Newsletter · · Ashutosh Agarwal

Custom Silicon Executives Go On Record as Memory Signs Take-or-Pay Deals - AI Accelerators: GPUs, Custom Silicon & Optics - Week of July 2, 2026

AI Accelerators for the week of July 2, 2026. OpenAI's Greg Brockman and Broadcom's Hawk Tan put the Jalapeno inference chip on the record, Micron's sixteen strategic customer agreements get decoded as five-year take-or-pay deals to 2030, and two power operators split over whether AI data centers should run off-grid or on the grid.

AI Accelerators: GPUs, Custom Silicon & Optics

Week of July 2, 2026: Custom Silicon Executives Go On Record as Memory Signs Take-or-Pay Deals

Issue 005, Thursday, July 2, 2026.

Last issue the optics operators did the talking. This week the microphone passed to three louder fights: custom silicon finally put executives on the record, the memory bull case migrated from the earnings print to the contract structure underneath it, and the power debate split cleanly between two operators who disagree about where the electron should come from. Below we keep insiders and pundits in separate columns. On-air figures are the speakers' own claims unless we source them otherwise, treat them as such.

1. Custom Silicon: Jalapeño Gets on the Record

Issue 003 covered the launch and Issue 004 gave it a post-mortem. What's new this week is that the principals spoke. On The AI Daily Brief (Jun 25), OpenAI president Greg Brockman (operator) framed the chip as an inference ASIC, "the first AI accelerator in a multi-generation compute platform," and part of "our long-term full-stack infrastructure strategy to make compute more abundant." He claimed a nine-month design-to-tape-out (which he called possibly the fastest ever for a high-performance ASIC), credited to AI-assisted design, and pointedly reaffirmed OpenAI "cannot get compute fast enough," i.e., no cut to Nvidia orders. Alongside him, Broadcom CEO Hawk Tan (operator) said customer compute demand is "simply insatiable… much more than we can address. And this is not just 26, not 27. We're seeing that same and even elevated demand in 28 as well." That 2028 duration is the real tell, it's what let Micron sign multi-year contracts (see below).

The skeptics matter here. On ThursdAI (Jun 26), Alex Volkov (W&B/CoreWeave) reported engineering samples already running GPT-5.3, a target of ~50% inference cost reduction, and, crucially, that this is inference-only; Nvidia keeps training. His chip-fab contacts noted OpenAI "did not release any specs," and suspect the design actually started two-plus years ago, making "nine months" generous marketing. The honest read: a real second-source for inference tokens, not an escape from Nvidia.

The best framework came from Dylan Patel of SemiAnalysis on Training Data (Jun 30): "everyone will have their own ASIC program… tens of billions of dollars. In the case of Google, hundreds of billions of dollars a year." Google alone runs three distinct TPU architectures (with Broadcom, with MediaTek, plus a third). But labs keep GPUs for workloads that don't fit, "even if the majority of the pie goes to NVIDIA and TPU and Tranium," niches get carved out. His unit economics: Trainium all-in lands under ~$10M/MW-equivalent versus a SpaceX-Google GPU rental near ~$25M/MW/yr, a divergence that explains why custom silicon exists at all. Counterweight from The Information's TITV (Jun 25): analyst Matt Kinda notes Microsoft's and Meta's ASICs still show limited volume ramp after 2 to 3 years, chip quality is the gate, not the press release. And The Six Five (Jun 29) flagged Qualcomm's data-center debut, notably with a memory design that skips HBM.

2. Memory: Stop Watching the Print, Read the Contract

The tape is saturated with Micron; the insight this week is structural. On Chip Stock Investor (Jun 30), the Rossolillos decoded Micron's 16 "strategic customer agreements," five-year take-or-pay deals (2026 to 2030) with hyperscalers and Nvidia on their behalf. Take-or-pay means the customer takes the product or pays a pre-agreed cancellation minimum; roughly $22B of cash deposits are set to land on the balance sheet as a liability. The sharpest detail: price ceilings apply only to existing products, HBM4 and HBM5 are excluded, so the highest-value future memory carries no cap. When executed at current prices, these deals lock ~40% of Micron revenue. Their own bear case, stated plainly: 80 to 90% gross and 70%+ operating margins are "just not sustainable"; paying up now pulls in supply and moderates margins by ~2028 to 2029, and the stock can decouple from financials exactly as Nvidia has (still crushing it on the P&L, lagging memory and optics on the tape).

