MUMBAI: OpenAI is turning up the heat on AI hardware, and this time the spice is silicon. The company is expanding beyond building AI models with Jalapeo, a custom chip developed with Broadcom to handle the inference workloads generated when those models respond to users.
Unlike training, where models learn from vast datasets, inference is the point at which a trained AI model processes a request and produces an answer. As that workload grows with the use of AI for chat, coding and reasoning, OpenAI is looking to make every response faster and less power-hungry.
Jalapeo was tested with GPT-OSS 120B, DeepSeek R1 and Kimi K2.5 1T. OpenAI said the chip delivered 1.5 to 1.9 times more AI work per watt than the comparison systems used in its tests.
It also recorded 1.7 to 3.6 times lower end-to-end latency, while performance on highly interactive workloads was as much as 4.1 times higher, according to the company.
OpenAI said Jalapeo was developed through a co-design approach spanning the hardware and software stack, rather than as a standalone processor. The work covered the chip, memory, networking and systems required to serve AI models.
Broadcom worked with OpenAI on the custom processor, while Celestica contributed to boards, rack infrastructure and manufacturing hardware.
The development cycle was also relatively short. OpenAI said the journey from initial design to manufacturing tape-out took about nine months, with its own AI models used during parts of the engineering and optimisation process.
The focus on inference reflects a growing challenge for AI companies: building powerful models is only part of the equation. Running them millions of times for users can make computing costs, energy consumption and response times just as important as raw model capability.
Despite developing its own silicon, OpenAI is not cutting ties with external chip suppliers. The company plans to deploy Jalapeo in small volumes by the end of 2026, followed by a broader rollout in 2027. It also plans to develop subsequent generations of its custom hardware.
That leaves Nvidia and other computing partners firmly within OpenAI’s wider infrastructure strategy. Jalapeo is instead intended to add specialised hardware designed around OpenAI’s particular workloads, especially where energy efficiency and latency can directly affect operating costs and the user experience.
OpenAI is not alone in moving deeper into custom AI silicon. Google has developed its Tensor Processing Units, while Amazon has its Trainium and Inferentia chips for specialised AI workloads.
For OpenAI, Jalapeo marks a broader shift in strategy: rather than simply buying computing power off the shelf, it is beginning to design more of the machinery underneath its AI systems itself.
