OpenAI is no longer treating compute as something it simply rents from the outside world. Its first custom AI chip, Jalapeño, is a signal that the next frontier model race will be fought as much in inference economics as in model demos.
According to Axios, OpenAI has begun testing the Broadcom-built chip and plans to use it for customer queries later this year. The first version is designed for inference: the expensive, always-on work of answering ChatGPT prompts, running Codex tasks, serving API calls, and eventually powering agentic products that may sit in loops for minutes or hours. That distinction matters. Training grabs the headlines, but inference is where AI companies meet their margins every day.
The obvious headline is Nvidia dependence. OpenAI has relied heavily on Nvidia GPUs, and the entire AI industry has spent the last three years learning what happens when the most important input to a product is scarce, expensive, and controlled by a small set of suppliers. A custom inference chip does not make Nvidia irrelevant. It does give OpenAI another lever: hardware tuned for the serving patterns it knows best.
That is the real story. OpenAI can design around its own workloads instead of forcing every query through general-purpose accelerators. If Jalapeño performs well, the company can optimize for latency, throughput, power use, networking, and cost per answer in ways that are hard to achieve with off-the-shelf chips. For a product like Codex, where users may kick off long coding sessions, small efficiency gains can compound into a major business advantage.
The move also makes OpenAI look less like a pure software lab and more like a full-stack AI platform. Google has TPUs. Amazon has Trainium and Inferentia. Microsoft and Meta have their own silicon programs. OpenAI now has a credible path to owning more of the stack beneath ChatGPT, even while it remains deeply tied to Microsoft, Broadcom, and the broader cloud supply chain.
There are reasons to stay sober. OpenAI has not published hard performance numbers, broad deployment still has to happen, and custom chips are unforgiving: great silicon strategy can still run into manufacturing limits, software maturity problems, or a newer Nvidia generation that resets the benchmark. The first chip is also aimed at inference, not the giant training runs that define frontier model jumps.
Still, Jalapeño is worth watching because it points to where AI competition is headed. The labs that win will not only have stronger models; they will have cheaper, faster, more reliable ways to serve them at massive scale. For builders, that means the cost curve of intelligence may increasingly depend on hardware decisions made far below the API layer. For investors and enterprise buyers, it means model quality is only half the question. The other half is whether the provider controls enough infrastructure to keep prices, capacity, and product speed from being dictated by someone else's bottleneck.