The debate underneath: is this price or volume? The Morning Market Briefing (Jun 25) put the bear framing bluntly, data-center memory revenue up ~8x, but "price doubling every quarter while volume grows only low single digits." The All-In panel (Jun 26) took the structural side: DRAM is the bottleneck, only three firms make HBM, 2026 HBM is sold out, and DRAM could be 30 to 40% of hyperscaler capex next year. The most useful signal came from Andreas Sloth Christensen on Real Vision (Jun 29): DRAM spot has kept accelerating with no drawdown even as token-expenditure indices roll over, a divergence worth watching, which he attributes partly to demand front-loading. Forward curve, per Tech Brew Ride Home (Jun 29): Jefferies sees memory +40 to 50% in Q3 and +30 to 40% in Q4, with Ming-Chi Kuo pegging 15 to 20% of consumer memory capacity shifting to data centers.

3. Power: Two Operators, One Disagreement

Where Issue 004 gave you VoltaGrid's behind-the-meter build for Oracle, this week framed the actual strategic split. On 20VC (Jun 29), Bloom Energy CEO K.R. Sridhar (operator) made the off-grid case: a 2GW greenfield data center takes 12 to 18 months minimum because copper, cooling, and tradespeople are all constrained, "power should not be a bottleneck if you rely on us." He cited delivering 50+MW in 55 days for an Oracle customer, and argued Bloom's solid-state cells ramp in milliseconds to match GPU pulse loads versus a turbine's "zero to 60 takes a few seconds." Directly opposite, on Let's Talk Energy (Jul 1), Microsoft (operator) called behind-the-meter a supplement, not the primary answer, it prefers grid-connected supply plus large renewable PPAs, and will pay for the transmission and substation upgrades itself. Two credible operators, opposite conclusions; that's the debate to underwrite, not resolve.

The constraint everyone agrees on: it's sold out. Marc Andreessen on The a16z Show (Jun 29, VC), turbines booked four years out, cooling gone, GPUs short, today's models "dumber versions" than physics allows. Lenovo CFO Winston Cheng (operator) on Odd Lots (Jun 27): modular builds in 6 to 9 months, but land and grid power are the gating items and component shortages run 2 to 3 years, pushing him to Southeast Asia and the Middle East for power.

4. The Financing Question Nobody Quite Answers

Bears sharpened the math. On Thoughtful Money (Jun 28), Louis-Vincent Gave (pundit) cited McKinsey's $6.7T of AI capex through 2030, requiring ~$2T/year of revenue to justify, twice the entire global advertising industry, with semis now 18% of the S&P. More concrete: Telltales (Jun 28) flagged Oracle at $56B capex against a $24B free-cash-flow deficit and a planned $40B debt raise, and Next in Tech (Jul 1) warned the neocloud layer is increasingly funded by GPU-collateralized debt with "not a lot of transparency as to which are using GPU-based financing." The bull rebuttal is Dylan Patel's: high leverage on high growth is fine for equity holders, and every rented GPU/TPU/Trainium can be turned around to sell tokens at positive margin, he notes Anthropic was net-income profitable ex-SBC in Q2 with Opus-token margins north of 80%. Both can be true; the fault line is credit quality at the smaller neoclouds, not the hyperscalers.

Negative Space (The Signal in the Silence)

  • Optics stepped back from the mic. After last issue's Coherent operator lead, the only optics content was one technical episode (The Deep Edge Podcast, Jun 25: silicon photonics at 5 to 8 pJ/bit vs copper's 15 to 25). The stocks are, by multiple accounts, "on a tear," but no operator spoke. Watch for the disconnect.
  • AMD stayed off the accelerator tape, again. Two dedicated MI350/MI400/Lisa Su searches returned nothing. Two issues running with no AMD operator or analyst voice is itself a data point.
  • Nvidia roadmap: no new detail. "Best quarter ever for chip stocks," but no fresh Blackwell/Rubin operator color, and NVDA is lagging memory and optics.
  • Google/Amazon/Microsoft/Meta custom chips: discussed only through Dylan Patel's lens; no first-party operator interviews this week.