OpenAI and Broadcom on June 24 unveiled Jalapeño, the first custom chip from the maker of ChatGPT, marking the company’s formal entry into silicon design and its most concrete step yet to escape dependence on Nvidia’s supply. Jalapeño is an application-specific integrated circuit built for one job: running large language models. It is also the first product of a multi-generation computing platform the two companies are developing together, according to people familiar with the matter.
The chip is designed for inference — the stage of AI where a trained model answers questions — rather than for training, the far more compute-intensive stage where models learn from data. That choice is strategic. Inference demand grows with usage, and OpenAI’s services generate enormous inference workloads that today run on Nvidia hardware rented from cloud providers. A chip tailored to that workload could cut the company’s largest operating cost, the price of serving every query ChatGPT answers.
OpenAI and Broadcom emphasized flexibility in their announcement. Jalapeño, they said, is designed to adapt to a wide range of large language models, not only OpenAI’s own. The claim matters for Broadcom, which sells custom silicon to multiple customers and builds each design with the customer’s specific needs in mind, and it signals OpenAI’s intent to sell or license its chip design to other companies rather than keep it exclusively in-house. A person close to the company said OpenAI sees silicon as a product line in its own right, not merely a cost-saving measure.
The partnership pairs OpenAI’s model expertise with Broadcom’s record in custom accelerators. Broadcom has designed specialized chips for Google’s TPU program, Meta’s AI infrastructure, and a string of other customers, and it has built a business around ASIC design that has made it one of the quiet giants of the AI supply chain. Its engineers bring experience in high-bandwidth memory integration, networking, and the packaging technology that AI accelerators depend on — exactly the areas where a new chip entrant is most likely to stumble.
For OpenAI, the chip is the latest in a series of moves to control its own compute. The company has signed multi-billion-dollar commitments with cloud providers, explored data-center projects, and negotiated with chip suppliers, all in service of the same goal: enough computing power to keep its models ahead of rivals. Designing its own silicon gives it bargaining power over pricing and supply that no contract can provide. Nvidia’s GPUs remain the default for AI training, and OpenAI said Jalapeño will not replace them for that purpose; the chip attacks the inference layer, where the volume of work is largest.
The economics of inference explain the urgency. Every ChatGPT query, every API call, every model-generated image burns compute, and OpenAI’s usage has grown so fast that its compute bill has become one of the largest line items in its business. Industry estimates put the cost of serving a single large model response at fractions of a cent, but multiplied across billions of daily queries, the total runs into the billions of dollars a year. A chip that delivers the same answers with better power efficiency or lower memory costs would move that number directly into profit, which is why OpenAI’s investors have pushed for silicon independence even as the company’s model team has pushed for more of everything.
The timing reflects the state of the AI industry’s supply chain. Demand for AI accelerators has outstripped supply for two years, and the companies building the biggest AI systems have concluded that relying on a single supplier for the industry’s most important component is a structural risk. Amazon has its Trainium and Inferentia chips, Google has its TPUs, Microsoft has its Maia accelerators, and Meta has its MTIA line. OpenAI, with the most widely used AI product in the world, was the conspicuous holdout; Jalapeño closes that gap.
Analysts said the significance of the announcement lies less in the chip’s specifications, which the companies disclosed only partially, than in the shift it represents. “The largest AI companies are becoming chip companies,” said one semiconductor analyst. “The question used to be whether custom silicon could match Nvidia’s performance; the question now is how fast the custom designs ramp.” Broadcom executives have said they expect custom AI accelerators to capture a growing share of the market, and the OpenAI deal gives that thesis its highest-profile proof point.
The road ahead is long. Jalapeño will need to be manufactured, tested, and deployed at scale, a process that typically takes quarters even after a design is taped out, and OpenAI will need to build the software stack that lets its models run efficiently on the new hardware. Broadcom’s role in that process — from design to supply — makes it one of the most important beneficiaries of OpenAI’s ambition. The chip’s name, chosen by engineers in the two companies’ joint program, is a joke about heat and spice; the stakes it carries are not.
