Anthropic in Early Talks to Directly Lease Up to 1 GW Data Center from Stream Data Centers, Majority Owned by Apollo Global Management
Anthropic is in early discussions with Stream Data Centers, majority-owned by Apollo Global Management, to directly lease up to approximately 1 GW of computing capacity and deploy tensor processing units designed by Broadcom and Google to reduce reliance on traditional cloud billing and unit costs. Google may also provide credit guarantees for some projects. The details have not been publicly confirmed by either party, and discussions are still in the early stages.
This aligns with previously disclosed pathways. The company has signed over a dozen letters of intent for U.S. data centers, totaling more than 1 GW of capacity, which are non-binding documents for due diligence. Apollo and Blackstone previously financed approximately 1 GW of tensor processing units valued at around $35 billion through special purpose vehicles, with hardware entering a leasing structure via Google and Broadcom, with initial deployments at sites like Fluidstack. Google has provided guarantees for rents at multiple data centers, with public statements indicating potential exposures in the billions of dollars.
The computing puzzle also includes multi-GW next-generation tensor processing units from Google and Broadcom, to be delivered starting in 2027; Amazon's Trainium with a cap of about 5 GW; and AMD's accelerator cards at about 2 GW. The goal is to increase online capacity from approximately 1.4 GW by the end of 2025 to around 9.5 GW by the end of 2027. On the equity side, Google has committed to additional investments, with the company valued at approximately $965 billion after a round in May 2026, and has secretly submitted IPO documents. Annualized revenue has been reported at around $30 billion, with over a thousand enterprise clients in the million-dollar range.
Directly renting power and managing cabinets means that model companies transition from being cloud tenants to co-signers of power and chips. Cloud providers will still offer some capacity, but pricing power shifts from hourly instances to decade-long leases and guarantees. Power queuing and substations, rather than software interfaces, become the next bottleneck.
The market mechanism rewrites cloud gross margins into infrastructure spreads. Buyers are model labs looking to lock in machines and reduce unit inference costs, while sellers are data center platforms holding land, power, and private loans. The event-driven aspect comes from parallel discussions on direct leasing and guarantee structures. Beneficiaries include Google, Broadcom, and Apollo, who can package chip residuals and rental guarantees, while pressured parties are general clouds that can only sell accelerator card hours at catalog prices. Funding shifts from operational cloud fees to debt and decade-long leases of special purpose vehicles.
Source: Public Information
ABAB AI Insight
Anthropic has previously placed training and inference on Amazon and Google Cloud, exchanging equity, capacity, and IPO narratives. Directly leasing 1 GW acknowledges that cloud markups have become a major item on the cost curve. Assets like Stream, owned by Apollo, provide collateralizable cabinets and power, not another set of managed consoles. Google sells chips and may also guarantee rents, effectively using its balance sheet to back its competitors' physical layers, solely to spread tensor processing units into more data centers.
The capital path is "chip vendors commit to purchase, asset management issues debt, cloud vendors guarantee, labs pay rent." Spending $35 billion to buy about 1 million processors and then lease them back shifts capital expenditures from model companies' annual reports to special purpose vehicles. Writing leases and guarantees into disclosures before going public is intended to show investors capacity rather than just the risks of renewing cloud contracts. The flow of money is towards power and custom silicon, not new chat windows.
This is analogous to Netflix building its own content library, retailers building their own logistics, and airlines using leasing companies instead of just buying hours from manufacturers. The AI industry is transitioning from renting cloud services to controlling power: software differences still exist, but the physical layer is starting to determine who can lower unit costs.
The structural change is a reconstruction of the supply chain. The cloud has shifted from being the only window to an optional channel, with labs directly signing contracts with power developers. The mechanism is: model revenues can now support GW-level commitments, but cloud gross margins eat into expansion speed; whoever can separate and recombine chips, rent, and guarantees can bypass catalog pricing. Control is not expressed as owning every building, but as who’s name is on the lease and who’s rating is on the guarantee letter.
ABAB News · Cognitive Laws
When cloud bills are expensive enough, renting power becomes more cost-effective than renting instances.
Guarantee letters can determine who dares to hand over cabinets more than equity.
The control that model companies want is ultimately written on the power station queuing list.