The AI chip market has become a three-way race as AMD and Google close in on NVIDIAs dominance. Here is how the latest AI accelerators compare for training and inference workloads.

NVIDIA H200

The reigning champion with 141GB HBM3e memory and 4.8 TB/s bandwidth. Dominates training workloads with the CUDA ecosystem advantage. Price: ~$30,000 per chip (when available).

AMD MI350

AMDs latest challenger offers 192GB HBM3e — more memory than H200 — at a significantly lower price point. ROCm software stack has improved dramatically, with PyTorch support reaching near-parity with CUDA.

Google TPU v6 (Trillium)

Available exclusively on Google Cloud, TPU v6 offers exceptional inference performance at lower cost-per-token than GPU alternatives. Best for serving large language models at scale.

Market Impact