NVIDIA, Microsoft, AWS split AI infrastructure
- Amazon and Anthropic expanded their alliance in April, Microsoft rewired its OpenAI deal days later, and AWS kept adding NVIDIA capacity — clarifying the AI stack. - The biggest number is Anthropic’s more-than-$100 billion AWS commitment for up to 5 gigawatts, alongside OpenAI’s 2-gigawatt AWS deal and Azure changes. - AI infrastructure is no longer one-company territory — chips, clouds, and model access are splitting into separate profit pools.
AI infrastructure is settling into layers. NVIDIA supplies the scarce chips. Cloud platforms package those chips into usable capacity. Model companies decide where to train, where to serve, and who gets enterprise distribution. The new thing is that the boundaries are getting clearer — and the money is getting split more explicitly across Amazon, Microsoft, OpenAI, Anthropic, Google, and NVIDIA. ### What changed this spring? April brought two moves that made the structure visible. On April 20, Anthropic and Amazon expanded their partnership: Anthropic said it would spend more than $100 billion over 10 years on AWS technologies, secure up to 5 gigawatts of capacity, and keep AWS as its primary training and cloud provider for mission-critical workloads. A week later, on April 27, Microsoft and OpenAI amended their partnership so Microsoft stayed OpenAI’s primary cloud partner, but OpenAI gained freedom to serve products across any cloud provider. (anthropic.com) ### Why does that matter? Because the old picture was too simple. People talked as if one model company would pair with one cloud and one chip vendor. But the actual market now looks more like a mesh. Anthropic runs broadly across AWS, Google Cloud, and Azure. OpenAI is still deeply tied to Microsoft, but now also has a major AWS distribution and infrastructure deal. That means enterprise customers are buying access paths, governance, and capacity — not just a single model. (anthropic.com) ### Where does NVIDIA fit now? Basically at the bottom of almost all of it. Even when clouds push custom silicon like AWS Trainium, NVIDIA still anchors a huge share of frontier and enterprise AI capacity. AWS said in March it would add more than 1 million NVIDIA GPUs across regions starting in 2026, including Blackwell and Rubin systems. Microsoft’s Azure expansion for Anthropic was also explicitly framed around NVIDIA-powered infrastructure. So even when clouds compete, NVIDIA often still gets paid first. (anthropic.com) ### Is AWS just a landlord here? Not anymore. AWS is trying to move up the stack. Anthropic’s deal is not just raw servers — it includes Trainium2 through Trainium4, Graviton, Bedrock access, regional inference expansion, and direct Claude Platform availability inside AWS accounts. OpenAI’s February partnership with Amazon pushed the same direction: AWS becomes the place where enterprises run OpenAI-powered “stateful” agents and access OpenAI Frontier through Bedrock. (aws.amazon.com) That is cloud bundling, not commodity hosting. ### What is Microsoft protecting? Distribution and enterprise control. Microsoft gave OpenAI more freedom, but kept Azure first position unless it cannot support the required capabilities, preserved IP licensing through 2032, and kept itself wired into OpenAI’s growth. That looks less like surrender and more like a reset: Microsoft wants to be the default enterprise wrapper around frontier AI, even if the models themselves travel more freely across clouds. (anthropic.com) ### Why is Anthropic the interesting swing player? Because Anthropic is proving that a frontier lab can be multi-cloud without becoming cloud-neutral. It chose AWS as its primary provider, committed huge spend there, kept Claude available on Bedrock, Vertex AI, and Azure Foundry, and had already announced a big Azure-NVIDIA partnership last year. That makes Anthropic both a model company and a bargaining chip in the cloud wars. (blogs.microsoft.com) ### So who captures the margin? Three groups, not one. NVIDIA captures value when demand forces everyone to buy accelerated compute. Clouds capture value when they turn that compute into managed enterprise platforms with security, billing, and compliance. Model companies capture value when customers insist on a specific model family or workflow layer. The catch is that each group is trying to creep into the others’ turf. (anthropic.com) ### Bottom line? The AI stack is not consolidating into a single winner. It is unbundling into chips, clouds, and models — with each layer trying to own the customer relationship before the others do. (anthropic.com) (aboutamazon.com)
