OpenAI has published its first public benchmarks for Jalapeño, the inference chip it designed with Broadcom, and the company says the silicon beats Nvidia's current Blackwell-generation systems on the measure that increasingly decides how much it costs to run a large language model at scale: how much useful work a chip delivers per watt of electricity.
According to figures OpenAI released this week, Jalapeño delivers 1.5x to 1.9x more AI work per watt at peak throughput than the best commercially available Nvidia systems, with 1.7x to 3.6x lower end-to-end latency. For interactive workloads — the kind that power a live chatbot session rather than a batch job — OpenAI says the gap widens to 2.1x to 4.1x. In testing across GPT-OSS-120B, DeepSeek R1 and Kimi K2.5, the chip reportedly hit roughly 1,400 tokens per second per user on GPT-OSS and topped 700 tokens per second on a single DeepSeek R1 request, according to details reported by The Decoder.
Jalapeño has been in development since roughly mid-2024, with fabrication beginning last November. It exists today only as engineering samples — OpenAI plans to deploy it inside its own infrastructure by the end of the year, and the company says it is already at work on second- and third-generation versions with Broadcom.
A comparison drawn on OpenAI's terms
The chip is inference-only; it cannot train models, leaving that work to Nvidia hardware for the foreseeable future. Analysts have also flagged that the headline comparison is not entirely apples to apples: Jalapeño uses newer HBM4 memory, while the Blackwell systems it is measured against still run HBM3e. Nvidia's forthcoming Vera Rubin platform, which also uses HBM4, is the fairer match — and even there, OpenAI says Jalapeño produces more output tokens per megawatt, despite Rubin using a multi-token-prediction optimization Jalapeño has not yet adopted. On total cost of ownership per token, OpenAI's own figures put the two roughly even.
Jalapeño smokes every other chip.
Dylan Patel, Founder and Chief Analyst, SemiAnalysis, via The Decoder
Patel, whose firm closely tracks AI hardware, separately wrote that it is unusual for a company's first attempt at custom silicon to be competitive at all. "Usually first generation chips aren't competitive, but OpenAI is beating Nvidia Blackwell and even Rubin," he wrote, adding that the results raise questions about how durable Nvidia's software advantage remains once a customer of OpenAI's scale can stand up new models quickly on its own hardware.
Nvidia's earnings loom over the timing
The benchmarks landed the same day Nvidia reported fiscal second-quarter results: $96.22 billion in revenue, up 106% from a year earlier and ahead of the roughly $92 billion Wall Street had expected, with data-center revenue climbing 117% to $89 billion. Even so, shares slipped in after-hours trading, extending a pattern in which Nvidia stock has fallen after four of its last five earnings reports despite beating estimates each time. Investors have grown more attentive to custom silicon from OpenAI's own customers: Google's TPU line, Amazon's Trainium3 and Meta's MTIA are all now fielded internally, and Broadcom — Jalapeño's manufacturing partner — has reported more than $30 billion in AI chip bookings in a single quarter, according to figures cited by 24/7 Wall St.
None of that displaces Nvidia in the near term. Rubin systems are already shipping to customers, while Jalapeño remains a sample part running inside OpenAI's own labs. But for a company that has spent the past two years as Nvidia's single largest customer, a credible in-house alternative — even a partial, inference-only one — changes the leverage OpenAI carries in future chip negotiations, and adds to the list of hyperscalers no longer entirely dependent on one supplier for the hardware behind generative AI.
Elsewhere, a separate rundown of the benchmark disclosure reached similar conclusions about what a credible, competitive in-house chip could mean for AI hardware markets going forward.
