GPU capacity in Japan · Go-live August 2027
We're building liquid-cooled GPU sites in Japan where the grid connection is already confirmed by the utility. Reserve whole NVL72 racks now. Reservations are non-binding.
Why reserve now
New data-center connections around Tokyo and Osaka queue for years. Our 2 MW sites sit in regions where the utility has already replied that it can supply power, so the racks get built on a fixed schedule.
Utility supply replies in hand for each site. No multi-year queue.
Rack-level contracts for 3–5 years. Your price and allocation don't move.
Dedicated, physically separated racks. Useful for co-productions with Japanese studios.
Factory-built container modules, assembled on site while civil works run in parallel.
Specifications
| Compute | 12 × NVIDIA GB300 NVL72 (864 GPUs), 72 GPUs per NVLink domain |
|---|---|
| IT load | About 1,722 kW per site; site demand under 2 MW |
| Cooling | Direct liquid cooling with closed-circuit cooling towers; dry operation below about 15 °C outside |
| Efficiency | Design PUE about 1.06 annual average, 1.09 at peak |
| Power path | A feed through UPS with lithium-ion batteries; B feed from the grid |
| Build | Six 40 ft container modules per site, single storey |
| Locations | Chubu (Aichi Prefecture) and Kyushu (Kagoshima Prefecture), Japan |
| Service | Dedicated rack clusters; bare-metal access; network and storage specified with each customer |
Availability, stated plainly. To keep cost and build time down there are no diesel generators. Batteries carry the site through short outages; if a grid outage lasts more than about 20 minutes, jobs are checkpointed and the GPUs shut down in order. That suits training, rendering and batch generation. It isn't designed for always-on, latency-critical inference.
Sites
2 MW, 12 NVL72 racks. Grid supply in about four months from order.
2 MW, 12 NVL72 racks. Built in parallel with Chubu.
Two more 2 MW sites with grid replies, and about ten more in review, including one up to 30 MW.
Pricing
Final pricing depends on the number of racks, the term and payment terms. Reserving now holds this range.
Who it's for
Rendering, generative-AI shots and custom style models, with footage and models kept in Japan.
Wholesale rack capacity or power-and-facility space when you're short of megawatts.
Annual reservations that fit grant and fiscal-year budgets, dedicated or shared racks.
Team
CEO
Twelve years in AI, from machine learning to deep learning. Founder of an AI software company.
CTO
Software engineering and data-center infrastructure. Founder of a software company.
COO · Tokyo
Solar, storage and data-center projects in Japan, from grid connection and permits to commissioning, across Hokkaido, Akita, Aomori, Nagano and Fukuoka.
Timeline
Reserve capacity
A reservation is not a contract. It holds a place in the allocation and the price range above. We'll reply within two business days, and send a short letter of intent if you'd like to firm it up.
Prefer email? hello@infitoken.com
FAQ
No. It's an expression of interest. Binding terms are only set in a service agreement, which we'd sign closer to go-live.
Long-term pricing well below on-demand GB300 list prices, a confirmed build schedule, and data residency in Japan, which matters for work with Japanese studios and partners.
Round trips across the Pacific add on the order of 100 ms. That doesn't affect training, rendering or batch jobs. Large data sets can move over dedicated links or by bulk transfer.
Racks are dedicated and physically separated. For studios that require it, we plan to complete an MPA Trusted Partner Network assessment before go-live.
Yes. Annual reservations are available and can be timed to grant or fiscal-year budgets.