OpenAI has revealed the OpenAI Jalapeño chip, its first custom processor for running AI models, unveiled with Broadcom on Wednesday, 24 June 2026, and built specifically to make large language models faster and cheaper to serve.
What the OpenAI Jalapeño chip actually does
Jalapeño is what engineers call an inference chip, meaning it handles the moment a trained model answers your prompt rather than the heavy training that comes first, as reported by CNBC.
OpenAI designed it from scratch around the way its own models behave, instead of buying general-purpose hardware off the shelf.
The chip was built with Broadcom and manufacturing partner Celestica, and it is a reticle-sized ASIC, which in plain terms means a very large, application-specific piece of silicon tuned for one job.
That job is squeezing more answers out of every watt of power the data centre burns.
Why nine months is the headline number
OpenAI says it took Jalapeño from initial design to manufacturing tape-out in just nine months, which it believes is the fastest ASIC development cycle ever pulled off in high-performance chips.
For context, custom silicon projects usually crawl through several years before a wafer ever reaches a factory floor.
The partnership behind the chip dates back to October 2025, when OpenAI and Broadcom first announced they would build custom hardware together. Going from that handshake to a finished design in well under a year is the kind of pace that usually only happens when a company is desperate for compute.
Early testing points to performance per watt that lands substantially better than current state-of-the-art.
The design wins partly by cutting down data movement, the quiet energy drain that comes from shuttling information between memory and compute, and balancing those resources more tightly than off-the-shelf parts manage.
What the OpenAI Jalapeño chip means for the AI race
The move pushes OpenAI toward owning its full stack, the chips, the models and the serving systems, in a strategy that echoes the way Apple controls both its silicon and its software.
Reducing reliance on outside chipmakers also softens the supply crunch that has shadowed the whole AI sector.
OpenAI calling its product an Intelligence Processor rather than a GPU is a deliberate flex, a way of saying this silicon is purpose-built for thinking machines rather than borrowed from the graphics world.
The branding matters less than the economics, though, because cheaper inference is what keeps chatbots affordable to run.
Jalapeño is slated to start rolling out in OpenAI’s systems from late 2026, and the company has framed it as the first chip in a multi-generation platform rather than a one-off.
Whether it dents the dominance of established chipmakers will only show once those first units go live.
