SILICON VALLEY — Anthropic, the developer of the Claude AI model, has signed computing capacity contracts worth $517 billion, or about 700 trillion won, in less than a year, according to a tally of its deals. Running AI models requires enormous computing capacity, not only for performance but also for inference. Korean chipmakers that supply AI servers, a core component of computing infrastructure, are expected to benefit.
The Information reported on the 6th that the computing capacity purchase agreements Anthropic signed over the past 11 months total 14.8 gigawatts in power terms and $517 billion in value. Anthropic will spend that amount over the next 10 years.
Before October last year, Anthropic had secured only 1 to 2 gigawatts of computing capacity. It has added at least 14.8 gigawatts in the 11 months since. That much capacity would require running more than 10 of Korea's most advanced nuclear reactors, of the APR1400 class. It is comparable to the electricity consumed by a metropolitan area of 26 million people. Computing capacity is measured in power units because operating the data centers that deliver it requires vast amounts of electricity.
The dollar figure has also jumped in the space of a year. In December last year, Anthropic told investors its server leasing costs through 2029 would total $180 billion. The $517 billion to be spent over 10 years equals roughly 24% to 36% of the annual AI capital spending reported by big tech companies such as Amazon, Microsoft and Meta. Those budgets, however, also cover general-purpose cloud services, logistics, consumer hardware and office facilities, so measured on computing alone, Anthropic's investment is not far behind big tech.
The largest of the deals struck over the 11 months is an 11-gigawatt arrangement under which Amazon and Google supply capacity to Anthropic, worth more than $300 billion over 10 years. In November last year, Anthropic agreed to pay $30 billion to lease 1 gigawatt of Microsoft's Azure servers, and in May this year it signed a contract with SpaceX worth up to $45 billion.
Anthropic has also begun buying AI chips directly and building its own data centers. In April this year, it agreed to purchase tensor processing units, or TPUs, produced by Broadcom and Google, followed by a chip purchase agreement with AMD in July. In November last year, it committed $50 billion with cloud startup Fluidstack to build its own data centers in Texas and New York.
Investors have focused on the scale of Anthropic's computing capacity ahead of an initial public offering expected in late October or early November. Securing sufficient capacity is essential to staying ahead in the race to develop and operate AI models.
OpenAI, Anthropic's biggest rival, is also moving aggressively as it prepares for a listing early next year. OpenAI plans to spend up to $750 billion for 30 gigawatts by 2030. Its Stargate data center joint venture was delayed by disputes among the partners, but the company recently restarted its project in Georgia with OpenAI leading the design and construction itself.
The impact on Samsung Electronics and SK hynix is also drawing attention. In July, Anthropic disclosed that it had signed supply agreements with both companies. Samsung Electronics is expected to play a role in foundry work for Anthropic as well as in manufacturing the company's in-house chips designed by Broadcom. SK hynix has signed a contract to supply HBM4, the latest generation of high-bandwidth memory, to Nvidia, a key Anthropic partner.
