OpenAI says its first chip beats Nvidia's Blackwell on efficiency

OpenAI has published benchmarks claiming that Jalapeño, its first custom-designed AI chip, outperforms Nvidia's Blackwell systems on efficiency — a first-generation part beating the incumbent's flagship at the specific job it was built for.

According to OpenAI's figures, Jalapeño delivers 1.5 to 1.9 times more AI work per watt at peak throughput than the best commercially available systems, with 1.7 to 3.6 times lower end-to-end latency and up to 4.1 times higher performance on interactive workloads. The tests ran three open models of very different sizes: GPT-OSS 120B, DeepSeek R1 670B and the trillion-parameter Kimi K2.5. The research firm SemiAnalysis verified some of the benchmark runs on-site, and its assessment carried unusual weight: first-generation chips are usually not competitive, it noted, and OpenAI's is beating Nvidia.

The caveats matter as much as the numbers. Jalapeño is an inference-only part — it runs trained models but cannot train them, which leaves the training market where Nvidia is strongest untouched. SemiAnalysis also pointed out that the fairer comparison is not Blackwell but Nvidia's newer Vera Rubin platform, since both it and Jalapeño use HBM4 memory; Jalapeño still comes out ahead on token throughput per megawatt there, but Rubin ships optimizations the OpenAI part lacks. And the chip is not yet a product: it exists as engineering samples, not hardware deployed at scale.

The speed is the striking part. Working with Broadcom, OpenAI went from first design work to a final blueprint in roughly nine months, and from start to fabrication in about sixteen — a schedule that, if the benchmarks hold up in production, compresses what has historically been a multi-year moat into a single product cycle.

The timing sharpens the message: the claims landed in the week of Nvidia's earnings report, and they extend a pattern rather than start one. Google has designed its own TPUs for a decade, Amazon builds Trainium, and Anthropic said this month it is assembling a custom silicon team of its own. Nvidia still sells the hardware that trains essentially every frontier model. But inference — the business of serving models to users, which scales with every new customer — is exactly where its biggest customers are now building their own alternatives, and Jalapeño is the first such part whose maker claims, with a third party partially concurring, to have beaten Nvidia on the first try.