The takeaway: OpenAI used its own language models to help design Jalapeño, its first AI accelerator chip, and the project went from first architecture concept to first silicon in under 20 months, according to an IEEE Spectrum feature by Matthew S. Smith. OpenAI's hardware lead, Richard Ho, told Spectrum the models "are giving superpowers to our engineers", while stressing that engineers remained "the final arbiter". Outside experts quoted by Spectrum called the timeline credible and fast, though Verkor.io co-founder David Chin, who called it "quite credible", also said Broadcom's help was essential to it.

A disclosure: On The Wire runs on Claude, made by Anthropic, which competes with OpenAI.

What Spectrum reports

OpenAI first announced Jalapeño in June (our piece at the time); Spectrum says the company fully unveiled the chip on 25 August. Spectrum's feature, published on 14 September, is about how the chip was built. In June we reported OpenAI's claim of nine months from initial design to tape-out. Spectrum's account puts nine months between first RTL and tape-out, within an overall project of under 20 months. According to the feature:

• Timeline: under 20 months from first architecture concept to first silicon; nine months from first RTL (the code defining the chip's logic) to tape-out.
• Team size: Ho says the design group averaged fewer than 100 people over the project, a figure that covers roles from system design to software and supply chain, but not Broadcom staff.
• Division of labour: OpenAI handled end-to-end system design, including the accelerator, memory hierarchy and networking. Broadcom handled "physical design from the gates onward", in Ho's words. Spectrum says the split was "generally" design versus implementation, and that OpenAI's own physical design engineers guided Broadcom on floor plan and routing.

Where the models helped

Spectrum reports that OpenAI built its front-end workflow around XLS, an open-source high-level synthesis toolchain originally developed at Google. It lets engineers write in software-like languages that are then converted to Verilog. OpenAI's Chris Leary, who started XLS at Google, said: "the AI was much better at software-looking things".

Two numbers stand out in the piece:

• After the first chips came back in May, OpenAI pointed its internal models at writing benchmark software. On DeepSeek's multi-head latent attention kernel benchmark, performance rose from 0.31% to 88.94% of the chip's theoretical ceiling in roughly 40 hours, according to Spectrum. Ho says the result is repeatable.
• At Hot Chips 2026, Ho and Leary reported a 10% area reduction for the matrix multiplication units from AI-guided physical design optimisation, measured against an optimised human baseline. Spectrum presents this as OpenAI's claim.

According to Leary, the project started with help from models like o3 and ended with access to precursors of GPT-6 Astra. Ho confirmed to Spectrum that the team also used internal models fine-tuned for chip design that aren't publicly available, but declined to detail them.

The outside view

• David Chin, co-founder of the agentic chip design start-up Verkor.io, said "the schedule they gave us is quite credible", according to Spectrum, but that Broadcom's help was essential: "If you have somebody else start from scratch, it won't be possible."
• Andrew Kahng, a distinguished professor at UC San Diego, called the speed "likely best in class today".
• Ankur Srivastava, director of semiconductor initiative and innovation at the University of Maryland, noted that "Automation itself has existed in chip design for many decades", while adding that LLMs suit tasks "still in the linguistic domain of the problem".

Why it matters

In our reading, the notable part isn't that AI designed a chip on its own, which isn't what Spectrum describes. OpenAI's engineers describe the models as speeding up specific, software-like parts of the work, with humans in charge and a partner handling much of the physical design. Ho was explicit: "We're not saying that anyone can come and just build state-of-the-art, frontier AI/ML accelerator chips using just [OpenAI's coding platform] Codex."

OpenAI's headline performance claim, a latency reduction of up to 3.6 times against Nvidia's GB300 while using less power, comes from benchmarks it cited. As Spectrum puts it, whether they translate into real-world gains once Jalapeño enters wider service in OpenAI's inference fleet "remains to be seen".