OpenAI just published benchmarks for Jalapeño, the custom chip it built with Broadcom in nine months. Running tests against NVIDIA's GB300 system, Jalapeño delivered 1.5 to 1.9 times more work per watt and cut response time by 1.7 to 3.6 times across major models including DeepSeek R1 and others.
What it does
Jalapeño runs trained AI models — the part where ChatGPT answers your question. It doesn't train new models from scratch, so OpenAI still depends on NVIDIA for that. Each Jalapeño chip uses 700 watts compared to NVIDIA's 1,200-watt system, packing 216 GB of memory at speeds up to 15.4 TB per second.
When you'll see it
OpenAI plans small internal deployments by late 2026, with larger rollouts in 2027. The company won't sell or rent the chip to anyone else. Two more chip generations are already in development.
Why it matters
Every major AI lab — Google, Amazon, Microsoft, Anthropic — is now building its own chips instead of depending entirely on NVIDIA. Cheaper inference means lower costs per answer, but it also means tighter lock-in: the lab that runs your AI now controls the full stack, from software down to silicon. When OpenAI owns the servers and the chips inside them, switching providers gets harder. You're not just renting the model anymore — you're renting the entire machine it runs on.