OpenAI Just Built Its Own Chip — and It Used AI to Design It in 9 Months | Edition 292

Nine months. That's how long it took OpenAI and Broadcom to build Jalapeño — OpenAI's first custom chip for running AI models. A typical custom chip takes two to five years from design to tape-out. OpenAI did it in nine months, and they used AI to help design it.

Broadcom's press release described it plainly: "what may be the fastest ASIC development cycle ever." That's not marketing hedging. That's an engineer admitting they broke a record they weren't sure they'd break.

The nine-month timeline is possible partly because AI models were used to accelerate the chip design process itself. That's the feedback loop worth paying attention to: AI helping design the infrastructure that runs AI. The chip was built for gigawatt-scale deployment — OpenAI and Microsoft plan to run enough of these to consume roughly a gigawatt of power by 2026. That's a number that illustrates how enormous inference demand has become.

On performance, OpenAI says Jalapeño delivers performance per watt substantially better than current state-of-the-art inference hardware. Detailed benchmarks haven't been released yet. Claims that appeared in secondary coverage about specific percentage cost savings over Nvidia weren't found in primary sources and weren't used here.

OpenAI has been almost entirely dependent on Nvidia for compute. Every API call, every model run, every ChatGPT response has passed through Nvidia hardware. Jalapeño changes that for the inference side — the part you actually touch when you use ChatGPT or the API. Whether this efficiency gain reaches users as lower prices is a business decision, not a guaranteed outcome. OpenAI hasn't committed to specific price cuts tied to this chip. What's confirmed is a substantial performance-per-watt improvement — the path from there to your API bill is still OpenAI's to decide.

The practical significance isn't just cost. Owning your inference silicon means OpenAI controls the full stack for the product they sell. That's a different kind of leverage — over their own roadmap, over capacity planning, and over what they can optimize for — than renting compute from a supplier.

Nine months to tape-out is a headline. The method — using AI to design the chip that will run AI — is the story underneath it. Whether Jalapeño eventually lowers what you pay for API access depends on decisions OpenAI hasn't announced yet. What's already true is that the company no longer needs Nvidia's permission to control its own inference stack. That changes the leverage in a relationship that the entire industry has been watching for three years.