OpenAI's Chip Beat Nvidia's: The Edge Is in the Watts

On Tuesday, August 25, OpenAI published measured results for the first processor it designed itself. Against Nvidia's GB300, Jalapeño came out ahead on tokens per kilowatt and on latency: the published ranges are 1.5x to 1.9x and 1.7x to 3.6x.

The same day, Nvidia closed up 2.19% at $213.05.

Put the two side by side and the first explanation is that the market missed the news. A second: it saw the news, and separated the ratio's numerator from its denominator.

A Processor Designed in Nine Months

Jalapeño is not new: OpenAI and Broadcom introduced it on June 24 without performance figures.

What was shared was physics — a compute die of roughly 840 mm², close to EUV lithography's 858 mm² reticle limit, six memory stacks, TSMC's 3nm-class process. Concept to tape-out took nine months, by the companies' account the fastest ever for a chip built for one job (an ASIC).

Jalapeño's one job is inference — running a trained model, not training one.

The frame is the October 2025 agreement to deploy 10 gigawatts of OpenAI-designed accelerators by 2029. Broadcom chairman Hock Tan said at the launch:

"This is just the beginning of a multi-generation roadmap."

Who Ran the Measurement, and Where

The figures come from InferenceX, SemiAnalysis's benchmark suite, on three open models: GPT-OSS 120B, DeepSeek R1 670B, and Moonshot AI's trillion-parameter Kimi K2.5. Workload: an 8,000-token prefill, a 1,024-token decode.

Where it happened matters too. SemiAnalysis writes that the runs took place in OpenAI's lab with OpenAI engineers present, that OpenAI provided the results, and that it did not execute the full suite itself:

"we verified the InferenceX runs in person in the lab"

None of this makes the result wrong. It makes it different evidence than a measurement reproduced in an independent lab.

Two details cut the other way. The 8,000/1,024 shape is, in SemiAnalysis's own words, a "much easier workload to tune for" than their long-context suite. And Jalapeño was measured with single-token prediction while competitors had multi-token prediction on — worth a 3x-plus improvement in cost per token.

Where the Advantage Comes From: The Denominator

"Performance per kilowatt" is a division: tokens over power. A ratio grows two ways — bigger numerator, smaller denominator. Jalapeño is rated at 700W, the GB300 at 1,400W: the denominator is exactly half.

This is not an accusation: in a data center the scarce input is electricity, not chips, so per-kilowatt performance is the right denominator. But a reader who sees "beats Nvidia" understands one chip outperforming another; the measurement does not say that.

The denominator moves the answer again: results were normalized to published TDP, and Jalapeño's measured sustained draw in the lab was 550W, not 700W. All-in facility power narrows the gaps.

Headline figures came from the top of the range; the bottom is part of the same measurement.

The Wrong Opponent

The GB300 it was set against is a 2025 Nvidia design, a memory generation behind. In the token-generating phase of inference the bottleneck is bandwidth, not capacity — so this compares two generations of memory.

Nvidia's HBM4 part is Vera Rubin. SemiAnalysis calls the Blackwell comparison "somewhat incomplete and unfair" and says Vera Rubin is the proper baseline.

In its July 2026 measurements, Vera Rubin NVL72 delivered roughly 5.4x the throughput per megawatt of the GB300 NVL72 at the highest interactivity setting, around 4x at a middle one. Divided by total cost of ownership, it and Jalapeño are head-to-head.

The last row goes missing from most commentary: Jalapeño cannot train a model. OpenAI's hardware lead:

"Nvidia is a really good partner, and we continue to need a lot of Nvidia."

The Fifteen Months Between a Benchmark and a Shipment

SemiAnalysis says Jalapeño production ramps through 2027, most output scheduled for the fourth quarter. The measurement was published in August 2026: volume shipment sits fifteen months out.

Another constraint waits: the entire 2027 memory capacity of Samsung, SK Hynix and Micron is already contracted. Every copy of Jalapeño wants six HBM4 stacks.

A rough upper bound: if the full 10 gigawatts were accelerator power, it would take 14.3 million chips at 700W each — 85.7 million HBM4 stacks. The real number is lower: much of a data center's electricity goes to networking, host CPUs and cooling.

The performance argument is fought at the accelerator layer; the binding constraint sits one layer below, in memory allocation.

What the Market Actually Said

Pinning Nvidia's 2.19% gain on Jalapeño alone would be wrong. The Monday before, the Nasdaq fell 0.8% as semiconductor names sold off ahead of the report, and on Tuesday several brokerages raised price targets.

What can be said is narrower: the market did not price the benchmark as an event that changes Nvidia's income statement. Unpriced news is news not yet seen as an order or an allocation of capacity.

Nvidia reports after the close on August 26; watch gross margin and the composition of data center revenue, where custom-silicon pressure would show up first.

This piece draws on the announcements published by OpenAI and Broadcom, on SemiAnalysis's analysis built on its InferenceX benchmark suite, on Bloomberg's August 25 report, and on Tom's Hardware's technical coverage. The performance figures were supplied by OpenAI, and the party publishing the benchmark stated that it did not independently run the full suite; that distinction is preserved throughout. Power and memory capacity figures for Vera Rubin have not been disclosed publicly. Nvidia's August 25 close was $213.05. This is not investment advice.

Correction (September 13, 2026): prose tightened for rhythm; figures and sources unchanged.