OpenAI has spent two years quietly building its own AI chip. This week, at the Hot Chips 2026 conference, it finally showed the numbers — and let the analysts at SemiAnalysis into its lab to check them.

The chip is called Jalapeño, and it is not a general-purpose GPU. It is an inference ASIC: silicon designed to do one job — running large language models in production — as cheaply per watt as possible. OpenAI co-designed it with Broadcom, built it on TSMC's N3P process, and paired it with next-generation HBM4 memory.

The claim

On the metric that matters most to a company spending billions on compute — performance per watt — OpenAI says Jalapeño beats Nvidia's Blackwell "across almost all scenarios." SemiAnalysis, which verified the results in person using its own inference benchmark suite, reports the chip reaching nearly double a GB200's throughput per megawatt — a power-normalised efficiency measure, not raw speed — on one open-model workload, and more than 700 tokens per second for a single user running DeepSeek's R1.

For OpenAI, the appeal is obvious. Every token it serves today runs on Nvidia hardware, at Nvidia's margins. A chip it owns, tuned to its own models, is a route off what the industry has started calling the "Nvidia tax."

The caveats — and they matter

This is the part a headline usually drops.

The numbers are OpenAI's own. SemiAnalysis verified them in person — which is more than a press release, but it is not independent third-party benchmarking. The analysts themselves call the Blackwell comparison "somewhat incomplete and unfair": Jalapeño uses HBM4, and Blackwell does not. The like-for-like rival is Nvidia's next chip, Rubin — which isn't out yet.

And it is not shipping. Jalapeño taped out in November 2025; what appeared at Hot Chips is engineering samples. Volume production ramps "gradually through 2027," with most output landing at the end of that year. In chip terms, this is a demo with a roadmap, not a product on a shelf.

Why it lands today

The timing is not subtle. Nvidia reports its own quarterly earnings tonight, with roughly $91 billion of revenue guided — a figure built almost entirely on everyone, OpenAI included, buying its chips. Jalapeño is a signal that the biggest customers are also, patiently, becoming competitors.

OpenAI is not breaking new ground here so much as joining a road others paved: Google has shipped its own TPUs for years and Amazon has Trainium in production. Others have found it harder — SemiAnalysis notes that Meta's and Microsoft's in-house AI-chip programmes have struggled to get off the ground despite longer runways. What is new is that the company generating the most demand for Nvidia's chips is now, credibly, building an alternative to them.

None of it dents tonight's numbers. Custom silicon is a 2027-and-beyond story, and Nvidia's lead in the meantime is vast. But the direction of travel is clear: the companies renting the most compute are all, quietly, building their way toward renting less of it.