Tokyo Tests Turning AI Data Center Heat Into Cooling and Hot Water

A Tokyo trial puts discarded server heat to work

Tokyo is backing a project to recover heat directly from artificial intelligence servers and use it for cooling buildings, supplying nearby bathing facilities and supporting factories. The demonstration runs from July 16, 2026, to March 31, 2028, with the metropolitan government contributing up to 520 million yen toward eligible costs.

Tokyo Electric Power Company Holdings, known as TEPCO, is leading the work with MESH-X and NHK Technologies at a data center in Tokyo. Their target is a practical problem growing alongside AI: much of the electricity entering computing equipment eventually becomes heat, which operators must remove to keep the machines running.

The Tokyo government announcement describes a system that uses evaporation to cool servers fitted with graphics processing units, or GPUs, while collecting their heat directly at a high temperature. GPUs are chips that perform many calculations at once, making them central to modern AI.

The project brings together two needs usually managed separately. Data centers need cooling, while neighboring buildings and businesses need heat. Recovering useful heat could help the same electricity serve both computing and local energy needs.

Japan is also testing another approach: moving AI calculations between cities or changing when they run to match electricity availability. Together, the experiments address how AI uses power, where that power is consumed and what happens to the energy afterward.

How WaBiCo captures heat from chips

TEPCO calls its approach WaBiCo, short for Watt Bit Cogeneration. The name connects watts, a measure of power, with bits, units of digital information. Its central idea is to treat powerful electronic equipment as both a computing resource and a source of usable heat.

The cooling technology places water under reduced pressure, around half atmospheric pressure, so that it boils at roughly 80 degrees Celsius rather than the familiar 100 degrees. Heat from the GPU drives evaporation, allowing the system to collect thermal energy at a temperature useful for other applications.

TEPCO describes a purpose built cold plate that directly cools the GPU and recovers its heat. The circulation approach uses heat itself to move heat, without a pump driving that process. Capturing energy close to the chip offers a way to obtain hotter, more useful heat than allowing it to disperse through a server room.

Temperature matters because different jobs require different grades of heat. Warm air that is expensive to collect has limited uses. Heat recovered at around 80 degrees Celsius can serve a wider range of building and industrial applications, depending on their equipment and temperature requirements.

The proposed uses include supplying nearby bathing facilities and factories, providing district heating and using recovered heat as an energy source for cooling server rooms or neighboring facilities. Cooling driven by heat uses thermal energy to operate a cooling cycle, rather than relying entirely on an electrically driven compressor.

What Tokyo is funding and testing

TEPCO's project adoption announcement says Tokyo selected the proposal on June 30, 2026. The company identifies itself as the lead operator and sets out the government's maximum contribution of 520 million yen as partial support for specified expenses.

Tokyo's industrial and labor affairs bureau places the demonstration within its 2050 Tokyo Strategy, under the goal of maximizing energy efficiency. The practical work combines development of an evaporation cooling system with testing in a real operating environment.

The project must connect heat recovery to a useful destination. For a bathing facility, that means a dependable supply of thermal energy. For a factory, it means matching the temperature and timing of an industrial process. For server rooms, it means fitting the recovered heat into a cooling system that supports reliable operation.

The announcements identify the demonstration site only as a data center in Tokyo. They give no measured heat recovery rate, electricity savings, emissions reduction, named heat customers or commercial launch date. Those results will determine how useful the approach becomes beyond the trial.

Cooling efficiency and heat reuse measure different gains

TEPCO aims to improve a standard data center efficiency measure called Power Usage Effectiveness, or PUE. It divides all the electricity used by a facility by the electricity used by its computing equipment. A result closer to 1 means less additional electricity is being spent on cooling and other supporting systems.

A hypothetical facility using 120 units of electricity while its computers use 100 would have a PUE of 1.2. Reducing the extra 20 units would improve that figure. The Tokyo project seeks such gains by changing how heat is removed from the chips.

Heat reuse provides another possible benefit: a neighboring user could meet part of its heating demand with energy recovered from computing. The electricity still powers the servers, while useful thermal energy leaves the facility for another purpose.

The two benefits need separate measurements. Operators need to track electricity used for cooling, heat actually delivered to customers and the energy those customers would otherwise consume. A working heat connection becomes valuable when the receiving facility can use the energy at the time it arrives.

More efficient chips meet expanding AI workloads

NVIDIA describes energy as the foundation of AI infrastructure. At the January 2026 Davos meeting, chief executive Jensen Huang compared the industry to a cake with five layers: energy at the bottom, followed by chips, infrastructure, models and applications.

NVIDIA representative Hiroki Saito describes energy as AI's first gateway. His account also illustrates the speed at which computing demands are changing. Using ChatGPT's November 2022 arrival as a starting point, he puts the processing requirement associated with the January 2025 DeepSeek reasoning model at 100 times that level, and the OpenClaw autonomous agent roughly a year later at 1,000 times.

Those examples concern changing kinds of AI work. Reasoning models spend computing effort working through problems, while agents can perform sequences of tasks. More elaborate tasks can consume more processing capacity even as individual chips become more efficient.

NVIDIA says its Rubin generation can perform 10 times as much agent processing as Blackwell using the same power. The practical value depends on how the equipment is deployed and which tasks it runs.

Saito also describes a change in operating priorities. Traditional data centers concentrated on staying available. An AI factory, NVIDIA's term for infrastructure built around AI production, must also maximize useful computing output. Better performance per unit of electricity and growing demand are moving at the same time, leaving total consumption dependent on both.

