OpenAI's first chip, 'Jalapeño', went from design to silicon in nine months

A jalapeño aimed at Nvidia's softest flank

OpenAI unveiled Jalapeño, its first custom processor, co-developed with Broadcom after the two announced a partnership in October 2025. The chip is built for inference, the work of running already-trained models for users, and not for training. That split matters: training is a one-time cost where Nvidia's lead is strongest, while inference is a recurring bill that scales with every query.

The headline claim is "significantly better performance-per-watt than current state-of-the-art alternatives," with the part still in testing and initial deployment slated for late 2026. President Greg Brockman framed the bet as workload knowledge: "We have a deep understanding of the workload... how can we build something that will be able to accelerate what's possible?" The other claim doing the work is that the design went from concept to silicon in nine months, "accelerated by OpenAI's models."

The thread doubts the origin story

That nine-month figure drew the sharpest fire. A commenter writing from a chip-design vantage point dismissed the whole framing.

"I call BS. It's probably a white label around existing Broadcom IP, impossible to go from zero to this kind of chip in nine months…" — a skeptic on the silicon timeline

The "accelerated by our models" line met similar suspicion as marketing without substance.

"I wish there was more about this. As is I kind of have to assume that this is just meaningless marketing, like saying development was accelerated by Microsoft Office…" — a reader parsing the press release

Others shifted to second-order economics. One in the thread argued that since inference is the recurring cost, an inference chip cuts at the thin part of Nvidia's moat, and tied the move to OpenAI's roughly $1.4 trillion in compute commitments. A founder who has shipped hardware tempered the optimism by listing the graveyard of effort behind every working accelerator.

"…building a high-performance chip is not easy. Look at companies like Groq, Amazon, and Google." — a hardware founder

A quieter detail surfaced in the comments that OpenAI's own post skipped: the wafers come from TSMC, surfaced by a security engineer in the thread. If the part holds up, the real test is not benchmarks but whether OpenAI can use it as leverage to renegotiate the commitments that pin it to Nvidia.