According to Bloomberg on August 3, 2026, with Japanese outlets following the next day: NTT Data plans to spend at least JPY 1.5 trillion (roughly USD 9 to 9.6 billion) through 2033 on data centre construction and equipment in Japan, taking computing capacity to 1 gigawatt — about four times its current scale — by adding roughly 750 megawatts over the next seven years.
The reporting also spells out the basis for the arithmetic: each megawatt of installed data centre capacity in Japan costs around JPY 2 billion to 2.5 billion, excluding AI accelerators such as Nvidia's. On timing, roughly 42 megawatts is planned for this year, with construction on a 100 megawatt facility starting next year in Tochigi prefecture, north of Tokyo.
(These are reported figures, disclosed on a "said to" basis. We quote them as published and don't extrapolate.)
Putting It in Regional Context
In our piece on APAC's record 26.5GW data centre pipeline we cited figures from the same period: Japan added 293 megawatts of operational capacity to reach roughly 1.8 gigawatts, second in APAC only to mainland China. So one operator adding 750 megawatts over seven years represents a substantial addition on top of the current national base — from a single company.
Three Conclusions That Actually Affect Ordinary Users
- This capacity is being built for AI, not for cheap general-purpose VPS. High-density racks, liquid cooling, and priority for large customers on long contracts — that's the shape of the current build-out. So don't translate "Japan is building a lot of data centres" into "Tokyo VPS prices will fall". General-purpose virtualisation and AI training and inference compete for the same power and land, and the pressure from that runs upward.
- New capacity is moving outside Tokyo. Power and land constraints in Tokyo are a documented limit, which is why new projects keep naming Tochigi and Osaka, and further out Hokkaido and Kyushu. The practical consequence: when you see "Japan node", ask which metro it's actually in. Paths and return routes from Osaka or Tochigi to mainland China, to Southeast Asia, and to users in Tokyo itself can differ from a Tokyo facility — and those differences are measurable in latency.
- No capital expenditure announcement measures latency for you. No investment story can tell you the RTT between your users and a specific machine. That only comes from testing.
Turning This Into a Selection Decision
People choose a Japanese node for one of two reasons: serving users inside Japan, or serving mainland China with a short, stable return path. The two are validated differently:
- Measure before concluding. ping and mtr give you the baseline — how to test VPS latency and packet loss has the commands and how to read them. For throughput and real download behaviour, use speedtest, iperf3 and real download tests.
- For mainland China, look at the return path separately. A good outbound path doesn't imply a usable return path; optimising latency from mainland China to a Japan VPS and CN2 GIA, BGP and China-optimised routes explained cover how to tell them apart.
- Then come back to the requirement. How to pick a low-latency data centre node for your workload works backwards from workload type rather than from how big a facility is.
If what you want is simply an instance in Tokyo, how to choose a Japan VPS: a complete guide to Tokyo nodes breaks down the variables — instance type, bandwidth, routing — and SharkCloud's Japan node plans page lists what's currently available. The relationship between gigawatt-scale industry build-out and the single machine you rent is actually quite plain: it shapes where power and racks go over the next several years, while your experience today is still decided by that specific machine's routing and its measured latency.