A Graphics Processing Unit is a specialized processor originally developed for graphics rendering, now repurposed as the primary compute substrate for neural network training and inference. GPUs are optimized for massively parallel operations — executing the same instruction across thousands of data elements simultaneously — which maps cleanly onto the matrix multiplication that dominates deep learning.
The GPU market for AI is almost entirely dominated by NVIDIA, whose data-center GPUs (A100, H100, H200, B200) train and serve most of the world's frontier AI models. AMD has competitive hardware in some configurations but lacks the software ecosystem (CUDA) that makes NVIDIA the path-of-least-resistance choice. Custom ASICs (TPUs at Google, Trainium at AWS, various custom silicon at Anthropic and OpenAI) are gaining share at frontier scale but remain a minority of deployed compute.
For most enterprise buyers, GPU capacity is now a critical supply- chain constraint. Access to H100s and H200s has been a gating factor for AI deployment plans since 2023, and the situation remains tight in 2026 despite substantial capacity expansion.