InvestWidely confirmed7 publishers3 min readPublished Updated
OpenAI's first chip beat merchant silicon in 16 months. The moat was never the fab
Jalapeno's lead is measured against last generation and the volumes are tiny, but a first-pass ASIC clearing Nvidia, AMD and Google parts reprices the design barrier, not the supply chain.
The Investor · Invest desk
What happened
- OpenAI showed Jalapeno, its Broadcom-built inference-only ASIC, at Hot Chips and let SemiAnalysis benchmark it in the lab on the InferenceX suite.
- OpenAI says the chip does 1.5 to 1.9 times more work per watt and answers 1.7 to 3.6 times faster than Nvidia GB200 and GB300 racks on three open models.
- SemiAnalysis says it beat every Nvidia, AMD and Google chip the firm has tested, which first-generation silicon normally does not manage.
- Ho says there is no rental market, no instance type and no plan to sell it, because OpenAI needs the compute itself.
Why it matters
- precedent The inputs are the same TSMC N3P and HBM4 supply the merchant vendors queue for, so the barrier OpenAI crossed was design and software integration, and every other buyer with heavy traffic now has...
- exposure Nvidia's exposure here is a customer quietly serving part of its own demand rather than a rival appearing on the price sheet, which is harder for accelerator buyers to see in any comparison table.
- cost With data center power rather than budget as the binding limit, a perf-per-watt gain converts directly into traffic OpenAI can serve, so the in-house part pays even while it carries a sliver of...
- contradiction The firm that ran the benchmarks says the Blackwell comparison is unfair to Nvidia, because the HBM4 peer is Rubin and Rubin systems are already shipping while Jalapeno is at engineering samples.
The two published accounts of how fast this happened do not match. SemiAnalysis dates the program from the middle of 2024 and counts roughly 16 months from first hires to manufacturing tape-out [3]. Richard Ho, OpenAI's hardware vice president, told reporters the company went from design to tape-out in nine months, using its own models to speed the work [4]. The difference is presumably team-building and early architecture versus the design phase proper, and it matters to anyone reading this as a template. Sixteen months is a schedule a well-funded buyer can plan against. Nine months is a claim about AI-assisted design that nobody outside OpenAI has audited.
The scale is smaller than the headline numbers suggest. A 128-chip rack carries 1.7 exaflops of 4-bit compute and 27.5 terabytes of HBM4, and a full pod runs 2,048 chips across a 16-rack scale-up domain [19][20]. That is about 13 petaflops and 215 gigabytes per package [29]. The part is rated at 700 watts but held at or below 550 on the workloads tested [18]. OpenAI's next deployment target is around 100 megawatts [20], which at 550 watts a chip is at most roughly 182,000 chips, or 89 pods, before a single watt goes to networking, cooling or hosts [26]. Set against the October 2025 Broadcom agreement to deploy 10 gigawatts of OpenAI-designed accelerators through 2029 [25], that target is about one percent of the program [27]. Ho said the GPU fleet is not being swapped out and that the compute strategy still leans on Nvidia and Cerebras [13].
So the merchant revenue at risk in the next four quarters is small. The repricing is in the evidence, not the volume, and it cuts both ways. In OpenAI's favour: SemiAnalysis says the perf per watt result was produced without multi-token prediction, while the competing configurations on its chart were each vendor's best-performing SKU with it [14].
Against it: the load-bearing caveat sits in SemiAnalysis's own copy. Every number came from OpenAI, and while the firm watched InferenceX runs in the lab, it did not execute the full suite and has not seen AgentX results [15]. AgentX is the suite it prefers for chip comparisons, because long-context multi-turn traffic stresses routers, prefix caching and offload infrastructure that a single-turn test leaves idle [16]. OpenAI's largest quoted margin, 2.1 to 4.1 times, is on the heavily interactive work that AgentX exists to measure [8]. The biggest claim is the least independently exercised.
The follow-on part, B0, is already in fabrication with roughly 25 percent better performance per watt expected [11]. Stack that on the claimed 1.5 to 1.9 times and a second silicon revision lands somewhere near 1.9 to 2.4 times GB300 [28]. It would still be a comparison against Nvidia's previous generation rather than Vera Rubin [6], and Vera Rubin is the part that other buyers can actually take delivery of.
What to watch
- An AgentX run on Jalapeno executed by SemiAnalysis rather than supplied by OpenAI, on long-context multi-turn traffic.
- Any published head-to-head between B0 and Vera Rubin systems now going to Nvidia's customers.
- Whether the roughly 100 megawatt next deployment target holds as B0 leaves fabrication and 2027 output is booked.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence50
- Adoption8
- Hype gap+40
- Incentives65
- Confidence55
Perspective Coverage
7 publishers- Builder
- Builder 30%
- Operator
- Operator 24%
- Investor
- Investor 46%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
OpenAI published its first benchmark results on Tuesday for Jalapeno, its Broadcom-built inference chip.
- [2]
Jalapeno was announced at Hot Chips, and OpenAI invited SemiAnalysis to its labs to inspect the chip and benchmark it with the InferenceX suite.
- [3]
SemiAnalysis said design work began in the middle of 2024 and went from initial team hiring to manufacturing tape-out in about 16 months.
- [4]
OpenAI said it used its own models to speed the work on Jalapeno, moving from design to tape-out in just nine months, according to hardware vice president Richard Ho.
