lambda Labs CEO Stephen Balaban told The MAD Podcast (Jun 18) that the "throw the GPUs out in five years" bears "are completely wrong," and, the line that matters, "we're now leasing [H100s] out at a higher rate now than we were originally in 2023." That is a direct operational read on residual values, and it cuts against the depreciation-cliff short thesis.

The spot market agrees. [Operator/market data] Compute Exchange's founder told Odd Lots (Jun 15) that B200 pricing fell after launch and then "came up. And now it's higher than the initial launch prices," with H100 up ~8% and A100 up ~10–15% over three months (drawn from 150k+ daily traded prices; figures unverified).

The bear is loud. [Pundit] Jim Chanos, on Monetary Matters (Jun 20), argued GPU "prices are up 40, 50%, even more as we speak" (unverified) and that neoclouds penciling out at single-digit "5%, 6%, 7%, 8%" pre-tax ROICs are "an equipment leasing company, a finance company, in effect." Note the tension: the same price strength that supports residuals is what Chanos says wrecks neocloud returns. Co-panelist Val Zlatev's structural point is worth filing, Nvidia "doesn't want to get beholden on four customers," so it deliberately "give[s] more supply to the new clouds" to feed competition for the hyperscalers.

On mix, [Pundit] The Information's Phoebe Liu (TITV, Jun 16 put Nvidia's inference share at "around 66% to 74% over the past year," inference now ~60% of workloads (derived analysis, not official; unverified). The training-to-inference pivot is finally visible, with coding the leading workload, per TechSurge (Jun 16).

2) Memory is the new constraint, and three operators said so

This is the week's highest-conviction signal because it came from operators, not opiners.

  • [Operator] HPE CEO Antonio Neri (Squawk on the Street, Jun 17): memory sees only "slight improvement in '27, through yields," and "true wafer capacity is not coming online until sometime '28, [or] the end of the decade." Asked whether the price surge is denting demand: "Not at all, actually an acceleration of demand."
  • [Operator] Intel CEO Lip-Bu Tan (No Priors, Jun 18): "right now memory is a bigger shortage, We are supply constrained," and flagged helium supply as an underappreciated input risk.
  • [Pundit] Val Zlatev (Monetary Matters, Jun 20) quantified the squeeze, memory "gone up four to five X," reasoning models and agents that "suck a ton of memory" as the driver, DRAM/NAND now ~50% of a PC/phone bill of materials versus ~20% (unverified).

Read-through: the marginal AI dollar is increasingly an HBM/DRAM dollar, and the constraint is fab capacity that does not arrive until 2028. Watch Micron and the HBM supply chain over the GPU names if you want exposure to the actual bottleneck.

3) Custom silicon, the financing is the moat now

The week's defining custom-silicon item was a balance-sheet maneuver, not a benchmark. [Pundit/journalist] The Information's Dakin Campbell (TITV, Jun 16) detailed Broadcom's $35B Anthropic deal: Anthropic rents the chips, and Broadcom "provided a backstop, if at the end of the lease [the chips are] worth less than what everybody thought, Broadcom will have to step in." The debt is tranched, A1/A2 carry the backstop and are "rated investment grade," while "the B tranche does not, your credit exposure is actually to Anthropic." Campbell's framing: "Broadcom was at risk of being left behind" versus Nvidia's customer-financing playbook, and this is its entry, against a balance sheet he pegged at "$65 billion in debt, less than $20 billion in cash" (unverified). This is the key debate: ASIC vendors are now underwriting residual-value risk to win sockets.

Sizing the prize, [Pundit] TechSurge (Jun 16) put ASICs at "mid-teens, on a revenue basis" of accelerators today, with a path to "25% or 30% of a much bigger pie," and noted Broadcom's own guide toward "$100 billion in AI revenues next year" (references company guidance; unverified). The honest framing from the same episode: "is the opportunity in front of us still bigger or is it not? Because if it's big, I think they both thrive. If it's not, they're both screwed."

Supply scarcity is reaching custom parts too: [Pundit] AI Breakdown (Jun 19) reported Amazon's Trainium 4 is "completely sold out, more than a year before this chip is actually shipping" (media-sourced; unverified). And the sleeper operator disclosure of the week: [Operator] Lip-Bu Tan told T. Rowe Price's The Angle (Jun 17), "I'm quietly building my GPU, build the different workload for different agentic AI," the first explicit confirmation Intel is building beyond Gaudi.

4) Optics, thin tape, but one signal worth its weight

Negative space is the headline here. Across the window there was no dedicated podcast coverage of the InfiniBand-vs-Ethernet debate, Arista, Credo's active electrical cables, Astera Labs, or Fabrinet/Lumentum strategy. For a theme this newsletter is built on, that silence is itself information, the optics narrative is not yet on the podcast agenda the way GPUs and ASICs are.

What did surface was high-quality. [Operator/inventor] UC Santa Barbara's John Bowers, a silicon-photonics co-inventor, on 632nm (Jun 16): at 100-terabit switch chips, copper "might literally have 20 dB of loss over, a meter," while photonic loss is "the same whether you go a millimeter or 10 kilometers." His structural call, "that top level of spine switches may be all optical, Google is leading this," points at optical circuit switching moving from research into the spine, and he pegged silicon-photonic transceivers as "a thousand times more reliable than VCSEL."

For the equity angle, the freshest name remains Coherent, though the relevant episode (Chip Stock Investor) dates to Jun 9, just outside the window: Nvidia's ~$2B equity stake, datacom/comms revenue +41% y/y, and a co-packaged-optics roadmap targeting 2H 2026, "the moment for these optical companies has finally arrived."

Positioning takeaways

  1. Own the bottleneck, not the headline. Three operators say memory, not GPUs, is the binding 2026–28 constraint. HBM/DRAM exposure has the cleanest scarcity story.
  2. Watch vendor balance sheets. Broadcom's backstop turns ASIC competition into a credit question. Residual-value risk is migrating onto suppliers' books, a quality-of-earnings flag, not just a growth story.
  3. Used-GPU residuals are firming, undercutting the depreciation-cliff short, but neocloud ROICs remain the live bear case. Don't conflate the two.
  4. Optics is early on the tape. The technical case (Bowers) is strong; the investor conversation hasn't caught up. We'll keep flagging when it does.

Next issue: Thursday, June 25. Nothing here is investment advice; speaker claims are flagged unverified where they are assertions rather than confirmed data.