OpenAI shows its first custom chip, Jalapeño, built for inference

OpenAI has shown its first custom chip, called Jalapeño, built with Broadcom under a partnership the two companies announced in October 2025. The chip is built for one job: inference, the work of running a trained model to answer a request. OpenAI is clear that it is not trying to replace Nvidia for the heavier work of pre-training, which will still run on Nvidia hardware. What Jalapeño is meant to do is run existing models more cheaply, and the company says early tests show significantly better performance per watt than the best alternatives available today.

The design is aimed squarely at OpenAI's own workload. The company says the chip is tuned for low operating cost when serving real-time coding models, the kind that power tools like Codex, where latency and price per request matter a lot. OpenAI also says its own models helped design the chip, one more sign of how far AI is moving into its own supply chain. You can read TechCrunch's report for the details.

For now this is a controlled bet, not a finished product. The chip is still in testing, and OpenAI has not given a deployment timeline, production volume, or pricing. The strategic point is vertical integration: by owning the chip, the data center, and the software, OpenAI wants to set its own cost curve rather than pay Nvidia's margin on every token it serves.

Why it matters

If you build on OpenAI's API, cheaper inference per watt is what eventually shows up as lower prices and steadier availability for coding models. But Jalapeño is inference-only and still in testing, so for now Nvidia stays central to training and nothing changes in your bill yet.