OpenAI unveiled Jalapeño - its own accelerator for inference (the computation that happens when AI answers a question). Once the cost of intelligence becomes an edge, the AI business stops being decided by leaderboards alone.
- OpenAI and Broadcom unveiled Jalapeño: OpenAI's first own accelerator for LLM inference (June 24).
- Early tests: performance per watt significantly above the current state of the art.
- Plan: gigawatt-scale deployment with partners, incl. Microsoft, from 2026, as a multi-generation platform.
One chip, handed personally to Sam Altman and Greg Brockman by the leadership of Broadcom. That's how they unveiled Jalapeño - OpenAI's first 'Intelligence Processor', built entirely to the measure of their vision for the future of LLM inference.
No need for mythology around silicon here. The fact is strong enough on its own. In OpenAI's words, the chip was designed from scratch around their deep understanding of how the models work: kernels (the basic computation programs), memory movement, networking and serving schemes. Engineering samples are already running in the lab at target frequency and power.
That shifts the game. OpenAI no longer just makes models and products on top of them - it's designing the infrastructure underneath too: chip architecture, kernels, memory, networking, scheduling and deployment systems. Own the whole stack (all the technology layers together), and every layer gets tuned against the rest.
My takeaway is elsewhere. I run our own servers instead of renting them, for exactly this reason - I want to know what's underneath me, not trust someone else's promise. Whoever designs the chip under their model stops renting their future from someone else's vendor. The rest of us - and we're among them - build on someone else's power bill and someone else's delivery schedule. That's a dependency you won't see in a leaderboard, but it decides who's still standing two hardware generations from now.