Newsletter · · Ashutosh Agarwal
Optics Speaks, Memory's Price Trap, and Power Goes Off-Grid - AI Accelerators: GPUs, Custom Silicon & Optics - Week of June 29, 2026
AI accelerator newsletter for the week of June 22-29, 2026. Coherent's CEO gives optics its first operator voice and calls the all-optical network 'just physics,' the post-Micron memory debate splits into structural re-rate versus a price-only choke point, and behind-the-meter power starts eating the nuclear narrative while smart money rotates off the hyperscalers.
AI Accelerators: GPUs, Custom Silicon & Optics
Week of June 29, 2026: Optics Speaks, Memory's Price Trap, and Power Goes Off-Grid
Issue 004, Monday, June 29, 2026
A twice-weekly read on the silicon, memory, optics and power feeding the AI buildout, sourced from what operators and investors are actually saying on the tape. Operator and insider commentary is kept separate from pundit opinion. On-air figures we could not independently verify are flagged.
The one-line version: Two issues ago we flagged optics as the theme with no operator on tape. This week we got one, and the Coherent CEO's "that's just physics" framing is the most confident supply-side statement in the complex. Meanwhile the post-Micron memory debate has split into "structural re-rate" versus "it's just price, not volume," and smart money is quietly rotating off the hyperscalers.
Optics: the negative space finally speaks
For three issues this section has been mostly empty boxes. That changed when Coherent's CEO sat for Squawk on the Street, 6/26, the first real operator voice we've had on AI networking.
The substance (operator): Coherent's Sherman, Texas fab makes the indium-phosphide light source for optical links between data-center processors, "one of the main constraints of the entire industry." They are doubling capacity year-over-year by next quarter and doubling again by end of next year, quadrupling total, because lasers have been the bottleneck. Data-center and communications revenue was "up over 40%" last quarter and the CEO expects that growth rate to accelerate; calendar 2026 is "largely booked," 2027 is filling, and "we now have customers booking out into 28," with some supply agreements running "through the end of the decade." On the copper-versus-optical question he was blunt: "as you crank up the data rate on any distance, you're going to eventually hit a limit with copper that forces you to switch to optical. And that's just physics… over time, virtually all of the network will be optical." Notably, he did not view on-package laser integration (CPO) as a threat to the business.
The pundit echo: on The Deep Edge Podcast, 6/25, Ray Mota framed the shift as "moving from the compute wars to the interconnect bottleneck," with the line of the week: "you simply cannot build a 100,000-GPU AI cluster on passive copper alone." His power math: electrical interconnects burn 15–25 pJ/bit, silicon photonics 5–8, in-package optical I/O under 5; 800G is baseline, 1.6T next. And Chip Stock Investor, 6/25 walked Coherent's 6-inch indium-phosphide wafer breakthrough (the industry was stuck at 2 to 4 inch) and the picks-behind-the-picks: Sumitomo Electric (substrates), Oxford Instruments (plasma etch), KLA and Onto (inspection); it cited an ~$2B Nvidia investment anchoring Coherent demand, an unverified on-air claim.
Read-through: the merchant optics supply chain is now sold out the way HBM is, multi-year, take-or-pay, capacity-gated. A better business than the cyclical components the Street remembers.
Memory: structural re-rate, or just a price spike?
Issue 003 covered the Micron blowout itself. This week the debate is the story, and it's a clean operator-vs-pundit split.
Operator data point: on Bloomberg Tech, 6/25, the takeaway was Micron guiding F4Q revenue to ~$51B versus a ~$43B Street number, with management admitting "no line of sight to the supply situation improving." Same segment: Qualcomm's CEO said the company is designing a memory architecture that does not require HBM to cut accelerator cost, the second HBM-skeptic design we've flagged after Cerebras, worth tracking as a real bear vector on HBM pricing power.
