Key points
▶ Anthropic is securing data centers and computing resources across the United States to meet rising demand for its AI services.
▶ According to an amended securities filing submitted by Anthropic on November 2, its long-term commitments for computing and infrastructure total $895.4 billion (about 1,240 trillion won). Of that amount, $447.5 billion is scheduled to be spent by the end of 2030.
▶ Anthropic is directly purchasing chips and cooling equipment and is also involved in designing data centers.
Anthropic, the artificial intelligence (AI) company, is moving to secure data centers across the United States to cope with surging demand for computing power. It is expanding beyond its traditional reliance on cloud providers to work with former cryptocurrency mining companies and data center specialists. The company is also considering directly purchasing semiconductors and cooling equipment, and converting shuttered power plants and factories into data centers.
The New York Times (NYT) reported on November 2 (local time) that Anthropic has signed dozens of agreements with companies including Amazon, Google, Microsoft and SpaceX since facing a computing shortage late last year. The company is contacting developers across the United States to secure data center sites, power and semiconductors, and is directly involved in building facilities.
Its investment commitments have also grown substantially. An amended securities filing Anthropic submitted to the U.S. Securities and Exchange Commission (SEC) the same day shows that its total commitments to purchase or lease computing and infrastructure amount to $895.4 billion (about 1,240 trillion won). Most are long-term contracts needed to develop AI models and operate services, with some extending through 2045.
Claude Code’s popularity surges
The race to secure computing capacity gathered pace late last year. Use of the AI coding tool Claude Code surged, but supplies of the servers and semiconductors needed to handle the demand could not keep up. Anthropic had to restrict access to some free services and adjust product prices.
Anthropic CEO Dario Amodei said at the NYT DealBook Summit in December last year that it was difficult to predict how much computing capacity the company would need in the future. The uncertainty around AI demand was so great, he explained, that it was difficult to avoid mistakes when deciding how much to invest.
Tom Brown, Anthropic’s chief computing officer, also identified computing capacity as a major bottleneck to AI development at a related Group of 20 (G20) event in September. Demand for AI services is growing quickly, he said, but there is a shortage of power and construction workers to support it.
Founded in 2021, Anthropic has long depended heavily on Amazon’s and Google’s cloud infrastructure. The two companies invested a combined $6 billion (about 8.3 trillion won) in Anthropic in 2023, and Anthropic put a substantial portion of the funds toward cloud services.
Brown, who has led the company’s contracting efforts, has also diversified its semiconductor supply chain. Anthropic uses not only Nvidia graphics processing units (GPUs), but also Amazon’s Trainium, Google’s tensor processing units (TPUs) and AI chips from AMD. The strategy is intended to reduce the risk of service operations being dependent on any one supplier’s chip availability.
Beyond Amazon and Google, Anthropic turns to SpaceX and mining companies
Anthropic has rapidly expanded its computing supply agreements since late last year. According to its amended securities filing, it has committed to spend at least $121.3 billion (about 168 trillion won) with Google through 2033 and at least $110 billion (about 152 trillion won) with Amazon through 2036. Its agreement with Microsoft is worth at least $44.9 billion (about 62 trillion won) through 2033.
Anthropic has signed a computing contract worth up to $120.5 billion (about 167 trillion won) through 2029 with SpaceXAI, SpaceX’s AI subsidiary. The agreement covers computing services using Nvidia GPUs and can be terminated with 90 days’ notice after an initial mandatory period. Its terms differ from the minimum spending commitments with Google, Amazon and Microsoft.
Anthropic has also partnered with “neocloud” companies that specialize in supplying AI computing resources, signing a $50 billion (about 69 trillion won) agreement with Fluidstack and a $41.3 billion (about 57 trillion won) deal with Lambda. Former Bitcoin mining companies TeraWulf and Hut 8 are also building data centers to provide computing capacity to Anthropic.
Speed is one reason Anthropic has brought these companies on board. Large companies such as Amazon and Microsoft need massive computing resources for their own AI services as well. Cryptocurrency mining companies, by contrast, may be able to convert sites and power infrastructure they already have into data centers, potentially saving time compared with building new facilities from scratch.
Eric Bernhardsson, CEO of computing services provider Modal, told the NYT that Anthropic is securing a substantial share of the computing capacity available on the market. Competition among AI companies is expanding beyond model development to a race to secure power and data centers first.
Designing data centers directly and buying chips and cooling equipment
Anthropic is not leaving facility construction entirely to its computing suppliers; it is also involved in design and equipment procurement. The company is hiring data center construction and supply-chain specialists and directly purchasing semiconductors from Nvidia, AMD and Broadcom. It is also securing supporting equipment, including cooling systems required for the chips, and supplying it to facility operators.
