OpenAI has its own chip now, and on the benchmark it beats Nvidia's Blackwell by up to 1.9 times on work per watt
OpenAI published the first live measurements of Jalapeño, the inference chip it designed with Broadcom: 700 watts per chip, 13.4 petaflops of MXFP4 maths, 15.4 terabytes a second of HBM4 memory bandwidth, 128 chips to a rack. Run on SemiAnalysis's public InferenceX suite against GPT-OSS 120B, DeepSeek R1 and Kimi K2.5, it returned 1.5 to 1.9 times more throughput per watt than the Nvidia Blackwell systems it was measured against, and cut end-to-end response times by 1.7 to 3.6 times. Design work started in mid-2024 and taped out in November 2025; the first units go into OpenAI's own data centres in very small volumes at the end of this year, with the real ramp through 2027.
Why it mattersNvidia's margins rest on there being nothing else to buy. A customer spending tens of billions a year building its own alternative changes that arithmetic before a single rack is installed. Two cautions travel with the numbers: SemiAnalysis puts Jalapeño level with Nvidia's next-generation Vera Rubin rather than ahead of it, and by the time the chip ships in volume it will be racing Nvidia's 2027 parts, not today's.
▶ Related video: OpenAI Just Built Its First AI Chip: Jalapeño Explained
Read in the app — free, in 9 languages
Related stories
Nvidia says its new rack does 30 times more agent work per megawatt — it ran the test itself, and the benchmark's authors have not signed off2026-08-25ByteDance has folded its coding tool and its agent builder into one app, and pointed it at the office
2026-08-25A lab lost the 40,000 chips it was going to train on. So it sold the factory to the company that makes them.
2026-08-25Xiaomi has designed three of its own chips, and the point of all three is running its models on your device instead of in a data centre
2026-08-24China's biggest single cheque for robot bodies: $900 million at a $6.3 billion valuation, for a humanoid due in production this year
2026-08-24