OpenAI’s Jalapeño Chip: Everything We Know About the First Custom AI Processor Built with Broadcom and TSMC

Performance Claims: Faster, More Efficient, and Much Cheaper

Early testing by OpenAI shows the chip delivers "substantially better" performance per watt than current state-of-the-art AI accelerators . Crucially, Broadcom prices Jalapeño at roughly half the cost of a comparable GPU, with multiple sources estimating overall cost savings at approximately ~50% . Broadcom's CEO, Hock Tan, stated that Jalapeño competes directly with Nvidia's Blackwell and Google's TPUs .

Manufacturing Process & Partners: Nine Months from Concept to Tape-Out

The chip was co-developed by OpenAI and Broadcom, which also manages manufacturing . The design went from initial concept to tape-out in just nine months, described as potentially the fastest ASIC development cycle ever achieved for high-performance advanced semiconductors . TSMC manufactures the chip on its 3nm (N3) process node, incorporating HBM3E high-bandwidth memory . Server maker Celestica will handle board, rack, and system integration around the processor .

Deployment Timeline: Prototype Now, Launch by End of 2026

A physical prototype was delivered to OpenAI on the announcement day . OpenAI targets deployment by the end of 2026, focusing initially on its own data center infrastructure . Jalapeño is the first step in a multi-generational compute platform, with a second-generation chip reportedly already planned for TSMC’s more advanced A16 (1.6nm) process node .

Financing and Infrastructure: Gigawatt-Scale Plans

Building out infrastructure for the custom chip is capital-intensive. OpenAI has secured gigawatt-scale data center capacity to support deployment . The company has also reportedly secured capacity allocation at TSMC for both the current N3 node and the future A16 node, implying significant pre-commitment spending .

Industry Context: Reducing Dependence on Nvidia

Jalapeño fits squarely into a major hyperscaler trend of building custom silicon to displace Nvidia GPUs. Google has its TPU series, Amazon has Trainium and Inferentia, Microsoft is developing its own AI chip, and Meta is pursuing custom inference silicon . Nvidia still commands roughly 80% of the AI chip market, but hyperscalers are increasingly motivated by cost, supply-chain control, and workload-specific efficiency gains . Jalapeño's ~50% cost reduction versus comparable GPUs directly addresses the economic pressure of running inference at massive scale . OpenAI views owning the silicon as essential to "build the full stack" — from models to hardware — and gain a competitive moat beyond software alone .