Bedrock, Vertex, Foundry

Here's what happens in practice. A bank wants a chatbot that answers customers using Claude, the model made by Anthropic. Instead of signing a separate contract with Anthropic, the bank opens Amazon Bedrock, because it already keeps its databases and security systems there. It clicks Claude in the list of models, pays through the account it already has with Amazon - and the model starts running inside its own infrastructure, without its data traveling out to a third company.

The idea is simple, and it's business - not charity. Amazon, Google, and Microsoft run enormous cloud infrastructures, the servers where websites, databases, and entire companies live. Building an AI model at Claude's level is expensive and slow, so all three invited the best outside models onto their storefront and take a cut of every sale. Google does it with its own model, Gemini, too, but sells Claude alongside it - profiting from both sides of the counter.

It's used by companies that already live inside one of the three clouds and don't want a second door out. If your whole system is on Amazon, you go through Bedrock. If it's on Google Cloud - Google's service for storing and processing data - you open Vertex. If your company runs on Microsoft and Office, you use Foundry, which until recently was called Azure AI Studio. The gain is that security, access, and billing all stay inside a familiar system - trust doesn't get spread out to a third party.

The differences between the three aren't in the models themselves - Claude on Bedrock is the same Claude Anthropic offers directly. The difference is in the packaging: who controls access, where the data sits, which other cloud tools it automatically pairs with.

The storefront now sells the wrapper around the model too. On October 28, 2026 Microsoft released the Copilot harness in Foundry, in public preview: the service runs the agent loop and the sandbox the agent works in, and you provide knowledge, procedures and tools. The harness time is free until December 1, 2026. From then it is billed by consumption, and by Microsoft's figures the sandbox is 0.0994 dollars per vCPU-hour and 0.0118 per GiB-hour. Models and tools are paid separately. Which models are supported, the page does not say.

Google's storefront now sells OpenAI models too. Since October 21, 2026 GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are in Google Cloud's "model garden", for now in a private preview for select companies. They are bought through Google Cloud Marketplace and show up on one invoice: they draw down the cloud commitments the customer has already made. Google lists zero data retention and confidential computing as coming soon. Google itself claims to be the only major cloud provider with access to all three leading closed models.

You're not buying a different product from three different storefronts - you're buying the same product, packaged by three different sellers, each with an interest in you never leaving their shop.

I look at how it's put together

The three cloud giants aren't competing over who builds the better model - they leave that to Anthropic, to OpenAI (the maker of ChatGPT), and to Google DeepMind (Google's AI lab). Here they're competing over who gets to be the storefront the money passes through. So when you see news that Claude 'is coming to Azure' or 'is now on Google Cloud', don't read it as AI news. Read it as distribution news - who controls the door companies walk through to reach the most in-demand model right now.

The visual is generated code-art, no third-party images.