OpenAI is moving into its own silicon. The company just revealed Jalapeño, a custom OpenAI inference chip co-designed with Broadcom. It runs large language models like ChatGPT more efficiently. Crucially, it also leans less on Nvidia. It then reaches OpenAI’s data centers in late 2026.
What Jalapeño actually is
OpenAI calls Jalapeño an “Intelligence Processor,” an accelerator built around its own vision for LLM inference. The key word there is inference. This chip runs already-trained models. It does not train them. In practice, that means serving ChatGPT responses faster and at lower cost. OpenAI claims performance per watt “substantially better than current state-of-the-art,” though it admits final testing is still underway.
Built fast, and partly by AI
The timeline is striking. OpenAI says it went from initial design to manufacturing tape-out in about nine months. Broadcom handled the silicon implementation, while OpenAI’s engineers shaped the architecture. Notably, OpenAI also used its own models to help optimize parts of the design. So the company is, in a sense, using AI to build the chips that will then run its AI.
Why OpenAI wants its own chips
The strategy is about money and independence. Running ChatGPT at scale is enormously expensive, and most of that cost flows to Nvidia GPUs. By designing its own inference silicon, OpenAI can tune the hardware to its exact workloads and trim its bills. It also reduces a heavy dependence on a single supplier. That same pressure is reshaping the whole industry, much as Nvidia rethinks its own data-center designs. For a company reportedly eyeing a public listing, lower compute costs matter even more.
What we still don’t know
Plenty stays vague. OpenAI has not named the manufacturing node, shared independent benchmarks, or said how many chips it will deploy. OpenAI also promises a fuller technical report in the coming months. Until then, treat the performance numbers as the company’s own claims. Even so, the message is clear. OpenAI wants to own much more of the stack that powers ChatGPT.
