Newsletter · · Ashutosh Agarwal

OpenAI Unveils Jalapeno Its Own Broadcom Inference Chip - Custom Silicon vs Nvidia Weekly - Week of June 29, 2026

Custom Silicon vs Nvidia weekly for the week of June 22-29, 2026. OpenAI and Broadcom unveiled 'Jalapeno,' OpenAI's first custom inference chip taped out in nine months, while Micron's blowout showed memory now eats 40 to 50% of hyperscaler capex and makes designing your own logic die worth the trouble.

Custom Silicon vs Nvidia Weekly

Week of June 29, 2026: OpenAI Unveils Jalapeno Its Own Broadcom Inference Chip

Forty-plus episodes touched our beat in the strict 7-day window (June 22–29), and unlike last week, the operators actually showed up. Hock Tan, Greg Brockman, and Andrew Feldman all went on the record. The catch: it was a two-story week. Story one is OpenAI's custom chip. Story two is Micron, which swallowed the tape and is really the same story wearing a different hat.

TL;DR

  • OpenAI unveiled "Jalapeño," its first custom inference chip, co-designed with Broadcom, taped out in 9 months, claimed ~50% cheaper than a typical GPU for inference, gigawatt-scale deployment starting late '26. The seventh hyperscaler-class buyer to go custom.
  • Hock Tan said the quiet part out loud: "every model maker… will eventually go towards building, designing, building their own silicon." And he is not seeing Nvidia demand fall, his six customers are "insatiable" through 2028.
  • Micron's blowout is the read-through that matters: memory is now 40 to 50% of hyperscaler capex, and the squeeze is what makes designing your own chip worth the trouble.

What's new

Ranked by what moves a book, operator design-win color first, pundit framing last.

1. Jalapeño is real, and an OpenAI President plus a Broadcom CEO put their names on it. In a CNBC exclusive on Squawk on the Street (2026-06-24), OpenAI's Greg Brockman said the chip "is a real performance improvement on performance per watt… performance per dollar," designed "from end to end to tape out… in 9 months, which is real, just like blazing fast for this industry." Broadcom CEO Hock Tan laid out the schedule: "beginning it late '26… small prototype development in '26, scaling up over time in '27 and going full bore definitely in '28." A detailed technical report is promised "over the next couple of months," so the perf-per-watt number is still a claim, not a benchmark.

2. Tan's thesis, stated plainly, and the part the bears miss. Same interview: "every model maker, LLM frontier model developer, will eventually go towards building… their own silicon… simply because they can do it much better." But crucially, this is additive, not substitutive. Brockman: "It's very additive. We feel that we cannot get compute fast enough to keep up with demand… NVIDIA is our training partner." Tan said demand across his "only 6" customers is "simply insatiable… not just '26, not '27… elevated demand in '28 as well," and that AI revenue was "2x… more than 2x what '25 was," with "'27 will be 2x at least what '26 is." Custom silicon for inference; Nvidia still owns the training run.

3. The 9-month tape-out is the actual signal, not the chip. On The Information's TITV (2026-06-25), NEA's Mustafa Nimuchwala called it "a classic example of… buy the rumor, sell the news," but flagged the real takeaway: "the velocity side." Jalapeño "has eight stacks of HBM… leading edge TSMC capacity and they're doing co-loss and advanced packaging. So this is not a bottleneck optimized chip. This is a performance optimized chip." Translation: OpenAI is fishing in the same constrained HBM/CoWoS/leading-edge pool as Nvidia, it isn't routing around the bottleneck, it's competing for it.

4. Cerebras gave the operator counter-read. Cerebras CEO Andrew Feldman, quoted on Bloomberg Tech (2026-06-24): "This market will not be consolidated around GPUs. There will be ASICs… from hyperscalers… from labs, and then… companies like Cerebras." Record revenue of $191M, up 92% YoY; and a genuinely differentiated supply story, "we don't use HBM… we don't use [CoWoS]… we're at the five nanometer node," i.e. immune to the very constraints choking everyone else. The market still punished it: Cerebras fell ~16% on the Jalapeño headline (20VC), reading OpenAI's in-house move as a threat to its ~$20B OpenAI order.

5. Micron is the tell. The blowout, revenue ~$41.5B, gross margin ~85%, next-quarter guide ~$50B, ~$100B of locked long-term contracts, dominated dozens of shows. The number that matters for us: hyperscaler capex is now "40 to 50 percent… going to memory" (Mustafa, TITV). DRAM contract prices rose "90 to 95% in Q1 alone," and Tim Cook called it a "100-year flood" in memory costs (20VC). When memory eats half your bill of materials, squeezing the logic die with a custom design is suddenly worth a year of engineering.

