Anthropic Is Paying Its Rivals Billions for Compute — and Rewriting the Rules of AI Infrastructure

Anthropic is negotiating to pay Meta $10 billion for compute capacity. It already pays xAI — Elon Musk’s AI company — $1.25 billion per month for access to the Colossus 1 data center in Memphis. It has a 20-year, $19 billion lease with infrastructure firm TeraWulf. And it committed $50 billion to custom-built facilities with Fluidstack last November.

All of this is happening while Anthropic still runs substantial workloads on Amazon Web Services and Google Cloud, its original backers and cloud partners. The company is not leaving the public cloud. It is layering an entirely new category of infrastructure spending on top of it — and a growing share of that spending is going to direct competitors.

Meta, which is weighing the $10 billion proposal, develops Llama, a direct rival to Anthropic’s Claude. xAI builds Grok. Yet both are now potential infrastructure suppliers to the company trying to beat them. The old rulebook — where cloud providers supplied compute and AI labs consumed it — is being replaced by something far messier.

The sheer scale of these commitments tells its own story. Anthropic’s total infrastructure obligations, if all four deals materialize as planned, approach $100 billion across multiple decades, multiple geographies, and multiple counterparties — some of whom are simultaneously trying to beat Anthropic in the market. That’s not normal corporate procurement. That’s a company betting its entire future on the assumption that whoever controls the compute controls the outcome.

The deals, broken down

The numbers here are worth sitting with because they signal a structural shift, not just a big round of procurement.

Anthropic’s agreement with xAI, disclosed in a SpaceX offering filing from May 2026, covers the Colossus 1 facility in Memphis through May 2029. Either party can terminate with 90 days’ notice — a clause that cuts both ways. xAI can walk away if it needs the capacity for its own models, and Anthropic can walk away if it finds a better deal or if the relationship sours. That kind of mutual optionality is unusual in infrastructure contracts and reflects how fluid the compute market has become.

The TeraWulf lease is more conventional in structure but staggering in scale: 401 megawatts at a campus in Hawesville, Kentucky, with initial capacity expected in the second half of 2027 and projected contracted revenue of $19 billion over 20 years. TeraWulf told investors it expects the deal to be supported by investment-grade credit, which means banks are betting on Anthropic’s ability to pay — and on the underlying assumption that demand for frontier AI compute will keep growing for decades.

The Fluidstack deal, at $50 billion, is the largest of the four and the least detailed in public filings. Anthropic described it as custom-built infrastructure in Texas and New York designed for its own workloads. That language — “custom-built” and “for its own workloads” — suggests facilities purpose-designed for training and inference at Anthropic’s specific scale, not generic data center capacity that could be repurposed.

And then there’s Meta: an offer to buy up to $10 billion in computing capacity over two years from the company behind Llama. As of July 17, Meta was still weighing the proposal. If it goes through, it will be the most explicit example yet of a frontier AI lab paying a direct competitor to supply the infrastructure it needs to compete.

Why this is happening now

The simplest explanation is that there isn’t enough compute. Every frontier lab — Anthropic, OpenAI, Google DeepMind, xAI, Meta — is racing to train the next generation of models, and the models keep getting bigger. The public cloud providers, for all their scale, can’t build data centers fast enough to keep up. Power constraints, chip shortages, and regulatory bottlenecks mean that even the most well-funded labs face a physical ceiling on how much compute they can access through conventional channels.

But the deeper story is about risk allocation. Building a modern AI data center involves three things that tech companies are historically bad at: construction, financing, and permitting. A 400-megawatt campus in Kentucky doesn’t just need chips and servers. It needs power purchase agreements with utilities, environmental impact assessments, local government approvals, and construction crews who can pour concrete on schedule. These are not software problems. They are industrial problems, and they move at industrial speed — which is to say, slowly.

By spreading its infrastructure bets across Meta, xAI, TeraWulf, Fluidstack, AWS, and Google Cloud, Anthropic is reallocating risk. If one partner hits a permitting snag — OpenAI’s Project Jupiter with Oracle is reportedly facing exactly this — Anthropic still has five others. If construction runs late at one site, capacity comes online at another. The diversification that makes sense for an investment portfolio turns out to make sense for compute infrastructure too.

The downside is complexity. Managing capacity across six different providers, each with different contract terms, different APIs, different geographic locations, and different political relationships, is an operational headache. And the contracts themselves create dependencies that are hard to unwind. Anthropic’s Claude cannot easily switch from Colossus 1 to a TeraWulf facility mid-training-run — the data has to be moved, the software stack has to be reconfigured, and the training job might need to restart from a checkpoint. The infrastructure choices it makes today will constrain its model development for years, regardless of how the technology or the competitive landscape evolves.

The competitor-as-supplier paradox

The Meta proposal is the most revealing of the four deals because it challenges the basic assumption that AI competition means companies don’t do business with each other. In every other industry, competitors routinely supply each other. Samsung manufactures iPhone screens. Amazon hosts Netflix’s infrastructure. Boeing and Airbus share suppliers. But the AI industry has largely treated competition as a zero-sum contest — your model versus mine, your API versus mine, your users versus mine.

The Meta-Anthropic conversation suggests that’s changing. Meta has invested tens of billions in AI infrastructure, much of it built for Llama training and inference. If that infrastructure has spare capacity — and given the scale of Meta’s buildout, it almost certainly does — selling access to a competitor turns a cost center into a revenue stream. For Anthropic, buying from Meta is faster than building from scratch. The logic is sound even if the optics are strange.

The xAI deal follows the same pattern but is already live. Elon Musk’s company is collecting $1.25 billion per month from the company building Claude while simultaneously developing Grok. The irony is obvious. The business logic is harder to argue with: xAI has capacity, Anthropic needs it, and the deal makes money for both sides.

Whether this becomes the norm depends on two things. First, whether the compute shortage persists. If supply catches up with demand — and the billions pouring into data center construction suggest it eventually will — the incentive to buy from competitors diminishes. Second, whether the relationships stay transactional or become strategic leverage points. A 90-day termination clause means xAI can pull the plug on Anthropic with three months’ notice. That’s not enough time to migrate a major training operation to a new facility. The contracts create mutual dependency, but the dependency isn’t symmetrical — and whoever holds the capacity holds the leverage.

What it means for the rest of the industry

Anthropic is the most aggressive diversifier right now, but it won’t be alone for long. OpenAI has its own infrastructure challenges (the Oracle permitting issues are public) and its own relationship with Microsoft, which is also a competitor of sorts given Microsoft’s investments in its own models. Google DeepMind sits inside a company that builds its own TPUs and data centers, giving it a structural advantage that the independent labs don’t share.

The trend points toward a future where the winners in AI are not necessarily the companies with the best models, but the companies that are best at securing compute. Model architecture, training data, and algorithmic innovation still matter — but they matter less if you can’t get enough GPUs to train at frontier scale. The competition is shifting from “who can build the smartest model” to “who can build the most reliable supply chain for intelligence” — and the winners are already placing their bets.

That’s a fundamentally different kind of race, and it plays to the strengths of companies that are good at infrastructure, not companies that are good at research. OpenAI, for all its research prowess, is reportedly dealing with Oracle permitting delays. Google DeepMind benefits from Google’s own TPU supply chain. Anthropic’s dealmaking suggests it understands the new rules — and it’s paying competitors to buy itself time. Whether its competitors — and its suppliers — understand them too will determine the shape of the AI industry for the next five years. The model builders who figure out infrastructure first will have more runway than the ones who figure out architecture first — and right now, Anthropic looks like it’s betting the company on that exact premise.