- [5]
SemiAnalysis said comparison to Blackwell is somewhat incomplete and unfair, because Jalapeno really competes against chips like Rubin that also use HBM4, and Vera Rubin systems are shipping to customers now while OpenAI has nothing beyond engineering samples of Jalapeno.
- [6]
Jalapeno was tested against Nvidia's GB300, not the newer Vera Rubin generation.
- [7]
OpenAI said Jalapeno did 1.5 to 1.9 times more work per unit of electricity and answered requests 1.7 to 3.6 times faster than the Nvidia GB200 and GB300 systems it was measured against, on GPT-OSS 120B, DeepSeek R1 670B and Moonshot AI's Kimi K2.5 1T.
- [8]
For the most hands-on tasks requiring much back-and-forth, OpenAI said the gap grew to between 2.1 and 4.1 times better.
- [9]
SemiAnalysis said first generation chips are generally not competitive, but Jalapeno beat every Nvidia, AMD and Google chip it has been able to test on multiple top open source models.
- [10]
The chip will be deployed in small volumes by the end of 2026 before ramping through 2027.
- [11]
A next-generation version, B0, is already in fabrication and is expected to improve performance per watt by roughly 25 percent.
- [12]
A second-generation chip is approaching tape-out and work on a third generation is underway.
- [13]
OpenAI is not swapping out its GPU fleet; Ho described Jalapeno as one piece of a compute strategy that still leans on very good partners at Nvidia and Cerebras.
- [14]
SemiAnalysis said the headline throughput per all-in utility megawatt result was achieved without multi-token prediction, while the other chips on its chart were the best performing configurations of each SKU, all with multi-token prediction.
- [15]
All the numbers came from OpenAI; SemiAnalysis verified the InferenceX runs in person but did not run the full InferenceX suite and has not seen AgentX results.
- [16]
SemiAnalysis says AgentX is its preferred suite for comparing chip performance because long context and multi-turn datasets stress routers, prefix cache mechanisms, cache management and offload infrastructure that single-turn 8k1k tests do not.
- [17]
The chip is built on TSMC's N3P process with HBM4 memory offering 15.4 terabytes per second of bandwidth, which SemiAnalysis said is currently the highest among shipping or near-shipping accelerators.
- [18]
The chip is rated at 700 watts but held at or below 550 watts on the workloads tested.
- [19]
OpenAI will deploy the chip in racks of 128, with a full pod running 2,048 ASICs; a 128-chip deployment delivers 1.7 exaflops of 4-bit compute and 27.5 terabytes of HBM4, with each package carrying 15.4 terabytes per second of memory bandwidth.
- [20]
Up to 2,048 Jalapeno chips can be connected across a 16-rack scale-up domain, with OpenAI's next deployment target at around 100 megawatts.
- [21]
OpenAI has the budget and the floor space but is currently limited by data center power, which makes tokens per megawatt the metric that counts, and inference runs continuously across ChatGPT and the API so shaving watts per token compounds.
- [22]
On DeepSeek R1 the chip exceeded 700 tokens per second per user at concurrency of one, and Kimi K2.5 and GPT-OSS reached roughly 1,400 tokens per second per user, using single-token prediction without speculative decoding.
- [23]
SemiAnalysis confirmed Jalapeno's GSM8k eval results were on par with Nvidia chips.
- [24]
There is no rental market, no instance type and no plan to sell Jalapeno; asked whether OpenAI would offer it to others, Ho said the company is struggling to have enough compute for its own needs.
- [25]
The chip fits into OpenAI's October 2025 deal with Broadcom to deploy 10 gigawatts of OpenAI-designed accelerators through 2029.
- [26]
A 100 megawatt deployment implies at most about 182,000 chips, or roughly 89 pods of 2,048, before any power is allocated to networking, cooling or host systems.
- [27]
The roughly 100 megawatt next deployment target is about one percent of the 10 gigawatt Broadcom program.
- [28]
A 25 percent performance per watt gain on B0, applied to the claimed 1.5 to 1.9 times advantage, would put the revision at roughly 1.9 to 2.4 times GB300 on the same measure.
- [29]
Per package, a Jalapeno rack works out to roughly 13 petaflops of 4-bit compute and about 215 gigabytes of HBM4.
Sources
7 independent publishers whose own reporting we read for this story.
- cnbc.comOpenAI’s Jalapeño AI chip brings new 'threat' to Nvidia margins as custom silicon gains ground
1 article · August 26, 2026
- cryptopolitan.comOpenAI says its Jalapeño chip beats Nvidia's GB300 on inference
1 article · August 25, 2026
- newsletter.semianalysis.comA detailed look at Jalapeño, OpenAI's ASIC developed with Broadcom in 16 months, which beat Nvidia, AMD, and Google chips on multiple top open-weight models
1 article · August 25, 2026
- proactiveinvestors.comOpenAI's new AI chip outperforms Nvidia's GB300 in efficiency tests, company says
1 article · August 25, 2026
- seekingalpha.comWall Street Lunch: OpenAI's JalapeñO Tops Nvidia's Blackwell In Some Inference Tests
1 article · August 25, 2026
- semafor.comOpenAI says its AI chip rivals Nvidia in inference
1 article · August 26, 2026
- ventureburn.comOpenAI Claims In-House Chips Outperform Nvidia in Early Testing - Ventureburn
1 article · August 26, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.