OpenAI built its own AI inference chip, and for now, it's keeping it to itself.
Richard Ho, OpenAI's head of hardware, told Tom's Hardware Premium that the Jalapeno ASIC, developed with Broadcom, is built primarily for OpenAI's own compute needs. At Hot Chips 2026, the company still showed benchmarks pitting Jalapeno against Nvidia's Blackwell GB200 and GB300 chips, running its open-weight GPT-OSS model alongside DeepSeek R1 and Kimi K2.5. Ho said the demo was meant to counter the idea that Jalapeno only works with OpenAI's own models. But he added that OpenAI's internal compute demand is growing so fast that filling it will keep the company occupied for a good long time before any external customers get a look.
That's the real story here: efficiency and supply, not a chip launch. Ho called efficiency, not benchmark wins, the actual design goal, since a more efficient inference engine stretches power-constrained data centers further. He also framed keeping model codesign in house as a way to avoid leaking research IP to third-party silicon partners, something an outside foundry deal can't fully prevent no matter how many NDAs are signed.
Ho didn't rule out wider availability - as he put it, "you could use it for anybody, honestly" - but until OpenAI's own compute appetite is satisfied, Jalapeno's only customer is the company that built it.