OpenAI and Broadcom just unveiled their first custom AI chip. It’s called Jalapeño, it’s a pure inference accelerator, and it was designed from scratch specifically for large language models — not adapted from an existing GPU architecture.
What Jalapeño does
The chip is optimized for the workloads behind ChatGPT, Codex, and the API. According to OpenAI, early tests show performance per watt substantially better than current state-of-the-art. A detailed technical report is expected in the coming months.
What makes it interesting: Jalapeño reduces data movement and balances compute, memory, and networking resources to achieve utilization much closer to theoretical peak performance. Broadcom contributes the silicon implementation and networking technology (including Tomahawk chips).
Nine months from design to tape-out
Perhaps even more remarkable than the chip itself is the timeline. Development from initial design to manufacturing tape-out took just nine months. OpenAI calls it the fastest ASIC development cycle ever achieved in high-performance semiconductors. And part of that speed came from OpenAI’s own models helping with the design and optimization process.
There’s a beautiful circularity here: AI models helping design the hardware that will run the next generation of AI models.
The strategic picture
For OpenAI, Jalapeño signals a clear intent to control the full stack — from products to models to hardware. Greg Brockman called it part of a ‘long-term full-stack infrastructure strategy’ to make compute more abundant and affordable.
Initial deployment is planned for late 2026, in partnership with Microsoft and others. Long-term, the goal is powering gigawatt-scale data centers.
For Nvidia, this is a direct challenge. Their GPUs still dominate the inference market, but when OpenAI deploys its own silicon into its own data centers, the power dynamics shift. Google has been doing this with TPUs for years — now OpenAI is following the same path.
My take
This story shows where things are heading: the big AI companies don’t want to depend on a single chip supplier anymore. Whether Jalapeño actually delivers on its promised performance remains to be seen. But the development speed alone — nine months, assisted by their own models — is impressive and demonstrates how AI is accelerating its own infrastructure.
Sources: