Anthropic Wants the Keys to Its Own Compute

SITUATION REPORT: Anthropic is moving to lock down more of the physical infrastructure behind Claude. Anthropic is in talks to cement control over more data centers, The Information reports. The report doesn’t make every detail of the talks public. The direction is still clear. One of the three leading frontier labs wants a firmer hold on the buildings, power and chips that run its models, and less exposure to capacity it only rents.

This is significant because in 2026, compute is the scarcest input in AI. Talent and ideas still matter. But the labs that can guarantee capacity years ahead are the ones that can train the next model on schedule and serve customers without rationing.

📍 How Anthropic Got Here

Until recently, Anthropic mostly got compute through cloud partners and didn’t own much of it:

1. Amazon. Its largest backer and main training partner. Anthropic trains on AWS Trainium chips through Project Rainier, a huge cluster Amazon built for it.
2. Google. A major investor and TPU supplier. In late 2025, Anthropic announced plans to use up to one million Google TPUs.
3. Microsoft and Nvidia. A late-2025 deal added Azure capacity and Nvidia hardware, with Anthropic committing to large Azure spending.
4. Its own buildout. Also in late 2025, Anthropic announced a $50 billion plan to build US data centers with Fluidstack, starting in Texas and New York.

That multi-cloud setup gave Anthropic flexibility and a lot of chips. It also left the company depending on partners who have their own priorities, their own customers and, in Google’s case, a competing model family.

⚙️ What “Control” Means Here

For readers outside infrastructure, owning compute isn’t the same as renting it. When an AI lab rents cloud capacity, the provider decides a lot: where the chips sit, how clusters are set up, when upgrades happen and who gets priority when supply runs short.

More control usually means some mix of:

1. Long-term leases or ownership stakes in specific facilities.
2. Dedicated capacity that can’t be handed to another tenant.
3. A say in design, including power, cooling and how big the clusters get.
4. Predictable costs over several years instead of market-rate pricing.

What stands out here is the timing. Anthropic’s revenue has climbed fast on enterprise demand and Claude Code. Rate limits and capacity crunches have been a recurring pain point for heavy users. Control over more data centers goes straight at that bottleneck.

🎯 The Competitive Picture

Anthropic isn’t alone. OpenAI has spent the past year and a half signing enormous infrastructure deals through Stargate and with chipmakers and cloud providers. Meta and xAI are building their own giant campuses. Every frontier lab is reaching the same conclusion: if you don’t control your compute, you don’t fully control your roadmap.

Anthropic has been more careful with money than some rivals. That’s part of its pitch to investors. So a push toward firmer control over data centers suggests leadership now sees the risk of being short on capacity as bigger than the risk of committing too much capital.

📌 Implications for Practitioners

1. Capacity could get better. If the talks turn into deals, expect fewer outages and looser rate limits over time, especially for API and enterprise customers.
2. Watch the partner dynamics. Amazon and Google are both investors and suppliers. Any move toward independence will be watched closely in Seattle and Mountain View.
3. Pricing stability. Owned or long-leased compute can lower costs per token over time, which could show up in API pricing.
4. Energy and siting news will follow. Data center deals now go hand in hand with power deals. Expect announcements about locations and power supply.
5. Don’t expect overnight change. Data center capacity takes months or years to come online. The benefits land in 2027 and beyond.

🔭 Outlook

The AI race is now as much about concrete, power lines and chips as it is about model architecture. Anthropic’s reported talks show it doesn’t plan to fall behind on that front. The terms, partners and scale of any final deal will show how far the company is willing to go to own its future. Full details are in the original report from The Information.