Newsletter · · Ashutosh Agarwal

AI Accelerators - GPUs, Custom Silicon & Optics - Issue 003, June 25, 2026: Jalapeño Lands, Micron Roars, Depreciation Cliff Denied

Operator testimony dominates this issue for the week of June 25, 2026: OpenAI and Broadcom name OpenAI's first custom chip Jalapeño, Micron beats even the bull case, and Lambda Labs argues the GPU depreciation cliff everyone modeled is not happening.

AI Accelerators: GPUs, Custom Silicon & Optics

Issue 003, June 25, 2026: Jalapeño Lands, Micron Roars, Depreciation Cliff Denied

Research window: June 18–25, 2026. Operator and insider commentary is separated from pundit opinion, and figures asserted on a podcast are flagged unverified.

The loudest tape since we started. Two operator events dominated: OpenAI and Broadcom put a name and timeline on OpenAI's first custom chip, and Micron printed a number the sell side calls the cleanest read yet on the memory-wall thesis. The quieter, more interesting buy-side debate: whether the GPU depreciation cliff everyone modeled is actually happening. The operators on the tape say no.

Custom silicon: "Jalapeño" gets real

The single biggest story is OpenAI's first custom processor, Jalapeño, co-developed with Broadcom and unveiled June 24. This is operator testimony, not speculation. On Squawk on the Street, OpenAI President Greg Brockman framed it precisely as an inference part that does not displace Nvidia: "This is a chip that is designed for LLM inference from the ground up… designed the whole thing from end to end to tape out. Out in 9 months, which is real, just like blazing fast for this industry." He was emphatic on the boundary: "NVIDIA is our training partner, that's where we do our massive-scale frontier training runs," adding that Jalapeño is "very additive. We feel that we cannot get compute fast enough to keep up with demand."

Broadcom CEO Hock Tan, on the same segment, used the moment to validate the custom-silicon flywheel: "every model maker, LLM frontier model developer, will eventually go towards building, designing, building their own silicon." His demand language was the most aggressive on the tape, "compute demand among our 6 customers… is simply insatiable," and "this is not just '26, not '27. We're seeing that same and even elevated demand in '28 as well" (both unverified speaker claims; note Tan's assertion that Broadcom has exactly six XPU customers). Deployment cadence per Tan: prototype in late 2026, scaling 2027, "full bore" 2028. On Bloomberg Tech, the anchor relayed a company claim of "a 50% lower cost versus a typical AI GPU", flag that as a vendor claim, not independently verified; Brockman himself declined to put a number on it, promising a technical report "in the next couple of months."

The other custom-silicon operator was Cerebras CEO Andrew Feldman, fresh off Q2 results, on Bloomberg Tech. His architectural point is the one worth filing: Cerebras sidesteps the three binding supply constraints. "We don't use HBM… There's a global shortage. It's extremely expensive. Lead times are long… The other constraints… are CoWoS, which is a process inside of TSMC. Again, we don't use it. And the third is capacity at the three nanometer node… We're at the five nanometer node." Reported figures (unverified): Q2 revenue $191M, +92% YoY; a contract "north of $20 billion with OpenAI" plus $1B with G42. The HBM shortage now confirmed from an operator's mouth matters more than the IPO bragging.

GPUs: the depreciation cliff that isn't (says the operator)

No fresh Nvidia roadmap detail this window, but the most valuable GPU insight came from an operator directly contradicting a popular short thesis. On The MAD Podcast, Lambda Labs co-founder/CTO Stephen Balaban said the naysayers calling for a five-year GPU write-off are "completely wrong," and offered hard operational color: Lambda's 2023-vintage H100s "are leasing those out at a higher rate now than we were originally in 2023" (unverified). His read on the real constraint is consistent with the power theme below: the bottleneck is "land powered shell… entitled to have a certain amount of megawatt commitment from a utility," not chip supply. He also confirmed Rubin is entering deployment, "B300s, and VR200s coming soon."

