The $4 Billion Question Nobody’s Actually Asking
When Amazon announced it had completed its $4 billion investment in Anthropic by early 2025, the headlines mostly focused on the size of the check. Biggest corporate AI investment in AWS history. Validation of Anthropic’s technology. Another data point in the ongoing arms race between hyperscalers. All of this is true, but it misses the actual story.
The real story is structural. Amazon didn’t just buy a stake in a promising AI company. It bought exclusivity. Anthropic agreed to use AWS as its primary cloud provider for training, inference, and commercial API services. That’s the buried lede, and it matters more than the dollar amount because it represents a tectonic shift in how the AI infrastructure market actually works.
I spent years at McKinsey analyzing consolidation in enterprise software. I watched the database market, the middleware stack, the analytics layer. Every time, the playbook was the same: whoever controlled the foundational infrastructure layer controlled the margin, the switching costs, and ultimately the negotiating power with customers. This deal proves that playbook is alive and well in AI, only accelerated.
The Ecosystem Lock That Nobody’s Being Honest About
Let’s be direct: the idea that AI startups can remain agnostic about their infrastructure layer is now officially dead. Microsoft tied OpenAI to Azure. Google committed $2 billion to Anthropic while keeping it on Google Cloud. And now Amazon has done the same. The three major foundation model companies are no longer neutrally positioned technologies that any developer can pick up and run anywhere. They are distribution channels for specific cloud platforms, full stop.
This creates what I’d call a distribution moat. When a startup wants to build on Claude, they’re increasingly incentivized to do so via AWS Bedrock because that’s where Anthropic runs, that’s where the tooling is deepest, and that’s likely where Anthropic’s own engineering team is spending most of its optimization effort. Same dynamic applies to GPT on Azure and Gemini on Google Cloud. The model providers have become channel partners for the clouds, whether anyone wants to say it out loud or not.
Andreessen Horowitz released a survey in 2025 that captured the cognitive dissonance perfectly. Seventy-one percent of AI startups listed cloud vendor lock-in as a top-three concern. Yet sixty-eight percent were already running their primary workloads on a single cloud provider. That’s not incompetence on the founders’ part. It’s rational decision-making in an ecosystem where the model providers have already made the lock-in decision for you.
Why This Matters More Than the Margin Story
The conventional analysis here is about Amazon’s unit economics. AWS can now offer Claude through Bedrock, capture higher margins, and deepen customer relationships. All true. But there’s a second-order effect that matters more: the consolidation of the infrastructure layer is accelerating, and the threshold to escape it is becoming prohibitively high.
According to Synergy Research Group Cloud Market Share 2025, AWS, Azure, and Google Cloud together controlled sixty-six percent of global cloud infrastructure spend in 2025. That’s up from roughly fifty-five percent just five years ago. We’re in the middle of a decade-long consolidation, and now the foundation model providers are actively participating in that consolidation rather than disrupting it. They could have remained neutral, built multicloud infrastructure, forced the cloud providers to compete for their workloads. Instead, they took the capital and accepted the constraints.
For a founder, this means something concrete: the decision about which foundation model to build on is now partially a decision about which cloud to build on. It shouldn’t be that way, but it is. And the economic gravity is going to make it worse, not better, over the next two years.
The Unremarkable Growth Number That Should Alarm You
Amazon disclosed at re:Invent 2025 that startups building on Claude through AWS Bedrock grew to over 50,000 active developers, up 3x from the prior year. Read that number carefully. Fifty thousand developers. That sounds large until you compare it to the number of developers building on OpenAI (which has crossed into the millions) or the number of developers who might theoretically want to build on Claude but can’t easily do so outside of AWS.
The growth rate is real. The lock-in is also real. Together they tell a story about how power in the AI infrastructure market is consolidating faster than most people think. AWS re:Invent 2025 — Anthropic and Bedrock Announcement framed this as a win for developers and startups. In some narrow sense it is. You now have easier access to Claude on AWS. But you also now have less leverage to negotiate terms if you want to use Claude on Azure or Google Cloud or your own infrastructure.
What This Means for Your Decision Tree
If you’re a founder evaluating which foundation model to build on, you’re not just evaluating model quality or price anymore. You’re making a structural decision about your long-term cloud footprint. That’s fine if you’ve already decided on AWS as your strategic cloud provider. It’s strategically expensive if you haven’t, because now you’re taking on distribution risk that didn’t exist two years ago.
The harder truth is that Anthropic, OpenAI, and Google didn’t have to structure themselves this way. They could have demanded multicloud neutrality as part of any capital raise from hyperscalers. They could have built their own infrastructure or partnered in looser configurations. Instead, they chose the concentrated capital route. It’s easier in the short term, and it’s almost certainly correct from a pure business perspective. But it closes off options, and in an industry moving as fast as AI, closed options become expensive liabilities faster than anyone expects.
The question I’d ask myself as a founder is not whether Claude is better than GPT or Gemini. The question is whether you’ve thought through the full cost of being structurally dependent on a specific cloud provider for your core AI inference layer. If you have and you’ve decided it’s worth it, that’s a legitimate strategic choice. If you haven’t, you should. Because the decision is being made whether you’re thinking about it or not.