The tenant buys the building — OpenAI's chip and the race to own the stack

Renting compute makes you a customer. Designing your own chip makes you an infrastructure company. OpenAI just crossed that line, and it tells you what the frontier now believes owning is worth.

OpenAI and Broadcom unveiled Jalapeno, a custom inference accelerator targeting 10 gigawatts of capacity by 2029. Taping out your own silicon is not a procurement decision; it is a declaration that inference at your scale justifies owning the chip, not just the model on it. The tenant is buying the building.

Why now, and why inference

The economics tip at scale. Below a threshold, merchant GPUs are cheaper than a multi-year custom-chip program; above it, the margin on your own inference silicon pays the design cost back many times. Optimizing specifically for serving language models, rather than general training, is OpenAI targeting the workload it runs most and pays most for.

Insurance as much as leverage

And it is a hedge against dependence. OpenAI has stacked roughly 26 gigawatts across NVIDIA, AMD, and its own Broadcom accelerators — a multi-vendor spread that means no single supplier holds its roadmap hostage. A custom chip is the ultimate version of that: the one supplier you can never be cut off from is yourself.

Every prior computing era ended with the platform owners integrating vertically into their own silicon. The frontier labs are running the same play, and Jalapeno is the moment one of them stopped being purely a tenant on someone else's compute.

OpenAI — OpenAI and Broadcom unveil LLM-optimized inference chip → · Data Center Dynamics — OpenAI partners with Broadcom for custom AI accelerators →