OpenAI has spent its entire existence renting its brains from Nvidia. Last week it started building its own. On June 24, the company and Broadcom unveiled Jalapeno, OpenAI's first custom-designed processor, a chip built not to train models but to run them, the high-volume work of answering the hundreds of millions of daily queries that flow through ChatGPT and Codex. The headline figure is not a benchmark but a calendar: from blank page to production-ready silicon in nine months, against an industry norm of two to three years. OpenAI says its own AI models helped with the design, a tidy loop in which the technology is now used to build the hardware that powers it.
The point of Jalapeno is economics, not bragging rights. Nvidia's graphics processors are extraordinary general-purpose machines, which is exactly the problem when your workload is narrow and gigantic. A chip shaped around one company's specific models, kernels and serving systems can squeeze out more performance per watt, and at the scale OpenAI operates, even a small efficiency gain compounds into enormous savings across data centers measured in gigawatts. Broadcom handled the silicon implementation and networking, the contractor Celestica built the physical hardware, and engineering samples are already running real workloads in OpenAI's labs. Crucially, the chip targets inference; OpenAI will keep leaning on Nvidia for the heaviest training, so this is diversification, not divorce.
What makes the moment matter is that OpenAI is not acting alone. Google has run on its own Tensor Processing Units for years. Amazon has Trainium and Inferentia and is now reportedly preparing to sell its custom chips to outside customers. Meta has its own accelerators. Just this month Qualcomm agreed to buy the AI startup Modular for nearly $4 billion in a direct swing at CUDA, Nvidia's software moat, while ByteDance is said to be lining up Chinese chip suppliers and a startup called Architect Labs raised seed funding to use AI to speed custom chip design itself. The era when every serious AI company simply queued for Nvidia GPUs is giving way to one where the largest buyers each want a chip of their own.
Broadcom is the quiet winner of this shift. Rather than competing with Nvidia head-on, it has positioned itself as the partner that helps hyperscalers build bespoke silicon, and the strategy is paying off spectacularly. Its AI chip revenue hit $8.4 billion last quarter, up 106 percent year over year, with the next quarter guided to $10.7 billion and a backlog of $73 billion. Chief executive Hock Tan has set a target of more than $100 billion in AI semiconductor sales by 2027. The OpenAI deal is both revenue and a marquee endorsement that could pull in the next wave of customers.
None of this dethrones Nvidia, whose CUDA software ecosystem and forthcoming Vera Rubin systems keep it dominant in training and locked into developer habits that custom chips do not easily break. But the AI hardware market is quietly splitting in two. Training, where raw power and ecosystem depth rule, remains Nvidia's fortress. Inference, where cost per query is everything, is becoming a contest. With hyperscalers planning to spend more than $700 billion on AI infrastructure this year, even a modest migration toward custom silicon represents a vast pool of money flowing away from the company that has defined the boom. Jalapeno is one chip. The pattern behind it is the story.