KDDI’s Osaka AI Data Center Turns Liquid Cooling Into A Power Test

KDDI Moves Liquid Cooling Into Osaka AI Infrastructure

KDDI is turning liquid cooling from a demonstration result into part of a commercial AI data-center buildout in Osaka, where dense GPU systems are making air cooling harder to defend.

The Japanese carrier worked with Mitsubishi Heavy Industries and NEC Networks & System Integration Corporation on immersion cooling tests at KDDI’s Oyama Network Center.

In 2023, the companies said server cooling energy use fell 94 percent compared with conventional air-cooled systems, while Power Usage Effectiveness dropped to 1.05.

KDDI has since moved beyond the lab.

The Osaka Sakai Data Center uses HPE systems and NVIDIA’s GB200 NVL72 rack-scale hardware.

KDDI is aiming WAKONX AI at startups and enterprises working on large language models, so the Osaka site is being positioned as a customer-facing AI platform rather than only a cooling showcase.

AI Racks Are Pushing Past Air Cooling

The pressure comes from electricity as much as chips.

Global data centers consumed roughly 415 terawatt-hours in 2024, and the International Energy Agency projects demand to exceed 1,000 terawatt-hours this year.

The same article puts a next-generation AI rack near one megawatt of power, around the level needed to supply 750 average American homes.

Conventional air-cooled facilities typically run at a PUE of around 1.7, which means cooling and overhead add another 0.7 units for each unit of server energy.

KDDI’s 1.05 test result narrows that overhead sharply, though it remains a controlled demonstration rather than multi-year hyperscale operating evidence.

The Efficiency Claim Still Has Limits

Immersion cooling solves one part of the energy problem: heat removal.

It does not reduce the electricity drawn by GPUs, and it adds service and maintenance complexity because submerged hardware must be lifted, drained and cleaned.

The dielectric fluids also create procurement and disposal questions.

That makes liquid cooling an infrastructure trade-off, not a simple upgrade.

Operators can cut cooling overhead, but they take on new capital costs, handling processes and retrofit limits in existing data centers.

Japan Treats Cooling As Industrial Capacity

Japan’s policy context raises the stakes.

PJM in the United States projects a six-gigawatt shortfall by 2027, while Japan has linked data-center energy demand to its 2050 carbon-neutrality target.

The local cooling market is also becoming a business signal.

IMARC Group puts Japan’s data center cooling market at $2.8 billion in 2025 and projects $7.2 billion by 2034.

The IEA also puts the global 2030 electricity-demand figure for data centers at 945 terawatt-hours.

KDDI’s unresolved burden is operational proof.

The company has a 94 percent cooling-energy claim, a PUE figure, an Osaka deployment and NVIDIA Blackwell hardware; the next test is whether those gains hold under sustained commercial AI workloads without making maintenance and fluid handling a new bottleneck.