OpenAI's Jalapeño chip joins wave of custom AI accelerators challenging Nvidia's grip
OpenAI is building a custom inference chip with Broadcom, part of a broader industry shift as Google, Apple, and SpaceX design their own AI accelerators to reduce dependence on Nvidia's dominant GPU lineup.
OpenAI is developing a custom inference chip called Jalapeño in partnership with Broadcom, the company disclosed this week. The move places OpenAI alongside Google, Apple, and SpaceX in a wave of tech giants designing proprietary AI accelerators to cut reliance on Nvidia's dominant GPU lineup.
The chip is purpose-built for inference workloads — the production phase of AI where trained models generate outputs for users. While Nvidia has held roughly 80 percent of the AI accelerator market for years, companies running billion-dollar compute bills are now betting that custom silicon can deliver better performance-per-dollar for their specific workloads.
Key developments
- 01OpenAI's Jalapeño is an inference-only chip developed with Broadcom. No parameter counts, timeline, or fab partner have been disclosed. The project signals OpenAI's intent to control more of its inference stack as ChatGPT usage scales.
- 02Google's TPU v6 has been in production since 2023, handling both training and inference for Gemini. Google claims 4.7× better performance-per-watt than competing GPUs on its workloads.
- 03Apple's M-series Neural Engine powers on-device inference for Siri and image processing across iPhones, iPads, and Macs. The M4 chip, released in 2024, delivers 38 trillion operations per second for machine learning tasks.
- 04SpaceX is designing a custom chip for Starlink satellite networking and autonomous systems, according to job postings that surfaced in early 2025. Details remain sparse, but the company is hiring ASIC engineers with AI accelerator experience.
- 05
