OpenAI unveils Jalapeño, its first custom chip with Broadcom
An inference processor built to cut the cost per query and loosen Nvidia's grip
Summary
Openai unveiled Jalapeño, its first custom-built inference processor, designed with Broadcom, TechCrunch reported. The chip targets inference, running trained models against user queries, rather than training. OpenAI says early tests show significantly better performance-per-watt than current state-of-the-art parts, but it remains in evaluation with no deployment date. The company framed the chip as one layer of optimizing its infrastructure stack, from architecture to memory and networking. Pre-training is expected to keep running on Nvidia hardware. The move follows Google's TPUs and Amazon's Trainium in building accelerators tuned to a lab's own workloads.
Why it matters
Inference is OpenAI's largest and fastest-growing cost. A chip that lifts performance-per-watt attacks the unit economics of every query and chips at Nvidia's pricing power, the same logic that pushed Google and Amazon to custom silicon. Owning the design also hedges supply during a GPU shortage.
What to watch
- When Jalapeño moves from testing to production deployment.
- Independent benchmarks against Nvidia and custom rivals.
- How much of OpenAI's inference shifts off merchant GPUs.