Anthropic is advancing the strategic plan of building its own server computing power, which is expected to significantly reduce its own computing power costs in the long run. Multiple industry insiders familiar with the plan revealed that the Claude large-scale model development company has signed more than ten letters of intent with a number of U.S. data center developers in recent months and plans to lease data center computer rooms. Relevant people familiar with the matter said that Anthropic management privately negotiated a cooperation plan: Google will provide financial guarantees for Anthropic’s rent payments, and Google will also participate in joint research and development of server chips that will be used in these computer rooms.
At this stage, Anthropic mainly obtains computing power through leasing chip servers from cloud vendors such as Amazon Cloud Technology and Google Cloud. However, the direct lease of data center projects marks a major shift in the company's strategy. Its competitor OpenAI has also recently promoted similar layouts.
OpenAI is negotiating a guarantee with Nvidia to provide cover for leasing a data center campus in Ohio with a total power consumption of up to 10 gigawatts. The project will saddle OpenAI with hundreds of billions of dollars in debt for hardware purchases and computer room rentals.
In comparison, the scale of Anthropic’s planned projects is currently smaller. People familiar with the matter said that the total amount of letters of intent signed recently exceeds 1 gigawatt of computing power.
According to the cooperation plan, if Anthropic is unable to pay the rent, Google will perform the contract on its behalf, thereby providing repayment guarantee for data center developers and their lenders. Anthropic's own balance sheet and credit rating alone are unable to meet the qualification requirements of financial institutions for lending money to large-scale data center projects.
Anthropic and OpenAI simultaneously planned to lease their own data centers, highlighting that chip manufacturers have an increasingly strong financial say in the wave of AI computing power infrastructure expansion. Nvidia, Google, and Broadcom, which jointly develops chips with Google, have participated in financing the purchase of computing hardware and provided guarantees for data center leasing. This closed-loop cooperation model has also caused industry risk controversy.
Chip manufacturers have their own competitive considerations in promoting such cooperation. Nvidia has long monopolized the AI server chip market and now faces a joint challenge from Google and Broadcom. The tensor processing unit (TPU) jointly developed by the two is the core computing power carrier of AI companies such as Anthropic. Anthropic, which currently ranks first in the industry in terms of revenue, has recently finalized a US$20 billion Broadcom TPU purchase order.
Broadcom also participated in the establishment of a new computing infrastructure fund funded by Apollo and Blackstone, which plans to provide financing for AI data centers with a total scale of more than 20 gigawatts, including some projects that support Broadcom's self-developed chips and serve Anthropic and OpenAI (OpenAI is teaming up with Broadcom to develop self-developed AI server chips). It is unclear whether this new fund will provide funding for the computer room project Anthropic is negotiating this time.
The letter of intent signed by Anthropic and the developer is not legally binding. During this period, the company can conduct due diligence on the partner and the land, and ultimately has the right to choose to shelve all or part of the leasing project.
However, the company has already prepared a large-scale computing power layout, and has formed a senior team including two former Google executives to advance lease negotiations, with the goal of locking in at least 10 gigawatts of self-owned computer room computing power in the next few years - a scale sufficient to meet all the electricity needs of ten cities such as San Francisco and Denver.
According to people familiar with the matter, this computer room leasing project is fully led by Tim Hughes. He previously served as chief development officer of Stack Infrastructure, a leading U.S. data center developer, and officially switched jobs this year to join Anthropic.
Leasing a data center usually requires the enterprise to purchase, operate and maintain all the server equipment in the computer room. In the long run, this will allow Anthropic to independently control the purchase price and supply channels of core hardware such as chips, server racks, and network equipment; at the same time, the company will also bear the huge financial pressure brought by high-priced hardware alone.
Both companies have released IPO listing plans, and the new direct rental debt of data centers will further increase the financial burden of OpenAI and Anthropic. In the next five to seven years, the two have committed a total of more than US$1 trillion in AI server rental expenditures to major cloud vendors. Anthropic also recently signed an agreement to rent SpaceX’s idle computing resources.
However, cloud vendor executives such as Microsoft CEO Satya Nadella have already predicted that leading AI companies will eventually build their own computing infrastructure.
Nadella said in an interview with "Strategic Insights" last week: "Objectively speaking, whether it is Anthropic or OpenAI, building self-built computing facilities is an inevitable choice in the long run, and the logic is very clear." (Additional note: OpenAI and Anthropic will not build their own computer rooms, but only adopt a leasing model)
Even if it starts leasing its own data centers, the computing power Anthropic rents from cloud vendors will still be significantly ahead of its own computer rooms in the short term. For example, the company has finalized a five-year computing power purchase agreement of approximately US$200 billion with Google, and the corresponding computing power far exceeds the direct lease of the computer room negotiated this time.
In recent months, various enterprises have accessed Anthropic's AI code tools and large models on a large scale, and the demand for computing power has skyrocketed, which has also forced enterprises to accelerate data center leasing cooperation.