OpenAI's Jalapeño inference ASIC stays in-house, with broader rollout left open

OpenAI has built a custom AI inference ASIC called Jalapeño, and for now it is reserved for the company's own use, according to a report from Tom's Hardware.

The same report notes that OpenAI has left the door open to a broader rollout. The company said it will have its 'hands full' with Jalapeño for 'a good long time', a remark that both explains the decision to keep the chip in-house for now and leaves room for a larger deployment later.

The details are thin so far. The report identifies Jalapeño as an inference part serving OpenAI's internal workloads, but it does not disclose a foundry partner, tape-out timing, production volumes, performance figures, or any schedule.

The reason this matters is that inference economics, not training, increasingly decide whether large-scale AI products make money. As agents, long-context requests and real-time multimodal calls multiply the volume of inference, the purchase, power and cooling costs of general-purpose GPUs become the heaviest fixed expense, and silicon tailored to a specific model's compute profile can deliver better efficiency on those workloads.

It is also a direction the industry has been moving in for years: companies with both frontier models and large-scale workloads increasingly shift part of their inference onto chips they design themselves, reducing reliance on any single general-purpose compute supplier.

What to watch next: whether Jalapeño stays an internal experiment or moves into production clusters, whether OpenAI publishes more architectural and performance detail, and whether the talk of being busy reflects a demand pipeline rather than a capacity constraint.