OpenAI's CFO says the firm's own models taped out Jalapeño in nine months

OpenAI is selling its models into chip design, life sciences and financial services, chief financial officer Sarah Friar said on 7 September at Goldman Sachs' Communacopia + Technology Conference in San Francisco. The proof point she used was internal. OpenAI's own systems, she said, helped design Jalapeño, the inference accelerator the company is building with Broadcom, and the chip was taped out in nine months. Tape-out is the handoff from design to the factory.

Jalapeño was unveiled earlier this year as the first chip in a multi-generation compute platform with Broadcom and Celestica. Richard Ho, who leads OpenAI's hardware programme, has said the architecture was built around the kernels, memory movement and serving patterns the company's models actually use. Hock Tan at Broadcom has tied the same project to gigawatt-scale halls with Microsoft and other partners. Friar's new claim is narrower and more useful: the design cycle itself ran on OpenAI software, and it finished in three quarters.

That is the pitch to semiconductor teams who now treat large models as a drafting tool rather than a chatbot. If a frontier lab can shorten its own tape-out, it can sell that workflow to firms that still measure chip projects in years. Friar placed the same argument next to two price facts. OpenAI cut the price of Luna, its lower-cost model, by 80 percent and saw usage rise about tenfold. Codex, the coding product, has 25 million users. Enterprise revenue grew 32 percent from June to July, against 20 percent growth in overall annualised revenue in the same window.

The cost fight with open-weight models

Open-weight systems from Chinese labs have been the cheap alternative for companies that do not want to pay frontier rates. Friar said OpenAI's lower-cost models are now cheaper than those alternatives on the major clouds. She did not publish a full price card at the conference. The 80 percent Luna cut is the number she was willing to put on a stage. A tenfold rise in usage after a cut that large is what a volume business is supposed to look like. It is also what a company with a large compute bill needs if it wants the enterprise line to keep outrunning consumer ChatGPT.

Friar told investors in August that enterprise had already overtaken consumer in OpenAI's revenue mix, after starting the year at roughly 60-40 the other way. Annualised revenue has been reported around $40 billion. Those figures sit behind the Communacopia remarks. They explain why she spent her minutes on chip design and Codex rather than on consumer features.

What nine months does and does not prove

A nine-month tape-out is fast for a new accelerator. It is not a yield number. It is not a performance-per-watt number against Nvidia's next part. Nvidia remains OpenAI's main training partner. Friar has said the next large training run will use Vera Rubin, with AMD and Cerebras already in the mix and the Broadcom line as the company's own hedge. Jalapeño is an inference bet. The nine-month story is meant to show that the same models that write code can also finish a chip.

Buyers in EDA and foundry flows will want to know which stages the models touched: RTL, verification, place-and-route, or the documentation around them. Friar did not break that down on the Goldman stage. The useful public fact is simpler. OpenAI is no longer describing chip design as a future market. It is describing a finished tape-out, a price cut that moved Luna's traffic, and an enterprise book that is growing faster than the consumer one. Those are the three numbers from 7 September. The rest is a sales calendar.