OpenAI has built its first chip from the ground up and it does one specific job – handling ChatGPT and Codex requests without routing every one of them through Nvidia hardware.
Called Jalapeño, the chip was developed with Broadcom and marks OpenAI’s entry into a group that every other major AI lab joined years ago, the group of companies that design their own silicon instead of buying all of it from outside vendors.
What Jalapeño Does
Jalapeño is an ASIC (Application-Specific Integrated Circuit), which means it was built to do one job well rather than many jobs adequately. And this job is inference, the process of running an already-trained AI model to answer a live user query. However, training new models from scratch, which is far more computing-intensive, will still depend on Nvidia chips for now.
OpenAI calls Jalapeño an “Intelligence Processor,” attesting its architecture was shaped around how its models actually move data during a response and not built as a general-purpose chip later adapted for AI work.
Richard Ho, who leads OpenAI’s hardware team, said the chip was tuned to “efficiently execute our most important workloads close to the hardware”s theoretical limits.” Engineering samples are already running real workloads in OpenAI’s labs, including a model called GPT-5.3-Codex-Spark.
OpenAI also says early testing shows the chip delivers far better performance per watt than current top hardware, and Broadcom CEO Hock Tan has said it performs in line with Nvidia’s Blackwell chips and Google’s own AI processors, while cutting costs by roughly half compared with standard AI GPUs.
“Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable for people and businesses, and can be used to solve more important problems,” said Tan.
Why OpenAI Needed Its Own Chip
The pressure behind this move is, no doubt, financial. OpenAI brought in about $13 billion in revenue in 2025 and lost roughly $21 billion in the process, a gap driven largely by the cost of compute.
Nvidia GPUs are expensive and they’re often reserved months in advance, and this scarcity has limited how cheaply OpenAI can serve its models to hundreds of millions of users. It also makes sense as every other frontier-scale AI company already builds some of its own chips, including Google who has its TPUs, Amazon with its Trainium, Microsoft with Maia, and Meta with its MTIA. OpenAI was the last of the major labs still buying all its silicon from outside the company.
Jalapeño Was Built in Nine Months
OpenAI and Broadcom took Jalapeño from initial design to manufacturing tape-out in nine months, a pace both companies describe as the fastest ever achieved for a high-performance chip of this complexity. Part of that speed came from OpenAI using its own AI models to accelerate parts of the design and testing process. This means the technology that will eventually run on Jalapeño helped build it.
Where the Chips Go From Here
OpenAI plans to deploy Jalapeño at gigawatt-scale data centers by the end of 2026, with Microsoft as the lead partner. Reports indicate Microsoft has committed to buying around 40% of the first production run. Both companies describe this as the opening chip in a multi-generation platform, with OpenAI aiming for 10 gigawatts of custom compute capacity by 2029.
However, Nvidia is not going anywhere soon. OpenAI still needs its GPUs for training, and a full technical report on Jalapeño’s real-world performance is yet to be published. So far, what has changed is OpenAI now having a say in its own hardware roadmap for the first time. Also, how well Jalapeño performs at scale will shape how OpenAI will be perceived in the broader AI market.
