OpenAI Publishes First Jalapeño Benchmarks, Claiming Up to 1.9 Times Nvidia's Throughput Per Kilowatt

OpenAI released the first published performance results for Jalapeño, the inference application-specific integrated circuit (ASIC) it co-developed with Broadcom. Across three models, GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T, the company reports 1.5 to 1.9 times more throughput per kilowatt at peak and 1.7 to 3.6 times lower end-to-end latency than the best commercially available systems it tested. For interactive serving, where a single user waits on a response rather than a batch job, OpenAI puts the advantage at 2.1 to 4.1 times.

The power comparison is the part engineers should read closely. Tom's Hardware reported that the 700-watt Jalapeño part was measured against Nvidia accelerators rated at 1,200 and 1,400 watts, so a per-kilowatt win partly reflects each chip's power rating rather than a claim that Jalapeño beats the competing hardware on raw output per package. A chip built for one company's serving stack, with one tokenizer, one batching policy and one set of quantization choices, has structural advantages a general-purpose GPU cannot exploit.

The numbers are vendor-published, which is the standard caveat. What sets them apart from a typical lab blog post is the testing method: the runs used SemiAnalysis's public InferenceX benchmark, and SemiAnalysis verified some of them on site. That is partial third-party attestation, not independent reproduction. No outside party has yet run its own workloads on Jalapeño silicon, and OpenAI is not selling the chip. It plans to deploy it in its own data centers later this year.

The Russian-language technical channel AI ML Big Data circulated the same figures within hours, an indication of how closely non-US engineering audiences track the cost side of frontier serving. Market reaction was contained. Seeking Alpha reported that investors did not appear alarmed, with Nvidia down just under 2 percent and Broadcom higher on the session.

Jalapeño was unveiled with Broadcom in June. The pattern it belongs to is now well established: Google with tensor processing units, Amazon with Trainium, Meta with its own accelerators, and now OpenAI. Each removes a share of inference demand from the merchant market permanently, because the chip exists to serve one buyer's traffic and cannot be resold to anyone else.