Jalapeño: OpenAI and Broadcom Unveil Their First Custom AI Chip, and Nvidia Can Feel the Heat

Nine months. That's how long it took OpenAI to design its first custom AI chip from scratch. Jalapeño, unveiled on June 24 alongside Broadcom, is an ASIC built exclusively for LLM inference, not a general-purpose GPU.

Jalapeño: a chip built for inference, not training

The distinction matters enormously. Inference is every ChatGPT reply, every Codex query, every agentic action. Last year, keeping ChatGPT servers responsive cost OpenAI a staggering $8.4 billion. With 900 million weekly users, that operational cost is projected to reach approximately $14 billion this year.

Broadcom CEO Hock Tan told Bloomberg the chip delivers roughly 50% cost savings per inference token compared to current-generation GPUs. OpenAI itself is more cautious, calling the performance-per-watt "substantially better" than state-of-the-art without providing hard benchmarks yet.

OpenAI wants to own the full AI stack

The chip went from initial design to manufacturing tape-out in just nine months. That speed reflects deep software-hardware co-development and the use of OpenAI models to accelerate parts of the design process. Yes, the AI helped design the chip that will run the AI. We're officially living in an Inception sequel.

Engineering samples are already running ML workloads in the lab, including GPT-5.3-Codex-Spark. TSMC handles manufacturing, Celestica builds the server racks. Tan expects small prototype deployment in late 2026, a ramp-up through 2027, and full-scale production in early 2028.

Nvidia isn't losing sleep just yet

Let's be clear: OpenAI is not breaking up with Nvidia. A $30 billion Nvidia investment in February 2026, plus an agreement to deploy 10 gigawatts of the Vera Rubin platform, keeps the GPU giant firmly at the center of OpenAI's training pipeline. Jalapeño trims the inference bill. That's a different fight.

Wall Street's reaction told the real story. Broadcom stock climbed roughly 2%, while Nvidia dipped just 0.26%. The market gets it: warning shot, not a kill shot. Google has TPUs, Amazon has Trainium, Meta has MTIA, and Mistral AI is designing its own chips too. Custom silicon is now table stakes for serious AI players.

Why Jalapeño matters to you

If OpenAI can halve its inference costs, API prices could drop, subscriptions could stabilize, and the most powerful models could reach far more people. That's the real promise behind the spicy branding.