This approach can reduce the time data center builders spend sourcing equipment separately. Asher Genoot, CEO of Hut 8, told the NYT that Anthropic and data center operators are working together like a single team.
Anthropic has also taken a direct role in securing sites. Since last year, it has met with data center developers in places including Texas and Indiana to check permitting progress and the possibility of connecting to the power grid. It has also considered converting shuttered coal-fired power plants and factories into data centers to make use of existing industrial facilities.
Securing a facility, however, does not mean it can be put into operation immediately. Unstable power supplies can damage expensive AI chips, and specialized workers are needed to install transmission equipment and cooling systems.
Fluidstack, for example, raised concerns about power reliability in some data center designs proposed by Anthropic. The two sides reached a solution through discussions, but the episode illustrates how challenging power management can be for AI data centers.
TeraWulf, which is building a data center in upstate New York, had difficulty finding electricians. It had to compete for skilled electricians with construction of a new stadium for the NFL’s Buffalo Bills, which is underway nearby. TeraWulf ultimately asked local unions to provide electricians willing to work out of the area and secured workers by covering the additional costs.
Securing smaller facilities, too
Anthropic has lowered its threshold for data center sites. It initially focused on large facilities, but as supply tightened, it began considering sites with a capacity of 200 megawatts (MW). It later became willing to accept facilities as small as 50 MW and pursued plans to make up for power shortfalls with batteries and natural gas generation.
A 200 MW facility qualifies as a large data center, but it is not enough to meet the total computing capacity AI companies need. Anthropic is in a position where it must combine facilities secured in multiple locations to reach the required scale.
John Tough, managing partner at data center investment firm Energize Capital, said Anthropic may have to keep securing smaller facilities and adding capacity because it does not have complete control over its own computing supply chain. If a large data center is difficult to secure, operating several facilities at sites already connected to the transmission grid can be an alternative, he said.
As a result, power supply and the timeline for connecting to the transmission grid are becoming important competitive factors alongside data center sites. Even if a company buys semiconductors, it cannot use them for AI services without the space and power to install and run them.
Growing pressure to raise funds
The race to secure computing capacity is adding to Anthropic’s financial burden. Its amended securities filing submitted on November 2 put total commitments related to computing and infrastructure at $895.4 billion. Of that amount, $447.5 billion (about 620 trillion won) is scheduled to be spent by the end of 2030, with most of the remainder due between 2031 and 2045.
Payments due over the next 12 months alone amount to about $69.8 billion (about 97 trillion won). Not all of the commitments, however, represent firm payment obligations. About $213.8 billion of the total may be terminated if conditions such as financing or permitting are not met, or is subject to certain termination rights.
Revenue is also rising rapidly. Anthropic’s revenue for the third quarter of this year was about $17 billion (about 23.5 trillion won), up sharply from about $1.4 billion in the same period a year earlier. But computing and infrastructure costs during the same period totaled about $9.9 billion (about 13.7 trillion won), or 58% of revenue. The company recorded an operating loss of $567 million.
The company is turning to an initial public offering (IPO) and external financing to cover its growing investment costs. According to the NYT, Anthropic is pursuing a listing this month that could raise as much as $100 billion (about 138 trillion won). The size and pricing of the offering, as well as the listing schedule, have not yet been finalized.
Anthropic has also arranged a separate financing structure to secure semiconductors. The $63 billion (about 87 trillion won) financing transaction announced on October 26 involves acquiring Broadcom AI chips and networking equipment and then leasing them to Anthropic. Bank of America, Citi and Morgan Stanley are participating in the financing, while Blackstone is providing subordinated financing.
Will the race to secure computing capacity extend to data centers overseas?
Anthropic is also considering developing data centers overseas to reduce the risks of concentrating its infrastructure in the United States. In an earlier securities filing, it said it had signed a letter of intent for the development of overseas facilities, but did not disclose the countries involved.
An AI company’s competitiveness depends not only on model performance but also on the scale of computing capacity available to operate its models. Even with enough semiconductors, shortages of power, sites and construction workers can hamper service expansion.
Anthropic is reportedly set to secure more than 3.5 gigawatts (GW) of computing capacity this year and add a similar amount next year. However, the actual operating dates for each facility and the capacity that will be available under the terms of the contracts have not all been disclosed.
The key challenge ahead will be whether Anthropic can bring its contracted computing resources online as planned and use its rapidly growing revenue to bear the burden of long-term investment.