OpenAI reveals its first AI processor: Jalapeño
TL;DR
OpenAI has introduced its first custom AI server chip with Broadcom: Jalapeño. The ASIC targets inference, not training, meaning ChatGPT replies, agent runs, and similar model requests. OpenAI describes Jalapeño as an intelligence processor and the first step in a multi-generation compute platform. Broad deployment is reportedly planned for late 2026.
Nauti's Take
OpenAI is moving the bottleneck in-house. For builders, the chip name matters less than the direction: inference cost, latency, and capacity are becoming product policy.
If you scale on OpenAI, you are betting on its server economics, not just its models.
Briefingshow
Jalapeño shows that OpenAI wants control beyond models and products, down into the cost base of its infrastructure. Inference is the expensive always-on part of AI, especially for agents like Codex. If custom chips make that cheaper and faster, pricing, latency, and product limits could shift directly.
Sources