NVIDIA Blackwell Reaches the Public Cloud: AWS EC2 P6-B200 Goes GA

Blackwell is the NVIDIA GPU generation that follows Hopper, and only the B200 and GB200 belong to it. The H200, which is often lumped in with them, is still Hopper: it is a memory upgrade to the previous generation, not a new architecture. There was never a single day on which Blackwell became widely available — it arrived cloud by cloud and region by region. The AWS milestone of 15 May 2025 is one of the clearest markers: the p6-b200.48xlarge instance became generally available with eight B200 GPUs. Access was narrow, though. At launch it existed in one region only, US West (Oregon), and it could be obtained solely through EC2 Capacity Blocks for ML — you reserve a block of GPUs for a defined window rather than starting one on demand like an ordinary server. The performance claim from AWS is measured: up to 2x the performance of P5en instances for training and inference. P5en is the H200 generation, so that comparison is against Hopper, not against the older H100. What changed in practice is that renting Blackwell no longer required a direct arrangement with NVIDIA. It still required booking capacity in advance in one corner of the world, so scarcity had moved rather than disappeared.