

The Jalapeño Chip: OpenAI Builds Its Own Silicon
OpenAI spent years buying Nvidia's chips. In August 2026, it announced one of its own.
On August 25, 2026, at the Hot Chips conference, OpenAI unveiled Jalapeño, the first custom inference chip it had designed in-house. Sam Altman summarised the announcement in six words: "we made a chip and it is fast." The chip was co-developed with Broadcom, which handled the silicon implementation and networking, and Celestica, which built the board and rack systems around it.
Jalapeño was built for one job only: running inference for OpenAI's own API traffic, not training, and not resale. OpenAI claimed the chip could match or beat Nvidia's Blackwell-class GPUs on inference efficiency, with roughly 50% cheaper inference than equivalent Nvidia hardware and up to 3.6 times lower latency on some workloads.
The announcement mattered less for the chip itself than for what it signalled: the largest single buyer of Nvidia's chips had decided the margin Nvidia captured on every GPU was now worth building around. Jalapeño is the first step in what OpenAI describes as a multi-generation compute platform, with initial deployment planned before the end of 2026.
Enterprises tend to treat "which GPU vendor" as a commodity decision made once at the infrastructure layer. Jalapeño is a reminder that at OpenAI's scale, inference cost is now the dominant line item in running a frontier model, larger than training cost, and controlling it end to end was worth years of chip design effort.
What it actually meant
Jalapeño proves inference cost, not training cost, is now the dominant expense of running AI at scale, and whoever controls the silicon controls the margin. OpenAI's decision to spend years and real capital designing a chip for one purpose, running its own inference, only makes sense if that cost line had grown larger than the cost of building the chip.
What enterprises get wrong today
Enterprises treat "which GPU vendor" as a commodity decision made once at the infrastructure layer and default to whatever a cloud provider offers. They then get surprised when inference costs at production volume dwarf the training budget they planned around, months after launch.
More from Moments
Ask about this story
OpenAI just did something almost no AI lab has done before: it built its own inference chip instead of buying Nvidia's. Want to know how Jalapeño works, why OpenAI built it, or what it means for the AI hardware market?