OpenAI has published the first performance results for Jalapeño, its first custom chip designed to serve AI models. The company says the system delivered both higher throughput per watt and lower end-to-end latency than the comparison hardware across three public models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. That combination matters because inference systems normally trade fast interactive responses against efficient batch processing.

The result is a real infrastructure signal, but it is not a customer-facing availability announcement. OpenAI says it plans to begin deploying Jalapeño in its own compute infrastructure by the end of 2026. TechCrunch, reporting from the announcement, says the first deployment is expected to be very small, with more significant volume in 2027. No public product, cloud instance, or purchase channel has been announced.

What OpenAI measured

OpenAI says it used SemiAnalysis' public InferenceX benchmark to compare complete serving systems at a matched user experience, rather than comparing isolated chips. It reports 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency across the three workloads. The company says Jalapeño's nominal 700-watt rating stayed at or below 550 watts under the tested workloads, while the comparisons use published power ratings for the rival accelerators.

  • The published comparisons cover public open-weight models, not OpenAI's unreleased frontier workloads.
  • OpenAI says the results are normalized from InferenceX measurements and published accelerator power ratings; that methodology should be read alongside the benchmark's own documentation.
  • The company attributes the design to a full-stack system: chip, memory, networking, serving software, and model-specific kernels rather than a processor alone.
  • OpenAI says Jalapeño was developed with Broadcom and will coexist with NVIDIA and other accelerators, rather than replacing all outside hardware.

Why the claims need careful reading

The benchmark itself is public, but the presentation and the conclusion that Jalapeño leads the field are OpenAI's. The comparison systems are commercially available hardware available at the time of testing; they will not stand still before Jalapeño reaches broader deployment. Performance also changes with model architecture, quantization, prompt length, batch size, power measurement, software version, and the latency target selected. A strong result on these three models is evidence, not a universal ranking.

OpenAI also says AI-assisted design helped move the chip from initial design to tapeout in nine months and that generated implementations outperformed earlier human-written implementations on selected model blocks. Those figures apply to selected development work, not to the total chip-development process or complete model performance. They are a useful indication of a faster hardware-software feedback loop, but need independent replication before they become a general claim about AI-designed silicon.

The reader takeaway

Jalapeño's importance is strategic: OpenAI is trying to control more of the cost, latency, and power envelope behind agentic products. If the gains survive qualification and production scaling, they could reduce the cost of serving long, multi-step tasks and give OpenAI another lever beyond buying more general-purpose accelerators. For users and enterprise buyers today, nothing has changed in availability. The next evidence to watch is production rollout, operating reliability, and independent tests against the hardware available when Jalapeño actually ships.