AWS and NVIDIA plan 2 million-GPU AI expansion; Anthropic commits $100B+ for 5GW
AWS and NVIDIA unveiled a sweeping 2 million-GPU AI infrastructure expansion on August 27, deepening a partnership that spans silicon, models and data tooling. The plan integrates NVIDIA Nemotron open models through Amazon Bedrock and SageMaker, and brings NVIDIA's accelerated-data libraries to AWS services — cuDF paired with EMR for GPU-accelerated dataframes and cuVS with OpenSearch for vector search.
The capacity commitments underneath are staggering. Anthropic committed over $100 billion over 10 years for up to five gigawatts of AWS capacity, cementing AWS as a primary compute backbone for Claude. Separately, OpenAI expanded its AWS deal by roughly $100 billion over eight years for about two gigawatts of Trainium-based capacity — notable because it pulls OpenAI compute onto Amazon's custom silicon, not just NVIDIA GPUs.
The scale reflects the week's dominant macro theme: compute is the moat, and it's increasingly financed through mega-deals and vendor entanglement. It echoes NVIDIA's parallel $105B lease backing for an 8GW OpenAI Ohio data center and its $500B third-party financing platform with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR. Critics on HN and elsewhere warn about circular financing — 'Nvidia is effectively co-signing loans used to purchase its own silicon' — fueling AI-bubble concerns.
Strategically, the AWS-NVIDIA tie-up keeps Amazon central to frontier labs even as it pushes Trainium to reduce NVIDIA dependence — a hedge visible in the OpenAI Trainium allocation. Watch build-out timelines against power availability (gigawatt-scale data centers strain grids), whether Nemotron gains traction on Bedrock against Anthropic and Meta models, and how the compute-financing web holds up if demand or model economics shift.
