OpenAI publishes the first Jalapeno benchmarks at Hot Chips 2026, and SemiAnalysis reports the 700W part beating Blackwell on throughput per megawatt

Sources: SemiAnalysis, Tom's Hardware, discussion, discussion
Summary: OpenAI published the first benchmark numbers for its in-house Jalapeno inference chip at Hot Chips 2026. SemiAnalysis reports the 700W part beating Blackwell on throughput per megawatt, states it verified the InferenceX runs in person, and notes it did not run the full suite and has no AgentX results. Every performance number was supplied by OpenAI, the chip is at engineering samples, and SemiAnalysis argues the Blackwell comparison is unfair because Jalapeno competes with Rubin, which is shipping.
Why it matters: A first-generation in-house inference ASIC reaching flagship-GPU efficiency changes who sets the price of served tokens, though the sourcing does not yet support the claim independently.
Follow-up: Watch for independent third-party runs and for AgentX results.