For years, every ChatGPT response came at a steep hardware cost — OpenAI's bills to Nvidia for server time ran into the billions. Today the company moved to change that equation. Together with Broadcom, it unveiled Jalapeño: its first purpose-built AI chip.
Jalapeño is an inference accelerator, meaning it handles serving responses to users rather than training models from scratch. That distinction matters for the bottom line. Early testing puts its operating cost at roughly half that of typical AI GPUs, with performance-per-watt metrics that outpace current state-of-the-art alternatives.
The development timeline stands out. From initial design to manufacturing tape-out took nine months — reportedly the fastest advanced ASIC development cycle ever recorded in high-performance semiconductors. OpenAI accelerated the work by using its own models to handle parts of the engineering process, a genuinely recursive move that compressed what traditionally takes years into under a year.
First silicon samples are now in OpenAI's hands for testing, with customer-serving deployment targeted for late 2026 and full-scale rollout in the first half of 2028.
For Nvidia, this is a clear warning sign. OpenAI was among its biggest Blackwell customers. Microsoft, Google, and Amazon built their own chips years ago, and now the company behind the world's most-used AI assistant is doing the same. Jalapeño probably won't replace Nvidia entirely in OpenAI's data centers — but it gives the company real leverage at the negotiating table. That's the kind of leverage it has never had before.
