OpenAI has taken its first step into custom silicon, unveiling a chip co-designed with Broadcom that it says will lower the cost and increase the speed of running artificial-intelligence models in production.
The processor, named Jalapeño, is built specifically for inference—the stage where trained AI models answer user queries. Unlike training chips, which prioritize raw throughput, inference accelerators balance speed with power efficiency, a distinction that has become critical as OpenAI deploys ChatGPT, its API, and enterprise tools at scale. The companies said mass production will begin in the second half of 2026 and continue through 2029.
The announcement lands at a moment when the biggest AI labs are racing to reduce their dependence on Nvidia, the dominant supplier of graphics processors used for artificial intelligence. While Nvidia remains the backbone of AI training infrastructure, its profit margins and pricing power have drawn increasing scrutiny from cloud customers and hyperscalers. Custom silicon offers a path to lower unit costs over time, though it demands substantial upfront engineering investment.
Broadcom has emerged as a leading alternative for custom AI chip design, already working with Google, Meta, and smaller AI startups. The company's stock has outperformed many peers this year as investors bet that the custom-chip business will grow faster than traditional networking equipment. For OpenAI, the partnership provides a degree of supply-chain control but also deepens its reliance on a single outside manufacturer during a period of intense geopolitical tension around semiconductor exports.
Industry analysts noted that Jalapeño is unlikely to displace Nvidia in OpenAI's training clusters anytime soon. Training the largest models still requires thousands of specialized processors working in parallel, an area where Nvidia's ecosystem and software stack retain a commanding lead. The immediate impact may be felt in operating expenses: if the chip delivers on its promise of faster, cheaper inference, OpenAI could improve margins on API calls and reduce the price per token for developers.
The move also raises questions about whether other major AI players—Anthropic, Google, or xAI—will accelerate their own custom-silicon efforts. So far, most have relied on standard hardware or co-design partnerships, wary of the capital and expertise required to run a semiconductor operation. If Jalapeño succeeds, that calculus could change quickly.
Image source: v3b.fal.media