Anthropic’s First Data Center in Japan Is a 500MW Bet on Osaka

NTT Data Group is in the final stages of talks to build a 500MW data center for Anthropic in Osaka Prefecture, as reported by Nikkei Asia. The construction cost is put at $6.3bn, which would make it the first dedicated facility the lab owns in Japan and one of the largest single loads in the country.

Neither company would comment, so every number here rests on people who were not named. The agreement is targeted for the end of fiscal 2026, with partial operations expected in 2028. NTT Data operates more than 150 data centers across 20 plus countries and already runs two sites in Osaka, so this is an expansion of existing power and fiber rather than a greenfield gamble.

The megawatts matter more than the price tag

Anthropic has signed agreements covering at least 14.8GW of compute since October last year, spread across leases and cloud contracts with Akamai, Amazon Web Services, CoreWeave, Fluidstack, Google Cloud, Lambda, Microsoft and AMD. A 500MW building is small against that pipeline. It is also the largest piece of dedicated capacity the company would own outright in Japan, which is the part Japanese customers will hear about, as Data Center Dynamics laid out from the same reporting.

Land is the cheap input. At 500MW the binding constraints are the interconnection queue, the generation contract behind the site, and how much of a local grid one tenant can claim without a fight. Japan has been slower than the United States to sign hyperscale power deals, and Osaka puts the load next to a dense industrial grid instead of a remote generation site.

Sovereign AI is becoming a real estate problem

Anthropic has not said why Japan. The wider pattern is easier to read than any one announcement. Regulated buyers in finance, health and the public sector want inference to run inside a jurisdiction they can name, and that turns a model vendor into a tenant with a lease and a substation. NTT Data gets an anchor tenant for a 500MW shell. Anthropic gets a domestic story to tell before the public offering it filed for confidentially in June.

Three questions if you are buying AI capacity this year. Can you name the substation behind your inference, not just the region string in a configuration file? Does your contract survive a queue delay at the interconnection, or does the price reset? And if a regulator asks where the tokens were generated, do you have an answer that is not a vendor diagram?

Related reading. Data Center Fires and Power Failures Doubled Since 2024, and the Power Gear Is Why and The Firmus IPO Died on Delivered Megawatts, Not on Price track the same seam between AI demand and the power actually available.