OpenAI pushes its AI into chip design following the Jalapeño experiment

OpenAI is accelerating its push into the enterprise sector by offering its artificial intelligence models for specialized tasks, including chip design, life sciences, and financial services. Its Chief Financial Officer, Sarah Friar, detailed this strategy on Monday, 7 September in San Francisco, during Goldman Sachs’ Communacopia + Technology conference.

The company aims to move beyond the generalist use of conversational assistants by tailoring its systems to specific professions. According to Sarah Friar, OpenAI is also experimenting with pricing models based on the results achieved by companies rather than solely on usage volume, in a context where clients demand more measurable returns on their AI investments.

To illustrate this approach, the executive cited Jalapeño, OpenAI’s first inference chip. She indicated that the company’s models were used to accelerate its design and enabled reaching the “tape-out” stage—the point when the final chip design is sent for manufacturing—in nine months.

OpenAI officially unveiled Jalapeño on 24 June alongside Broadcom. The company describes this processor as the first element of a multi-generation computing platform, with Broadcom handling silicon implementation and Celestica managing the integration of cards, racks, and systems.

The group plans an initial deployment of Jalapeño before the end of 2026. At the end of August, OpenAI stated that its initial tests showed better throughput per kilowatt and latency results than commercial systems used as benchmarks, while noting that several generations of chips are already in development.

A battle over AI costs

This push toward specialized uses comes as OpenAI faces growing competition from open or open-weight models, notably Chinese ones,as well as players like Anthropic. Sarah Friar said that the price of the Luna model was recently cut by 80%, which reportedly led to about a tenfold increase in its usage.

The CFO also claimed that deploying Luna could be less expensive than some Chinese models run via cloud providers, citing Z.ai’s GLM 5.3. This comparison is based on OpenAI’s estimates and was not accompanied by a publicly released independent audit during this presentation.

According to figures shared by Sarah Friar, Codex now has 25 million users. Enterprise revenue reportedly grew by 32% between June and July, compared to 20% for annualized total revenue over the same period, while consumer and enterprise activities each already accounted for about half of revenues by mid-year.