OpenAI is teaming up with Broadcom to develop custom AI accelerators, set for mass production in 2026. These “XPUs” will power OpenAI’s internal models, like GPT-5, instead of being sold to others. This move aims to reduce dependence on Nvidia’s GPUs.
The announcement aligns with Broadcom’s reveal of a $10 billion AI infrastructure order from an unnamed customer. Sources like the Financial Times and Wall Street Journal identify OpenAI as the buyer. Broadcom’s stock surged after the news, showing investor faith in its custom silicon growth.
Chip Specs and Timing
OpenAI’s first chip uses TSMC’s 3nm process for better performance and efficiency. It features systolic arrays and HBM memory, much like top AI accelerators.
This strategy mirrors efforts by Google, Amazon, and Meta, who build custom silicon for their AI needs. Broadcom has helped with Google’s TPU and Meta’s MTIA, as noted in The Register.
Benefits for OpenAI
Custom AI chips could lower costs for running models, improve supply chain control, and boost performance for GPT-5. OpenAI has said it needs more compute power for future systems. A dedicated accelerator helps meet that demand.
The $10B deal fits OpenAI‘s push to scale beyond limited Nvidia GPUs. This order ramps up Broadcom’s 2026 AI revenue and solidifies it as a top partner for custom AI chips. The stock rally followed the earnings report.
Broadcom also provides AI networking parts, key for large models. Co-designing compute and networking gives it an edge, as seen in its Google work.
Wider AI Chip Landscape
OpenAI’s in-house silicon could make AI systems more diverse, cutting Nvidia’s market share and influencing prices. Yet, experts say AI demand supports multiple players, even with more custom chips.
This OpenAI-Broadcom partnership marks a shift to integrated AI infrastructure. It ensures stable compute for GPT-5-level models and positions Broadcom as a leading design partner.
