OpenAI finally put some numbers behind Jalapeño, its custom inference chip built with Broadcom, and the message to Nvidia is blunt: your Blackwell systems just got lapped. Presenting at Hot Chips on Tuesday, OpenAI showed benchmark results from SemiAnalysis’ InferenceX suite where Jalapeño delivered more tokens per user and higher throughput per kilowatt than currently available state-of-the-art inference processors. Richard Ho, OpenAI’s head of hardware, didn’t mince words on a press call, saying the results show a “very, very significant performance advance over state of the art.”
But the timing deserves scrutiny. That comparison is against Nvidia Blackwell as it exists today. Jalapeño won’t even hit “very small volumes” until late 2026, with meaningful deployment pushed to 2027. Nvidia’s roadmap doesn’t stand still, and by then Blackwell will be old news. OpenAI is essentially benchmarking against a moving target with a two-year head start. The chip could still be impressive when it lands, but claiming victory over today’s hardware is a bit like declaring you’ll beat a marathon runner by starting the race tomorrow.
The technical story is more interesting than the benchmark flex. OpenAI leaned into its full-stack advantage, designing Jalapeño specifically to attack the friction points in inference that general-purpose chips tolerate poorly. The company says the chip minimizes delays during prefill and communication phases, which typically act as bottlenecks. Model state, including the KV cache, can be explicitly placed and kept local while the system activates the right mix of compute, memory, and networking for each phase. That’s the kind of optimization you only get when the people designing the chip are the same people building the models that run on it.
OpenAI has said Jalapeño will be a multigenerational platform, with products, models, chips, and memory developed in concert. That’s a meaningful departure from the traditional model where chip designers guess what workloads will look like years out. If OpenAI can iterate on its own silicon at the pace it iterates on models, the 2027 deployment timeline might just be the beginning of a much longer squeeze on Nvidia’s inference business. The real test isn’t a benchmark chart — it’s whether Jalapeño can scale fast enough to matter before the next generation of merchant silicon catches up.
