Anthropic Plans $6.3B Japan Datacenter But Power Is The Bottleneck
Anthropic and NTT Data are reportedly building a 500MW AI datacenter in Osaka, proving the next era of AI is a race for electricity.
Business · Source: Reuters
What happened
NTT Data Group is reportedly in final-stage negotiations to build Anthropic's first dedicated datacenter in Japan. The facility will be located in Osaka prefecture. Construction costs could reach one trillion yen, or roughly $6.3 billion. The two companies are aiming to sign an official agreement by March 2027. If everything goes according to plan, partial operations will begin as early as 2028.
The scale is massive. The datacenter is planned to have a total power intake capacity of up to 500 megawatts. This scale places it among the most substantial computing facilities in Japan. AI datacenters require high-performance servers packed with specialized GPUs for training and running models. Installing these advanced servers at scale is exactly what pushes the total investment into the billions.
Power is the real bottleneck. Pulling 500 megawatts from the existing electrical grid takes considerable time. To avoid long delays, the project will initially explore alternative energy sources. Reports indicate they are looking at on-site power generation and fuel cells. NTT Data, the third-largest datacenter operator globally, will also use advanced liquid cooling systems to manage the intense heat generated by the equipment.
Key facts
- $6.3 billion — Potential construction cost of the Osaka facility
- 500 MW — Planned power intake capacity
- 32.7 trillion yen — Japan's data centre investment target by 2035
- 1,600 MW — NTT Data Group's combined global datacenter capacity
Why it matters
The AI race is now an electricity race. Compute is no longer just about buying GPUs. It is about finding the physical land and securing the raw megawatts to plug those chips in. A 500-megawatt facility requires the same amount of power as a small city. When companies like Anthropic have to explore building their own power plants just to turn on their servers, it shows a fundamental shift. Infrastructure constraints will dictate which AI models scale and how fast they reach the market.
Japan is aggressively positioning itself as the landlord for the global AI boom. The Japanese government has designated datacenters as a strategic sector. They have set a massive target to mobilize 32.7 trillion yen in public and private investment for datacenters, cloud computing, and storage batteries by 2035. We are already seeing a surge in proposals. There is a $12.6 billion plan for a 500-megawatt facility in Akita prefecture and a $5.4 billion plan for a 350-megawatt facility in Kagoshima. Japan offers a clear benchmark for how economies can use public-private partnerships to attract foreign AI developers.
For builders
Energy tech is the new AI infrastructure
Datacenters cannot wait years for traditional grid upgrades. Startups building on-site power generation, fuel cells, and liquid cooling systems have a massive opportunity. Major AI labs and datacenter operators will pay a premium to get facilities online faster.
Geographic diversification of compute
Anthropic is actively looking beyond the United States to places like Japan and Australia. Governments are heavily subsidizing these moves to attract AI investments. Founders building AI infrastructure should look at Asian markets where public-private partnerships are accelerating deployment.
Cooling systems are mission critical
High-performance GPUs generate intense heat that traditional air cooling cannot handle. NTT Data is specifically leveraging liquid cooling systems that use circulating fluids. Hardware engineers who can innovate in thermal management will find eager buyers among top datacenter operators.
My take
Everyone is obsessed with who has the best model, but the real moat is electricity. Anthropic exploring on-site power generation just to turn on its servers shows how constrained this market really is. If you want to win in AI over the next decade, you do not just need smart researchers. You need to figure out how to generate and cool 500 megawatts of continuous power.
Original reporting: Reuters. This is my rewrite and opinion.