OpenAI unveils Jalapeño, its first custom AI chip

TL;DR:

• OpenAI has revealed Jalapeño, its first custom AI chip, designed with Broadcom and built for inference.
• Broadcom’s CEO says it matches Nvidia’s Blackwell and Google’s TPUs; deployment is planned by year-end.
• The move deepens a trend of AI labs designing their own silicon to cut costs and reduce reliance on Nvidia.

The scramble for AI computing power is pushing the biggest labs into chip design. OpenAI has unveiled Jalapeño, a processor it developed with Broadcom to handle inference — the work of answering user queries to tools like ChatGPT — and plans to deploy it before the end of the year.

Vertical integration accelerates

The pitch is performance and independence. Broadcom CEO Hock Tan said the chip is “as good as” Nvidia’s Blackwell GPUs or Alphabet’s tensor processing units, while OpenAI hardware chief Richard Ho said it was built to stay performant across “future iterations of LLMs”. OpenAI says samples are already running in its labs against its GPT-5.3-Codex-Spark model, with Canada’s Celestica building the server systems for OpenAI’s exclusive use. Engineers took roughly nine months to complete the design before handing it to TSMC — using AI, the company said, to speed parts of the process.

OpenAI is not alone. Meta, Amazon and Google have all turned to partners such as Broadcom and Marvell for in-house silicon, and Anthropic is weighing a chip of its own. The common thread is cost and scarcity: labs are straining to find enough horsepower for their most demanding models, and an alternative to Nvidia’s GPUs eases both the bill and the bottleneck. The economics remain awkward, though — Tan noted that surging memory demand has squeezed Broadcom’s margins on custom chips, which lean heavily on high-bandwidth memory from SK Hynix and Samsung, an issue echoed in Micron’s recent supply deal with Anthropic.

For the UK, where the government is backing university labs to make AI cheaper to run, the message is that compute efficiency is becoming a competitive battleground, not just an engineering footnote.

Looking forward

Jalapeño is described as the first step in a multi-generation plan, signalling that custom silicon is now core strategy rather than experiment. If in-house inference chips deliver on cost, the prices businesses pay to run AI could ease — but only for those with the scale to design their own hardware, widening the gap between the hyperscalers and everyone renting from them.