Sudden power swings create another problem

AI facilities face challenges beyond their total electricity use. During model training, hundreds of thousands of GPUs can change their power consumption together on timescales measured in milliseconds. Electrical equipment must absorb those rapid changes while keeping power stable.

Amber Villegas-Williamson, principal consultant at the Uptime Institute, which advises electricity suppliers and data centers on reliability, described the strain:

AI does create very unusual power demand,

Villegas-Williamson compared the wear to repeatedly pushing a car engine harder rather than maintaining a steady speed. Drew Baglino, a former Tesla executive and founder of Heron Power Electronics, said AI demand can briefly rise as much as 50% above a facility's design capacity. In his example, a data center designed for 1 gigawatt could draw 1.5 gigawatts for a split second.

Equipment failures associated with these stresses include cracked turbines and batteries needing replacement within weeks or months. Some facilities have reportedly achieved availability closer to 80%, while estimated revenue losses during outages range from thousands to hundreds of thousands of dollars per minute, depending on the operation.

The North American Electric Reliability Corporation found that about three quarters of load models covering more than 33 gigawatts of American data center demand were insufficient to represent dynamic behavior. These models help utilities predict how customers affect the power system, making rapid changes in consumption an important planning issue.

A planned campus of 2.67 gigawatts in West Texas illustrates the scheduling consequences. Developer Joulent added engineering time to meet Microsoft's requirement for 99.999% reliability, moving the start of power delivery from 2027 to 2028. At that reliability level, annual downtime would amount to about five minutes.

Moving calculations instead of moving electricity

Japan's second experiment tackles where and when AI consumes electricity. J-Power, Cisco Systems, East Japan Railway, Hitachi and Bitmedia announced on October 8, 2026, that they were undertaking a demonstration connecting AI data centers in Yokohama and Kitakyushu.

The trial began in August 2026 and is scheduled to run until February 1, 2027. It tests workload shifting, which means assigning computing tasks to another location or another time according to electricity and computing conditions.

A spatial shift sends a calculation to a different data center. A time shift schedules it for a different hour. The system considers renewable generation forecasts, electricity trading prices, supply and demand, available computing capacity and information about disasters.

One goal is to use renewable electricity that might otherwise be curtailed. Curtailment happens when solar or wind generation is temporarily reduced because the network cannot carry or absorb all the available power. Moving suitable AI work toward that electricity could increase its use.

The companies will also examine the extra processing time created by shifting work and how to preserve service quality. Moving a task has practical costs, so the trial measures both the electricity benefit and the effect on computing performance.

Railway fiber links the wider experiment

The five companies are also testing an all photonic network along the Joetsu Shinkansen. Optical transmission equipment at Echigo-Yuzawa, Jomo-Kogen and Takasaki stations will use existing railway fiber to examine communication quality under vibration and changing weather conditions.

The tests cover signal loss, transmission errors, delay and variation in delay. These measurements matter when computing facilities in different places need to operate together. The railway network experiment is a distinct part of the program alongside workload shifting between Yokohama and Kitakyushu.

J-Power leads the consortium and evaluates results. Hitachi builds the workload shifting system and optical test environment, Cisco supports network design and measurement, JR East examines railway communications equipment, and Bitmedia develops the system within the data centers.

The work follows the creation of a council in April 2026 that also includes JR West Optical Network and Nagoya Railroad. Its broader aim is coordinated operation of distributed AI data centers using wide area optical networks and workload shifting.

The main milestones show how the heat and computing experiments overlap:

• April 2026: Seven companies establish a council for optical networking and workload shifting.
• June 30, 2026: Tokyo selects TEPCO's heat recovery proposal.
• July 16, 2026: The Tokyo heat demonstration period begins.
• August 2026: The Yokohama and Kitakyushu workload shifting trial begins.
• February 1, 2027: The workload shifting demonstration is scheduled to end.
• March 31, 2028: The Tokyo heat recovery project is scheduled to end.

Japan wants more computing with less wasted energy

Japan is pursuing AI and semiconductor investment while working toward carbon neutrality by 2050. TSMC's factories in Kumamoto Prefecture form part of that industrial expansion. Renewable energy supplied around 22% of Japanese electricity generation in 2022, leaving the country heavily dependent on other sources, including fossil fuels.

TEPCO's own fifth comprehensive special business plan sets a goal of putting growth in metropolitan Tokyo's data center demand among the world's highest by fiscal 2040. Its heat project therefore sits alongside an explicit ambition to expand the sector.

Heat recovery and workload shifting address different parts of that expansion. WaBiCo seeks more useful output from electricity already consumed by servers. Workload shifting seeks a better match between computing demand and electricity availability. Equipment that handles rapid power swings supports the reliability both approaches require.

For nearby communities, the concrete opportunities are useful heat and a closer connection between data centers and local energy needs. For operators, the tests concern cooling costs, productive computing time and dependable service. The next milestone is measured performance: how much energy is saved or reused while the computers continue doing their work.

Key Points

• Tokyo is supporting a TEPCO heat recovery demonstration with up to 520 million yen toward eligible costs.
• The project runs until March 31, 2028, and targets cooling, bathing facilities, factories and district heating.
• WaBiCo uses reduced pressure evaporation to recover useful heat directly from GPUs.
• A separate trial shifts AI tasks between Yokohama and Kitakyushu and across operating times to match electricity conditions.
• Rapid fluctuations in AI power consumption are also creating equipment and reliability problems.
• Measured energy savings, recovered heat deliveries and operating performance will determine the practical value of the demonstrations.