OpenAI and Broadcom Unveil 'Jalapeño,' a Custom AI Inference Chip

On June 25, 2026, OpenAI and Broadcom jointly announced Jalapeño, OpenAI's first in-house AI inference chip. The chip was designed by OpenAI's hardware team and fabricated by Broadcom's ASIC division on a leading-edge foundry process. From first design review to tape out, the project took only nine months, an unusually short cycle for a chip of this scale. Engineering samples are already running GPT-5.3-Codex-Spark inference workloads inside OpenAI's labs.

The headline numbers explain why the launch matters. Jalapeño cuts inference cost by roughly fifty percent compared with running the same models on the current generation of data-center GPUs. Per-watt performance, the metric hyperscalers care about most, beats the best NVIDIA GPU OpenAI has deployed today. OpenAI says large-scale deployment is planned for late 2026, with Microsoft as the first external customer and a long-term target of ten gigawatts of compute powered by the chip by 2029.

Behind the chip is a deliberate shift in strategy. Until now, OpenAI has bought almost all of its training and inference compute from NVIDIA, with Microsoft Azure as the main cloud host. Building its own silicon for inference gives OpenAI three things at once: lower unit cost, more control over its roadmap, and a way to share silicon design wins with hyperscaler partners so they lock in to OpenAI's models. Crucially, the chip targets inference only. Training will continue to run on NVIDIA's most advanced GPUs, where the software stack is mature and the workloads benefit from huge memory bandwidth.

The wider industry signal is hard to miss. Two years ago, the AI hardware story was almost entirely an NVIDIA story. Today, OpenAI has joined Google, Amazon, Meta and Microsoft in designing custom silicon, with Broadcom and Marvell acting as the design partners of choice for the labs that do not want to build their own chip teams. The era of one chip vendor owning the entire AI compute stack is closing. For users, that competition is likely to mean faster models, lower token prices, and a long-term shift in who captures the profit from running frontier AI at scale.