OpenAI unveils Jalapeño, its first custom AI chip co-designed with Broadcom

OpenAI has unveiled Jalapeño, its first custom "Intelligence Processor," designed together with Broadcom, as the ChatGPT maker pushes to build out its own AI infrastructure and ease its dependence on Nvidia GPUs. The chip was handed to OpenAI CEO Sam AltmanDealroom has a profile for this one. Try Dealroom → and president Greg BrockmanDealroom has a profile for this one. Try Dealroom → by Broadcom CEO Hock TanDealroom has a profile for this one. Try Dealroom → — the first accelerator in a multi-generation compute platform the two companies are building together.

Jalapeño was purpose-built for inference — the task of crunching data to answer a user's query to a chatbot like ChatGPT. OpenAI designed it from scratch around the large language models that power modern AI apps. "It will be performant on, we think, all kind of future iterations of LLMs," OpenAI hardware chief Richard HoDealroom has a profile for this one. Try Dealroom → told Reuters. OpenAI says early testing shows performance per watt "substantially better than current state-of-the-art," while Tan said the chip is as good as Nvidia's Blackwell chips or the tensor processing units designed by Alphabet's Google, with early samples showing cost savings of roughly 50% versus typical AI GPUs.

The accelerator requires large amounts of high-bandwidth memory — supplied to Broadcom by SK HynixDealroom has a profile for this one. Try Dealroom → and Samsung ElectronicsDealroom has a profile for this one. Try Dealroom → — which weighs on Broadcom's margins on custom AI silicon relative to products like networking switches. Broadcom's Tomahawk networking silicon helps bring the platform to large-scale production.

OpenAI already has samples running in its labs, hitting target power and performance with its GPT-5.3-Codex-Spark model. Engineers completed the design in roughly nine months — what the companies call the fastest ASIC tape-out ever in advanced semiconductors, partly by using AI to speed up the process — before sending it to Taiwan's TSMC for manufacturing. CelesticaDealroom has a profile for this one. Try Dealroom → will build the server systems, which, like the chips, will be used only by OpenAI. The company plans to deploy Jalapeño by the end of 2026 at gigawatt scale alongside Microsoft and other data-center partners.

The move follows Meta, Amazon and Google, which have all turned to the likes of Broadcom and MarvellDealroom has a profile for this one. Try Dealroom → to design in-house chips, while AnthropicDealroom has a profile for this one. Try Dealroom → is weighing a custom AI chip of its own.

Read more: OpenAI · Reuters · Bloomberg · DatacenterDynamics