> ## Content Index
> Fetch the complete content index at: https://www.metatalks.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# OpenAI designed its first AI chip with in-house models it has not released
- URL: https://www.metatalks.ai/openai-first-ai-chip-designed-with-unreleased-in-house-models/
- Published: 2026-09-18T14:56:00.000Z
- Updated: 2026-09-18T14:55:59.000Z
- Author: Al
- Tags: News, Semiconductors, #newswire

**Jalapeño, the accelerator OpenAI unveiled in August, went from first architecture concept to first silicon in under 20 months, with an OpenAI team averaging fewer than 100 people.**

OpenAI's own large language models helped its engineers design Jalapeño, its first AI accelerator. Richard Ho, OpenAI's vice president of hardware, told [IEEE Spectrum](https://spectrum.ieee.org/llms-for-chip-design?ref=metatalks.ai) the team had access to internal models fine-tuned for chip design and not available to the public, and declined to say which. Engineers drove the work and remained the final arbiter, he said; nobody, he added, builds a frontier accelerator on Codex, OpenAI's commercial coding platform, alone.

The team built the front end, the logic-design stage, around what the models were good at. Its designers wrote logic in software-like code, and an open-source toolchain called XLS turned that into Verilog, the language chips are specified in. Chris Leary, who started XLS at Google and now works at OpenAI, said the models did much better on work that looked like software.

At Hot Chips 2026, Ho and Leary said AI-guided optimization had packed the chip's matrix multiplication units into 10 percent less area than an optimized human baseline. Leary said a later model, available to the team by the end of the project, writes Verilog directly, skipping XLS's translation step, and is close to operating proprietary design tools on its own.

OpenAI unveiled Jalapeño in full on 25 August. In benchmarks the company cites, the chip cuts end-to-end latency, the time from prompt to last token, by up to 3.6 times against Nvidia's GB300, which OpenAI currently relies on. IEEE Spectrum says it remains to be seen whether that advantage carries over once the chip is in widespread service.

First silicon arrived under 20 months after the project's first architecture concept, by OpenAI's count. Andrew Kahng, a UC San Diego professor, recalls a 2016 design-automation workshop where Ho, then at Google, argued that a chip's design time comes down to how many iterations a team can run in a day. Kahng called OpenAI's pace likely the best in class today.

Broadcom handled physical design from the gates onward and carried the chip through production. The OpenAI group averaged fewer than 100 people, Ho said, not counting Broadcom's staff. David Chin, a co-founder of the chip-design startup [Verkor.io](https://verkor.io/?ref=metatalks.ai), said the timeline is credible but that Broadcom's help was essential: a team starting from scratch could not have matched it.