The category covers several distinct hardware approaches competing on different tradeoffs: Nvidia's GPUs (Blackwell, Hopper) dominate general-purpose training; custom accelerators like Google's TPUs or Amazon's Trainium optimize a cloud provider's own cost structure; and wafer-scale chips like Cerebras's trade flexibility for raw inference throughput, powering high-speed tiers like OpenAI's Ultrafast mode for GPT-5.6 Sol. Export controls on advanced AI chips have become a direct lever of AI policy, as seen in restrictions affecting Chinese access to frontier-class silicon.