OpenAI said it has developed its first custom AI inference processor, a chip built with Broadcom and codenamed Jalapeño, designed to run large language models more cheaply and to reduce the company’s dependence on Nvidia’s GPUs. Fortune reported that the chip is a key step in OpenAI’s evolution from a model company into one that controls more of its own computing.
The processor is optimized for inference, the stage where trained models respond to users, which is the part of AI computing that grows fastest as models move into products used by billions of people. Training a model is a one-time cost; serving it is a recurring cost that grows with usage. Chips that make serving cheaper change the unit economics of the entire AI business.
The name continues OpenAI’s tradition of playful codenames, and the company’s engineers are said to enjoy the joke: the spice is aimed at Nvidia’s grip on AI computing. The chip will not replace Nvidia’s hardware overnight, but it gives OpenAI a lever in negotiations and a hedge against supply constraints.
Broadcom is the design partner, and the choice is telling. The company has built a large business designing custom AI chips for hyperscalers, including Google’s TPUs and Meta’s accelerators, and its engineering teams are among the few capable of delivering chips at the scale OpenAI needs. Broadcom’s custom AI designs are manufactured at Taiwan Semiconductor Manufacturing Co., the same foundry that builds Nvidia’s chips.
The economics of inference favor custom silicon. Nvidia’s GPUs are expensive, and the high-bandwidth memory bolted to them adds further cost. A chip designed specifically for the math of language models—matrix multiplication and attention—can deliver similar throughput at lower cost, particularly for high-volume workloads that tolerate some latency.
OpenAI’s computing needs are enormous. The company operates one of the largest AI infrastructure programs in the industry, and its spending on chips and data centers runs into the billions of dollars a year and is growing. Every workload it can move to cheaper silicon is money it can spend on more compute, more research or lower prices for customers.
The chip is not the first sign of OpenAI’s desire to diversify. The company has signed agreements with Oracle for data-center capacity, has considered AMD’s chips for some workloads, and has been in discussions with Broadcom about custom silicon for more than a year, according to people familiar with the matter. The announcement formalizes a relationship that had been rumored for months.
Analysts said the significance lies less in the chip’s specs than in the supply chain it implies. OpenAI, like its rivals, is building options: its own designs, partnerships with multiple chipmakers and a growing base of data centers it controls directly. That portfolio approach reduces the risk that any single supplier’s delays or price increases stall its products.
For Nvidia, the development is a warning shot. The company’s dominance rests on its GPUs being the default choice for AI, but its largest customers are all building alternatives. Microsoft has its own accelerator program, Google has TPUs, Amazon has Trainium and Meta works with Broadcom on custom designs. OpenAI’s entry completes the set: every major AI lab now has a path away from Nvidia.
The timing matters. Inference demand is surging as AI assistants move into phones, browsers and enterprise software, and the cost of serving those requests is the industry’s largest open question. A chip that cuts that cost by even a third would reshape the economics of every AI product built on OpenAI’s models.
Details of the chip’s performance were limited. OpenAI did not announce production dates or the share of its workloads the processor would carry, and executives said only that deployment would proceed in stages. The company said the processor is part of a broader effort to secure its computing future.
For investors, the announcement is a message about OpenAI’s trajectory. The company that once rented all of its computing is now designing its own silicon, a transformation that has taken less than two years. It is a bet that the biggest AI company should also be a chip company—or at least a chip buyer with real alternatives.
Nvidia has not stood still. Its newest accelerators are designed with inference workloads in mind, and its software stack makes it easy for developers to run models on its hardware with minimal tuning. The company has argued that custom chips pay off only at enormous scale, and that its general-purpose hardware wins on flexibility and speed of iteration. The coming quarters will test that argument as OpenAI’s chip moves from announcement to production.
The broader meaning for the industry is straightforward: custom silicon is becoming the norm, and the era when every AI workload ran on one vendor’s chips is ending. OpenAI’s Jalapeño is a small chip with a spicy name, but it carries a large statement about where the computing power of AI is headed.
