AMD and OpenAI announced a sweeping chip supply agreement that positions AMD’s Instinct MI325X accelerators as the backbone of new inference clusters rolling out across OpenAI’s global footprint. The deal includes priority access to AMD’s 3D-packaged accelerators, wafer capacity commitments through 2027, and a joint engineering lab in Austin focused on model optimization.
OpenAI COO Brad Lightcap said the collaboration enables the company to diversify beyond Nvidia while maintaining performance targets for GPT-5.1 and upcoming multimodal agents. AMD will dedicate a portion of its TSMC N3 capacity to the program and provide OpenAI with firmware hooks for real-time telemetry and power management.
Joint Roadmap Targets Faster Inference
The companies are co-developing a specialized runtime that layers OpenAI’s Triton compiler atop AMD’s ROCm stack. AMD CTO Mark Papermaster described it as a "north star" project to close the tooling gap with CUDA. Early testing reportedly shows 20 percent faster batched inference on GPT-4.5 compared with previous Instinct deployments.
Beyond software, AMD will help OpenAI stand up modular data centers built around liquid-cooled rack sleds that support both MI325X and forthcoming MI350 accelerators. The designs aim to slash deployment times to under eight weeks.
Regulatory Review on the Horizon
The contract still faces regulatory scrutiny in the United States and European Union, where watchdogs are studying the AI supply chain for signs of concentrated risk. AMD and OpenAI said they are prepared to make transparency commitments, including regular disclosures about chip allocation and energy consumption.
Financial terms were not disclosed, but analysts at Evercore ISI estimate the deal could be worth more than $8 billion over four years if OpenAI exercises all volume options. The companies plan to provide further updates at AMD’s Financial Analyst Day in March.