Amazon, Microsoft and Google hedge across the AI profit chain

The same company can win three ways

Amazon’s relationship with OpenAI and Anthropic is not just about investing in model makers. It also sells them computing capacity and helps distribute their models. Microsoft and Google occupy overlapping roles across the AI industry, too. That matters because the eventual profit pool may sit with the model makers, the infrastructure providers or the companies that control access to customers—and these partnerships give cloud platforms a stake in more than one outcome.

AI labs get capital, specialized computing capacity and access to cloud providers’ enterprise customers. The providers, meanwhile, have spent years building relationships with those customers. Their overlapping bets hedge against where the industry’s profits settle:

Capital comes back as compute revenue

In the AI infrastructure chain, equity investments and project financing can quickly turn into commitments to buy computing capacity. For frontier labs, compute is their largest expense, ahead of staffing, model training and market entry.

On April 20 this year, Amazon announced an additional $5 billion investment in Anthropic. At the same time, Anthropic committed to spending more than $100 billion on Amazon Web Services over ten years. It has comparable agreements with Microsoft and Google.

That capital gives a lab contracted access to computing capacity, not ownership of the infrastructure. A lab that builds its own data centers owns assets it can sell or pledge; Meta has taken that approach. A lab that commits to buying capacity instead has no comparable asset, but its payments continue for years.

For the provider, the return can come from equity as well as sales. Recently, Anthropic’s rising valuation brought Amazon more direct benefit than selling it compute. Of Amazon’s $62.6 billion in net income in the second quarter, $53.4 billion was non-cash profit, arising “primarily from our investment in Anthropic.” AWS, which sells computing capacity, reported $16.6 billion in operating income.

The timing of those returns is uneven: compute commitments stretch years into the future, while revenue contracts—from customer subscriptions to fees for AI agents—run for months.

A web, not a chain

These relationships do not operate in isolation. Map the commercial agreements across the industry and competitors appear on both sides of deals: supplying services to one another, distributing one another’s products and co-financing projects. Most companies in this matrix depend on at least four partners.

Anthropic’s compute strategy is a clear example. It runs Claude on Nvidia graphics processors, Google tensor processors and AWS Trainium chips. Using three infrastructure providers limits reliance on any one type of hardware and strengthens Anthropic’s position in future negotiations over capacity.

The pattern is to diversify upstream suppliers to protect access, then bundle options closer to the customer. That helps companies retain access to buyers—and influence over pricing.

The same push toward customers is visible beyond infrastructure. OpenAI has acquired companies across consumer hardware, developer tools, product analytics, medical records and media to build consumer products and applications around its models. Anthropic’s acquisitions have focused on developer tools and software infrastructure, including Bun and Stainless; in other enterprise sectors, it has expanded mainly through partnerships.

In 2026, both labs moved directly into business implementation:

Partnerships extend distribution; acquisitions determine what a model provider owns outright. Software and implementation services bring both labs closer to enterprise customers and can make their models harder to replace.

Where the bargaining power sits

Who earns the margin depends in part on how costly it is for a customer to switch suppliers. If switching is easy, competition pushes prices toward the cost of providing the service. If it is difficult, the supplier can keep more of the difference.

The current balance varies at each layer:

I think the more consequential contest is not simply over which model is best, but over who owns the customer relationship when models are close enough to substitute for one another. Cloud platforms can sell several models through one account; model providers can push back by embedding their systems in the work itself.

The strongest position belongs to companies that earn from more than one layer and are not tied to a single version of the market. A company can be an investor, supplier, sales channel and competitor to several rival model developers at once. That is less a bet on one winner than a way to collect value whichever layer captures the margin.