The debate

For ASICs. A seventh credible buyer just went custom, and an OpenAI President said the chip beats general-purpose silicon on cost. Tan's framing, every frontier lab eventually builds its own, now has OpenAI, Google, Amazon, Microsoft, and Meta as evidence. And AI-accelerated design (OpenAI used its own models to help tape out Jalapeño in 9 months) compresses the iteration loop that was custom silicon's biggest disadvantage.

For the merchant. The same operators undercut the panic. Brockman: it's "additive," Nvidia is still the training partner, and "we cannot get compute fast enough." Wedbush's Matt Bryson on TITV: "Microsoft… Meta… have been working on custom ASICs for two, three years now. And you're not seeing a lot of volume." On the moat, the read split: Bryson thinks "that software… mode that [CUDA] created has dissipated a bit, at least in terms of inference… a token is a token." But Bing Xu, MXNet co-creator now building GPU-kernel infra, told The Cognitive Revolution (2026-06-27) the opposite: "CUDA moat is definitely there and… even higher," because auto-optimization needs Nvidia-only tooling, "an accurate profiler… a reliable driver," where rivals have "a large gap."

Where the read comes out this week: Jalapeño confirms the direction (everyone goes custom for inference) without changing the destination (Nvidia keeps training and the systems crown). As Telltales put it (2026-06-28), one custom chip "doesn't break a software moat built over more than a decade, but it changes every procurement meeting." Long the silicon (NVDA, AVGO); the memory cartel (MU and Hynix) is the cleanest beneficiary of either outcome.

Stocks in play

  • AVGO Bull: it now co-owns a marquee OpenAI design win on top of its six XPU customers; AI chip revenue scaling toward >$100B next fiscal year (Telltales). Bear: trades ~68x trailing FCF, priced for flawless execution; Morningstar's fair value is $650 and they note management didn't raise 2027 guidance (The Morning Filter). Watch: the Jalapeño technical report due "next couple of months."
  • NVDA Bull: still OpenAI's training partner; Tan says merchant demand isn't falling. Bear: inference is "50 to 60% of revenue" (20VC), exactly where ASICs aim. Watch: whether Jalapeño's perf-per-watt report lands above or below GB300.
  • GOOGL / AMZN / MSFT / META Bull: each already runs in-house silicon (TPU, Trainium, Maia, MTIA); Odyssey CTO Jeff Hawke confirmed running world-model workloads "on [Amazon's] chips," "very close to their Trainium team" (TITV). Bear: Bryson's volume critique, years in, still light. Watch: Meta's capex guide moved to $125–145B (Telltales); any Maia/MTIA volume metric.
  • MRVL / ALAB / ARM Bull: every new XPU rack pulls connectivity and Neoverse content. Bear/Watch: no fresh Marvell, Astera, or Arm design-win disclosure this week, a coverage gap, not a negative.
  • QCOM Bull: projecting $15B of data-center sales by 2029 and bought CUDA-software challenger Modular for ~$4B all-stock (TITV). Bear: "two-to-three-year chip delay" vs Nvidia (Bloomberg Businessweek). Watch: revenue from the acquisitions, still absent.
  • MU Bull: ~85% gross margin, ~$50B Q guide, ~$100B locked contracts; the toll booth on every accelerator. Bear: it's still memory, Bryson expects "oversupply at some point." Watch: HBM4 qualification cadence.

Read-throughs

  • HBM, Micron / SK Hynix / Samsung. The week's real supercycle. SK Hynix is raising ~$29.4B in a US ADR debut (trading July 10), among the largest share sales ever (Bloomberg Tech). Memory is the input that makes custom logic worth designing around.
  • TSMC / advanced packaging. Jalapeño uses leading-edge TSMC plus CoWoS plus 8 HBM stacks, so OpenAI competes for the same CoWoS Nvidia needs. Cerebras' wafer-scale 5nm approach is the only escape hatch from that queue (Bloomberg Tech).
  • Chip financing. Bloomberg notes Broadcom is "standing up a chip financing vehicle" to help fund OpenAI's tens-of-billions commitment, the ASIC ramp is again being underwritten through capital markets, not benchmarks.
  • Connectivity / ASIC services / Arm. No fresh Marvell, Astera Labs, or Alchip color this week. Watch attach rates as the Jalapeño and TPU capacity lands in '27.

What changed vs last week

Last week (June 22) the ASIC story was a financing story, the $35B Apollo/Blackstone Google-TPU credit deal. This week it became a product story: an actual named OpenAI chip with two executives on camera, plus the disclosure that Broadcom is standing up a financing vehicle for OpenAI (the financing thread continues, now pointed at AVGO/OpenAI rather than GOOGL). Two real deltas: (1) operators showed up, last week was all pundits and a long-Nvidia neocloud CEO; this week Tan, Brockman, and Feldman all spoke. (2) The bull case got more concrete (a chip, a 9-month tape-out) and more bounded (Brockman: additive, Nvidia stays the training partner). Still missing, as ever: a published MLPerf-style benchmark from any ASIC, and hard volume from Maia/MTIA/Marvell/Alchip.