PROBLEM: International information middle electrical energy consumption is on monitor to exceed 1,000 terawatt-hours this yr, straining energy grids and threatening to derail each AI growth and nationwide local weather targets.
SOLUTION: Japanese telecom big KDDI, working with Mitsubishi Heavy Industries, has commercialized liquid immersion cooling for AI information facilities, slicing server cooling vitality use by 94 p.c and is now deploying it at industrial scale in Osaka.
AI is working out of energy. Not computing energy, however electrical energy. International information facilities have consumed roughly 415 terawatt-hours in 2024. With the Worldwide Power Company projecting that determine to exceed 1,000 terawatt-hours this yr. Information facilities may eat the equal to Japan’s total annual electrical energy use. A single rack of next-generation AI chips now attracts shut to 1 megawatt of energy, sufficient to produce round 750 common American houses.
Cooling these racks accounts for a considerable portion of that load. Standard air-cooled information facilities run at a Energy Utilization Effectiveness, or PUE, of round 1.7, that means that for each unit of vitality powering a server, one other 0.7 models are spent on cooling and overhead. As AI workloads develop denser and warmer, that overhead is turning into unsustainable.
KDDI Company, certainly one of Japan’s three main cell carriers, thinks they’ve give you the reply.
Servers in fluid
In 2023, KDDI, Mitsubishi Heavy Industries (MHI), and community integrator NEC Networks & System Integration Company accomplished demonstration testing of an immersion cooling system at KDDI’s Oyama Community Heart, in Tochigi Prefecture north of Tokyo. The strategy submerges servers in a tank of dielectric fluid, a non-conductive liquid that attracts warmth immediately from the chips with out electricity-hungry air con. The outcomes had been hanging: cooling vitality consumption fell 94 p.c in contrast with typical air-cooled methods, and PUE dropped to 1.05.
That 0.05 overhead, towards an trade norm of 0.7, represents a significant effectivity achieve for any operator working tons of of megawatts of AI compute across the clock.
The three corporations moved from demonstration to commercialization in 2023, with containerized deployments designed for versatile set up. Final yr, KDDI took the subsequent step, saying a full-scale AI information middle in Sakai Metropolis, Osaka Prefecture, constructed with HPE and powered by NVIDIA’s GB200 NVL72 rack-scale methods. The Osaka Sakai Information Heart combines air cooling and direct liquid cooling in a hybrid design concentrating on lowered environmental impression for AI coaching workloads. KDDI plans to serve startups and enterprises constructing massive language fashions by way of its WAKONX AI enterprise platform.
A business reply to an industrial drawback
The broader context issues. Information middle energy demand is rising sooner than grid infrastructure can assist. In the US, grid operator PJM initiatives a six-gigawatt shortfall by 2027. In Japan, the federal government has flagged information middle vitality consumption as a structural problem to its 2050 carbon neutrality goal.
Liquid cooling shouldn’t be a brand new concept. Google, Microsoft and Meta have all experimented with variations. What KDDI and MHI have demonstrated is a verified path from lab trial to Tier 4-certified business operation in a hyperscale setting, utilizing tools developed and manufactured in Japan.
Immersion cooling does have actual constraints. Servers submerged in fluid are tougher to service than rack-mounted air-cooled tools. Upkeep requires lifting {hardware} out of tanks, draining and cleansing elements, including time and complexity. The dielectric fluids used carry their very own procurement and disposal concerns. And immersion cooling addresses the vitality price of cooling particularly, not the entire energy draw of the AI chips themselves, which continues to rise with every new GPU era.
Japan’s liquid cooling push is not going to, by itself, resolve the worldwide information middle vitality disaster. However KDDI’s development from a container trial to an operational Osaka facility with NVIDIA Blackwell {hardware} illustrates how effectivity beneficial properties that started as analysis claims have gotten the baseline expectation for any new AI infrastructure constructed right now.
Proof and context
The 94 p.c cooling vitality discount comes from the February 2023 demonstration take a look at at KDDI’s Oyama Community Heart, collectively carried out by KDDI, MHI and NESIC, and printed in a press launch by all three events. The PUE of 1.05 compares towards a typical typical information middle PUE of 1.7. These figures come from a managed demonstration, not but from multi-year operational information at hyperscale. No unbiased third-party audit of the 94 p.c determine has been publicly printed; the declare ought to be handled as manufacturer-reported till verified.
Japan’s information middle cooling market reached $2.8 billion in 2025 and is projected to develop to $7.2 billion by 2034, based on IMARC Group, pushed by AI adoption and Japan’s carbon neutrality commitments. Globally, information middle electrical energy demand is projected to succeed in 945 terawatt-hours by 2030, based on the IEA.
KDDI shouldn’t be alone in Japan’s liquid cooling area. Fujitsu launched power-saving liquid cooling applied sciences for AI information facilities in 2024, integrating edge computing for fault detection. Toshiba showcased energy semiconductor options at APEC 2026, together with a top-side cooled TOGT package deal engineered for thermally demanding, excessive power-density functions, enabling warmth dissipation by way of the highest of the package deal to scale back thermal stress on the PCB. Globally, Google, Microsoft and Meta are pursuing nuclear and large-scale renewable procurement as complementary methods, whereas corporations together with ZutaCore are providing direct-to-chip liquid cooling methods that MHI has additionally integrated into its product lineup.
The restrictions are actual. Immersion cooling reduces the vitality price of conserving servers chilly, however doesn’t scale back the ability drawn by the AI chips themselves. Every new GPU era calls for extra electrical energy per rack, that means effectivity beneficial properties threat being outpaced by uncooked compute development. Liquid cooling infrastructure is extra capital-intensive upfront than air cooling. Adoption in legacy information facilities is gradual as a result of retrofitting current amenities is dear. And the dielectric fluids utilized in single-phase immersion methods require cautious dealing with and disposal, including operational complexity.
The core perception is easy. Standard air con was designed for a world the place server racks drew 10 to 50 kilowatts. AI racks now method one megawatt. The physics of air cooling at that density don’t work. Japan’s funding in liquid cooling as a business product, reasonably than a analysis experiment, displays an early recognition that the infrastructure assumptions of the cloud period can not merely be scaled as much as meet the AI period’s calls for.
For inquiries relating to this text, please contact [email protected].
