Nine months is not how long a datacenter accelerator takes to design. It is roughly half that, sometimes less than half. OpenAI told the Hot Chips conference on Tuesday that its hardware team covered initial design through tapeout on Jalapeño in that window, and it credits its own models for much of the compression.
The chip was announced last October and engineered in close partnership with Broadcom. This week OpenAI attached measurements to it for the first time.
Models Writing Silicon
The most striking disclosure has nothing to do with performance charts. OpenAI says AI-generated implementations of selected GPT-OSS attention and mixture-of-experts blocks outran versions written by its own specialists by 1.5 to 1.8 times.
That is a kernel-level result. It does not mean an entire model ran that much quicker. It does suggest the design target was a part that machines can program competently, not only human experts.
The team leaned on the same approach after fabrication. Using Codex with GPT-Astra and earlier models, engineers brought three open-weight models to high performance inside two months. None of the three had been in Jalapeño's original production plan. Shortened design, measurement and verification loops are what OpenAI credits for the schedule.
What the Benchmarks Say
OpenAI ran the part through InferenceX, the public inference benchmark maintained by SemiAnalysis. Three models were used: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T.
Across the trio, Jalapeño did 1.5 to 1.9 times more work per watt. End-to-end latency dropped by 1.7 to 3.6 times. On highly interactive workloads the company puts the margin at 2.1 to 4.1 times.
A far larger figure appears in the deck and needs careful reading. Measured at the previous best time-between-tokens operating point, OpenAI reported 8.6 to 104.3 times more work per watt depending on the model. That is a comparison held at a fixed responsiveness target rather than a general speedup.
Efficiency claims rest on each accelerator's published power rating. Jalapeño carries a 700-watt rating but never pulled above 550 during the runs. The Register reported a full system totals 128 chips, 1.7 exaFLOPS, and 27TB of HBM.
The Architectural Bet
Serving a language model is really two dissimilar jobs stitched together. Chewing through the initial prompt is compute-bound. Emitting tokens one by one leans on memory bandwidth instead. Shuttling data between cores and packages burns time in both.
Operators normally paper over that with batching, which lifts utilization at the cost of making individual users wait. OpenAI designed around the trade instead. Model state, including the KV cache built during generation, can be explicitly placed and pinned locally, while the networking fabric keeps more of a job inside one connected system.
Richard Ho, who leads hardware at the company, told reporters the numbers mark a very significant advance over the state of the art. His framing was that the part can serve many customers per unit of power while still answering quickly.
What It Does Not Settle
The baseline invites scrutiny. Jalapeño was measured against Nvidia Blackwell systems that ship today. Ho put deployment at the end of 2026 in very small volumes, with anything meaningful arriving during 2027.
Nvidia's Rubin generation should be fielded by then. Winning against present-day hardware is not the same as winning on your own launch date. The runs were also conducted by the vendor on its own silicon, which is ordinary industry practice and still not third-party verification.
OpenAI has said it will keep buying from Nvidia and other suppliers regardless. Two more Jalapeño generations are already underway, and the company frames the family as a multigenerational platform where models, chips and memory get planned together. The prize is not supplier independence. It is deciding for itself how the silicon moves as the models change.