The pundits split on what that re-rating means. On Monetary Matters, Val Zlatev (Analog Century) corroborated the price action, GPU rentals "down 20, 30% year on year" into December 2025, now "up 40, 50%, even more" (unverified), while Jim Chanos reframed neoclouds as "equipment leasing companies… penciling out at 7%, 6%, 5%, 8%… single digit ROICs pre-tax" (unverified), modeling a conservative 10-year GPU life. On RiskReversal, Imran Khan read Nvidia's below-market multiple as the market doubting earnings durability, not a bubble. The operator says the asset lasts; the bears say the returns don't. That is the debate to underwrite.

Memory & capex: Micron clears the bull case

Micron's Q3 (reported June 24) was the quarter the memory bulls were waiting for. Per Closing Bell, adjusted EPS came in at $25.11 vs. $20.78 expected, revenue $41.46B, non-GAAP gross margin 84.9%, with Q4 guidance of ~$31 EPS and ~$50B revenue (all figures as reported on-air, unverified against the release). Evercore's Julian Emanuel said Micron "beat even the bull case for almost all the metrics." Susquehanna's Mehdi Hosseini tied it to the structural call: "the memory wall playing out. Customers have no choice but to pay a premium," flagging that 14 of Micron's 16 strategic customer agreements are "at least approximately $100 billion" over their terms, and that the real 2028 risk is "the transition from training to inferencing." His durability point, ~$16B of free cash flow in the quarter, is the bull's whole argument that "this cycle is very different."

The one wrinkle: on Squawk, CNBC's Mike Santoli and Christina Partsinevelos cited a local report, explicitly unconfirmed by the company, that SK Hynix may be reallocating capacity from HBM toward commodity DRAM. Treat as chatter, not guidance, but watch it; it would complicate the "HBM sold out through 2027" narrative pundits keep repeating.

Power: the build moves off-grid

The capex story is increasingly a power story. On Squawk, David Faber detailed Microsoft's "Kilby" project, a 20-year deal pairing Chevron natural gas with GE Vernova turbines in West Texas for "as much as 2.67 gigawatts," power flowing "by 2028" (unverified), and said he expects "two more of these major announcements" between a hyperscaler and an E&P before Labor Day. The operator color came from Barry Cinnamon on The Energy Show: utility interconnects are "taking… seven to 10 years," training loads "spike… to 25 megawatts," and ~100% of input power becomes heat, which is why developers are bypassing the grid entirely. That is the structural case for behind-the-meter gas.

Optics & networking: still mostly architecture, not names

Optics remains the thinnest of the four buckets in terms of investable specificity. The only dedicated episode was The Deep Edge Podcast, where analyst Ray Mota (a pundit, not an operator) made the framing argument: "we are moving from the compute wars to the interconnect bottleneck." His efficiency hierarchy (unverified): electrical interconnects at 15–25 pJ/bit, silicon photonics 5–8, co-packaged optics under 5, with 800G "the baseline" and 1.6T "imminent." His blunt line, "you simply cannot build a 100,000 GPU AI cluster on passive copper alone," is the thesis. But it was conceptual; no operator, and not a single named optics vendor.

Negative space (this is signal)

  • AMD was absent. Zero substantive MI300/MI350/MI400 commentary surfaced, striking given the GPU noise. The bull case is getting no airtime.
  • Intel Gaudi and hyperscaler in-house silicon (Google TPU, Amazon Trainium, Microsoft Maia, Meta MTIA) went undiscussed as standalone topics; Jalapeño sucked the oxygen out of the custom-chip conversation.
  • InfiniBand vs. Ethernet got no dedicated debate, and named optics vendors (Arista, Astera Labs, Credo, Fabrinet, Lumentum, Coherent) drew no coverage. The optics trade is still being argued in physics, not tickers.
  • HBM4 specifically was never addressed, all memory talk was generic or HBM3E-by-implication.
  • The marquee tech shows (BG2, All-In, Acquired, Sharp Tech) published nothing relevant in the window. The signal came from CNBC and niche operator interviews, not the usual investor podcasts.