What AWS is — and why AI runs on it
Amazon Web Services is Amazon's cloud computing division: it rents out computing power, storage, and software services over the internet so that companies don't have to build and maintain their own data centers. For most of its history, AWS was the backbone of the internet economy — powering everything from Netflix to startups. In the current AI era, it has taken on a new role: the physical infrastructure that trains and runs the world's most powerful AI models.
Think of AWS as the electrical grid for AI. Just as factories don't generate their own electricity, AI labs don't (mostly) build their own data centers from scratch. They plug into AWS — and increasingly, AWS is building custom hardware specifically for them.
The Anthropic relationship: AWS as primary AI partner
The most significant AI relationship AWS has built is with Anthropic, the company behind the Claude family of models. What started as a $4 billion investment in late 2024 has grown into one of the largest technology commitments in history: Anthropic has agreed to spend over $100 billion on AWS over ten years, securing up to 5 gigawatts of computing capacity on Amazon's custom Trainium chips (Trainium 2 through 4). Amazon has invested a total of up to $13 billion in Anthropic, with up to $20 billion more possible.
This isn't just a purchasing agreement — it's a deep technical partnership. Anthropic engineers write low-level software for AWS's Trainium chips, helping optimize model training from the silicon up. In return, Claude is available to tens of thousands of enterprises through Amazon Bedrock, AWS's managed AI platform, with named deployments at companies like Pfizer, Intuit, and Perplexity, as well as the European Parliament.
OpenAI also lands on AWS
AWS isn't exclusive to Anthropic. OpenAI — Anthropic's main rival — signed a $38 billion multi-year deal with AWS in late 2025, and the two companies built a "stateful runtime environment" for AI agents inside Amazon Bedrock. This means OpenAI's models can run persistent, multi-step tasks (like an AI assistant that remembers what it was doing across a long project) directly within AWS's managed infrastructure. OpenAI's Frontier models and Codex are now generally available on AWS, giving enterprise customers access to both major AI families through the same cloud platform.
Amazon Bedrock: the AI marketplace
Amazon Bedrock is the managed layer where all of this comes together for business customers. Instead of figuring out how to run AI models on raw servers, enterprises can access Claude, OpenAI models, and Hugging Face's open-source models through Bedrock — with AWS handling the security, compliance, and billing. Salesforce, for example, integrated Claude into its Einstein CRM platform via Bedrock, and Hugging Face models are available in the Bedrock marketplace for customers who want open-weight alternatives.
Government and defense: a growing frontier
AWS has become the cloud of choice for sensitive government AI deployments. Claude models are approved for FedRAMP High and Department of Defense Impact Level 4 and 5 workloads through AWS GovCloud — the security-hardened version of AWS for classified and controlled information. Anthropic's $200 million Department of Defense agreement runs on AWS-hosted infrastructure, and Claude was made available to the U.S. Intelligence Community through AWS GovCloud as early as 2024.
The open-source ecosystem: Hugging Face
Beyond the frontier model labs, AWS has built a deep relationship with Hugging Face, the platform that hosts most of the world's open-source AI models. Hugging Face models can be deployed with one click into Amazon SageMaker Studio, run on AWS's custom Inferentia2 inference chips, and accessed through the Bedrock marketplace. This means AWS serves both ends of the AI market: the closed, frontier models from Anthropic and OpenAI, and the open-source community that builds on top of freely available weights.
A new kind of risk: physical infrastructure under attack
AWS data centers have historically been treated as background infrastructure — reliable, invisible, and safe. That changed in early 2026, when Iranian drone strikes damaged at least three AWS facilities in Bahrain and the UAE, disrupting cloud services across the region. The attacks coincided with revelations that AI models running on AWS infrastructure were being used in U.S. military targeting operations. The episode was the first known targeting of commercial cloud infrastructure during active conflict, and it raised new questions about the risks of concentrating critical AI workloads in a small number of physical locations.
Where things are heading
AWS is positioning itself as the neutral ground where competing AI ecosystems — Anthropic, OpenAI, Hugging Face, and others — can all reach enterprise customers. Its custom silicon (Trainium for training, Inferentia for inference) gives it a hardware angle that rivals like Google Cloud and Microsoft Azure are also pursuing. The scale of the compute commitments involved — measured in gigawatts and hundreds of billions of dollars — suggests that whoever controls the physical infrastructure of AI training will have significant influence over which models get built and how fast they improve.
