OpenAI and Broadcom have unveiled Jalapeno, OpenAI's first custom AI chip, an accelerator built specifically for running large language models rather than training them. Both companies confirmed the announcement, framing it as the first step in a multi-generation compute platform the two are building together. The point of a chip like this is narrow and important: inference, the work of actually answering a prompt, is where most of the cost lives once a model is deployed, and a chip tuned only for that job can do it more cheaply than a general-purpose GPU.
The most striking detail is how fast it came together. OpenAI and Broadcom say they went from initial design to manufacturing tape-out in roughly nine months, which they describe as possibly the fastest development cycle ever achieved for a high-performance chip of this kind. Part of what made that pace possible, by OpenAI's account, is that the company used its own models to accelerate parts of the design and optimization process. That is a quietly notable claim on its own: an AI lab using its current models to help build the hardware that will run its next ones.
On the technical side, the architecture is aimed at the thing that actually limits inference performance, which is moving data around rather than raw compute. Jalapeno is designed to reduce that data movement and to balance compute, memory, and networking so that real-world utilization sits much closer to the theoretical peak, where most chips fall well short. Early testing, again OpenAI's own, points to performance per watt substantially better than current state of the art. The plan is for initial deployment by the end of 2026 and expansion in the years after, with reports that Microsoft is set to take around 40 percent of the output.
The reason this matters goes beyond one chip. Nvidia's grip on AI has rested on selling the GPUs that nearly everyone trains and runs models on, at margins that turn every token served into a payment upstream. Google built its TPUs and Amazon built Trainium and Inferentia for exactly this reason: at OpenAI's scale, designing your own silicon is cheaper than renting someone else's forever. Jalapeno is OpenAI joining that club, an attempt to own more of the stack underneath its products so that serving intelligence costs less and depends less on a single supplier.
The honest read comes with limits. This is an inference accelerator, not a training chip, so it does not touch the part of the pipeline where Nvidia is most entrenched. The performance figures are OpenAI's own and have not been independently tested, the chip is not yet running at scale, and custom silicon has a long history of looking better on a slide than in a data center. But the combination of the nine-month timeline, a named hyperscaler buyer, and a clear strategic motive makes the signal hard to miss. The companies that can afford to build their own chips are doing it, and the economics of who pays whom in AI are starting to shift.