AMD MI300X and Intel Gaudi 3 Challenge NVIDIA GPU Dominance for AI

AMD and Intel are making significant inroads against NVIDIA's near-monopoly on AI training and inference hardware. AMD's MI300X GPU (192GB HBM3 memory, 5.2 TB/s bandwidth) is being adopted by Microsoft Azure, Meta, and Oracle Cloud as an alternative for large language model inference. Intel's Gaudi 3 offers competitive training performance at 30% lower cost per FLOP for certain workloads. Meanwhile, domestic hyperscalers are accelerating development of custom AI chips: Google TPU v5, Amazon Trainium 2, Microsoft Maia 2, and Meta MTIA 2 are all ramping production. The diversification is reducing cloud customers' dependence on NVIDIA H100 availability.