The bull structural case (pundit): All-In, 6/26 argued DRAM is now the bottleneck. The new wrinkle is Micron's strategic supply agreements: price floor and ceiling, covering ~50% of revenue with roughly four customers, with floor pricing "ahead of prior-cycle peaks from a gross-margin perspective." Only three firms can make AI-grade HBM DRAM ("as close to magic as science can get"), and the panel's punchline: "DRAM is probably going to be 30 to 40% of all hyperscaler capex next year." China's CXMT (going public) floods consumer-grade DRAM, relief for Apple, not for servers.
The bear case (pundit): The Compound & Friends, 6/26 put it sharply: Micron's market cap (~$1.3T, about to pass Meta) and earnings (it "made more profits than Nvidia made a year ago") are real, but "the bear case is Micron's incredible earnings and revenue growth are predominantly coming from them being a choke point and just raising prices. It's not volume. They're not making more. They can't." With 60%+ ASP growth, the question is durability, and whether customers engineer around the choke point. Micron itself guided DRAM/NAND tight "beyond calendar 2027."
Our read: the floor-pricing contracts are the genuinely new, structurally bullish fact; the volume-versus-price critique is the genuinely valid risk. Both can be true into 2027.
Power: behind-the-meter is eating the nuclear narrative
The freshest operator voice on power was Luke Saladyga on The QTS Experience, 6/25. His point: hyperscalers now prioritize generation source and transient-handling over traditional siting. VoltaGrid, a five-year-old company, has 8.3 GW fully contracted, deploying behind-the-meter generation for Oracle and others to absorb AI load swings the grid can't, and "nuclear will not solve near-term AI load-swing problems." Tell of how the math has shifted: data centers now run "three nines at the point of interconnect," accepting more downtime so capital goes into GPUs instead of redundant power.
The counterweight (operator, slightly older but unmatched): on Interchange Recharged, 6/16, the Large Public Power Council's Tom Falcone said the buildout is real "as long as five or six companies are each spending $100+ billion annually," but warned that actual data-center demand materializing is roughly half of announced requests, with one hyperscaler's real load about half its stated ask. The "bragawatts" gap is the single most important risk hiding inside the power story.
The rotation nobody's leading with
A through-line across the macro pods: capital is rotating out of the hyperscalers and into the bottleneck names. Forward Guidance, 6/19 flagged JPM data showing data-center capex growth decelerating from +80% YoY, with hyperscaler free cash flow falling, "I don't think the hyperscalers will be the play," even as high-yield spreads "barely budged" (the cycle itself is intact). The extreme expression: Limitless, 6/17 reported Leopold Aschenbrenner's fund holding ~$9B of puts across NVIDIA, ASML and Oracle, long memory/power/neoclouds, short picks-and-shovels. Reported second-hand; treat as unverified. Either way, even bulls are moving down the stack toward memory, optics and power.
Custom silicon: Jalapeño, the morning after
Brief follow-through on Issue 003's lead. The Information's TITV, 6/25 added texture: OpenAI's Jalapeño taped out in nine months with eight stacks of HBM plus advanced packaging on leading-edge TSMC, "performance-optimized rather than bottleneck-optimized," designed partly with OpenAI's own models; early reaction was about velocity, not blowout performance. On 20VC, 6/25, the Broadcom-sourced ~50% cost-cut and perf/watt claims got the skeptic treatment (unverified on-air figures), and Cerebras fell 16% on the news, a reminder that "everyone builds custom silicon" is also a threat to the merchant inference upstarts.
Negative space (still signal)
What the tape did not contain this week: any dedicated AMD MI300/MI350/MI400 episode; any standalone Intel Gaudi discussion (Intel's Lip-Bu Tan on No Priors, 6/18 was foundry/strategy, though his "inference-CPU ratio shifting from 1:8 toward 1:1" line is worth noting); any standalone Google TPU / Trainium / Maia / MTIA segment (custom-silicon attention is 100% OpenAI–Broadcom); and no dedicated InfiniBand-vs-Ethernet / NVLink or HBM4 / SK Hynix–Samsung episode. BG2, Stratechery, Acquired and a16z were again silent on accelerators. When the entire merchant-GPU-challenger and TPU complex goes quiet while optics and memory dominate, that is the positioning tell.