OpenAI has published its first performance results for Jalapeño, a custom chip designed for AI inference. Inference is the stage where an already-trained model processes a prompt and produces a response, so improvements in this part of the system can affect speed, capacity and electricity use.

OpenAI’s results are company-reported benchmarks, not an independent certification. Even so, the announcement matters because it shows how major AI companies are moving beyond buying standard accelerators and are beginning to co-design chips, software, networking and data-centre systems around their own workloads.

What OpenAI says Jalapeño achieved

OpenAI tested Jalapeño on a public benchmark called InferenceX using GPT-OSS 120B, DeepSeek R1 and Kimi K2.5. The company said the chip delivered 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 than the comparison systems across those tests.

Those figures should be read in context. Benchmark results depend on the models, operating settings and comparison hardware. They are useful indicators of OpenAI’s testing approach, but they do not prove that every model or every user will see the same improvement.

Why inference chips are becoming important

Training a large model is expensive, but serving millions of daily requests is also a major infrastructure challenge. AI agents can make that harder because one task may involve several sequential model calls. Lower latency can make a system feel more responsive, while better power efficiency can help an operator serve more requests with the same electricity budget.

OpenAI says Jalapeño was designed around the way language-model inference moves between prompt processing, response generation, memory and networking. The company argues that reducing data movement between those parts of the system is central to making interactive AI faster.

What happens next

OpenAI says it plans to begin deploying Jalapeño in its own computing infrastructure by the end of 2026. It also says it will continue using accelerators from NVIDIA and other partners for training and inference.

The bigger story is not that one chip will replace every alternative. It is that AI providers increasingly want more control over the hardware and software used to deliver their products. Whether that produces lower costs or better user experiences will depend on real deployment results over time.

Source

• OpenAI — Jalapeño’s first results show industry-leading speed and efficiency in AI inference

Last checked: August 26, 2026. Performance figures are OpenAI-reported benchmark results and deployment plans may change. For a correction, contact contact.globalledger@gmail.com.