OpenAI's Jalapeño chip joins Big Tech's break from Nvidia

OpenAI announced Jalapeño this week, a custom inference chip developed in partnership with Broadcom designed specifically for inference workloads. The move marks OpenAI's first public entry into custom silicon and signals that even the most GPU-hungry AI labs are now hedging against single-supplier risk. Google, Apple, and SpaceX have all shipped custom AI accelerators in recent years, but OpenAI's announcement underscores how structural the shift has become.

Nvidia holds roughly 80 percent of the AI accelerator market, a position that has made GPU access a bottleneck for scaling inference at the pace OpenAI's product roadmap demands. Custom inference chips offer cost and power advantages over general-purpose GPUs when optimized for specific model architectures, though they typically require multi-year development cycles and sacrifice the flexibility of Nvidia's CUDA ecosystem. Broadcom has become the go-to partner for hyperscalers building custom ASICs—the company has worked with Google on TPUs and Meta on training chips, giving it deep experience in the workloads that matter most to AI labs.

OpenAI has not disclosed a timeline for Jalapeño's production deployment, whether the chip targets GPT-series models specifically, or whether it will supplement Nvidia hardware or eventually replace it in certain workloads. The company also hasn't shared performance benchmarks or architectural details. What's clear is that the largest AI labs now see custom silicon as a strategic necessity, not just a cost optimization. Watch for OpenAI to detail Jalapeño's architecture and performance in the coming months—particularly how it handles long-context inference and whether it supports multimodal workloads. The real test will be whether custom chips can scale fast enough to match the pace of model releases, a challenge that has tripped up even well-funded hardware programs in the past.