How Much Does It Cost to Cool an AI Data Center?

Published on July 10, 2026 | Written by Alec Pow
This article was researched using 14 sources. See our methodology and corrections policy.

Cooling a large AI data center can cost $100 million to $150 million to build for a 100-megawatt campus, followed by roughly $20 million to $40 million per year for cooling electricity, maintenance, and water under a realistic planning scenario. At the extreme end, the cooling system for a one-gigawatt AI campus could cost $1 billion to $1.5 billion before the first chatbot request is processed.

Those numbers are easier to grasp by the day. A 100-megawatt AI campus may spend about $54,000 to $107,000 every day keeping its chips within safe temperatures after electricity, maintenance, and water are included. A one-gigawatt AI factory could spend more than $400,000 to $800,000 per day on cooling electricity alone.

OpenAI does not publish a complete daily cooling bill for ChatGPT. A published research model for an AI service handling one billion requests per day provides a useful scale comparison. Under that model, cooling electricity could cost about $13,800 to $27,600 per day for ordinary prompts, rising to roughly $31,000 to $62,000 per day when 10% of requests are long, computation-heavy queries.

AI cooling cost includes far more than fans. The invoice can contain liquid-cooled plates attached to GPUs, pumps, pipes, heat exchangers, chillers, cooling towers, dry coolers, water treatment, leak detection, backup equipment, controls, construction, and years of electricity. Freshwater use can range from a large initial fill for a closed-loop system to millions of gallons every day for evaporative cooling.

Article Highlights

Jump to sections

• A 100-megawatt AI campus may need $100 million to $150 million (at $30 per hour, earning that amount would take about 1.6 to 2.4 thousand years of full-time work, before taxes) of cooling construction.
• The same campus can carry a cooling operating bill near $20 million to $40 million per year.
• A one-gigawatt AI factory may need $1 billion to $1.5 billion of cooling infrastructure.
• A modeled service handling one billion AI prompts daily could spend $5 million to $10 million per year on cooling electricity under ordinary workloads.
• Long reasoning requests could push that modeled cooling-electricity bill toward $11 million to $23 million per year.
• A large water-cooled data center can consume as much water as a city of 50,000 people.
• Closed-loop cooling cuts ongoing water withdrawals, but a large campus may still need millions of gallons for its first fill and later maintenance.
• A water-cooled AI campus consuming 5 million gallons per day uses as much household water as roughly 61,000 Americans, based on the EPA average of 82 gallons per person per day.
• That same campus would consume about 7.6 Olympic swimming pools of water every day, or nearly 2,800 pools per year.

How Much Does It Cost to Cool an AI Data Center?

AI facility	What that size means	Cooling construction	Cooling electricity per year	Cooling electricity per day
1 MW AI cluster	A small dedicated AI room or several high-density racks	$1 million to $1.5 million (about 16 to 24 years of full-time work at $30 per hour)	$151,000 to $302,000	$414 to $828
10 MW AI facility	A regional AI building or major corporate cluster	$10 million to $15 million	$1.51 million to $3.02 million	$4,140 to $8,280
100 MW AI campus	A hyperscale campus large enough to draw power like a city	$100 million to $150 million	$15.1 million to $30.2 million	$41,400 to $82,800
1 GW AI factory	A multibuilding Stargate-class development	$1 billion to $1.5 billion	$151 million to $302 million	$414,000 to $828,000

The construction ranges are planning estimates, not public vendor price lists. KPMG modeled liquid-cooling technology, installation, and commissioning near €900,000 to €1.5 million per megawatt, depending on whether the project is a new installation or a difficult retrofit, in its liquid-cooling cost study. The table uses a rounded $1 million to $1.5 million per megawatt band to avoid presenting currency conversion and project differences as false precision.

The electricity figures assume cooling consumes power equal to 20% to 40% of the computing load. ASHRAE reports that data centers can spend 20% to 40% of their energy on cooling in its AI data center framework. Electricity is priced at the 2025 U.S. industrial average of 8.62 cents per kilowatt-hour, published by the U.S. Energy Information Administration in its electricity price data.

A 100 MW AI Data Center

A megawatt measures power, not the number of chatbots inside a building. One megawatt is one million watts used at a given moment. A 100-megawatt AI campus can feed around 100 million watts continuously into servers before cooling and electrical losses are counted.

Crusoe has described each 100-megawatt Stargate data center building as drawing enough electricity for a city of roughly 100,000 people. That comparison is based on power capacity rather than annual household consumption, but it explains why a single AI building can affect a regional electrical grid. Crusoe’s Abilene expansion is designed to reach 1.2 gigawatts across eight buildings and about 4 million square feet, according to its Abilene campus announcement.

If a 100-megawatt building needs another 20 to 40 megawatts for cooling, the cooling machinery by itself can draw enough power for a sizable community. Running that equipment around the clock produces an electricity bill of roughly $15.1 million to $30.2 million (about 242 to 484 years of full-time work at $30 per hour) per year at the national industrial rate used here.

The cooling bill is not fixed. A cool northern climate may use outside air or dry coolers for part of the year. A desert campus may need more fan, pump, compressor, or evaporative work during summer. A company with a private power contract may pay less per kilowatt-hour, while grid congestion, demand charges, backup generation, and new transmission can raise the effective cost.

Cost to Cool One Billion AI Prompts

OpenAI does not disclose enough information to calculate the real ChatGPT cooling bill. The missing pieces include the daily request count, mix of models, request length, GPU fleet, server utilization, data center locations, electric rates, and cooling systems.

A 2025 Microsoft-affiliated research paper modeled a large AI service handling one billion prompts per day. It estimated about 0.8 gigawatt-hours of total electricity per day for ordinary frontier-model queries. If 10% of requests were long reasoning queries, daily electricity rose to about 1.8 gigawatt-hours. The paper also estimated a median ordinary query at 0.34 watt-hours, while a much longer test-time reasoning query reached 4.32 watt-hours. The figures appear in the published AI inference energy study.

Applying ASHRAE’s broad 20% to 40% cooling share creates the following scenario:

One billion ordinary prompts each day:

• Total AI electricity: 800 megawatt-hours per day
• Illustrative cooling electricity: 160 to 320 megawatt-hours per day
• Cooling electricity bill: $13,792 to $27,584 per day
• Annual cooling electricity bill: $5.03 million to $10.07 million

One billion daily prompts with 10% long reasoning requests:

• Total AI electricity: 1,800 megawatt-hours per day
• Illustrative cooling electricity: 360 to 720 megawatt-hours per day
• Cooling electricity bill: $31,032 to $62,064 per day
• Annual cooling electricity bill: $11.33 million to $22.65 million

This is not a ChatGPT invoice. It is a published one-billion-query workload translated into a cooling-cost scenario. The useful lesson is that prompt length matters. One short question and one agent running many steps, tools, searches, and retries are not the same product from an energy or cooling perspective.

AI Chips Cannot Rely on Ordinary Air Conditioning

Traditional data center racks often operated near 5 to 10 kilowatts. Current AI racks can exceed 100 kilowatts, and ASHRAE says several-hundred-kilowatt racks are emerging. A few rows of modern AI equipment can concentrate the heat that once came from an entire room of ordinary servers.

Uptime Institute says liquid cooling is commonly used above 50 kilowatts per rack or when high-performance equipment has specialized thermal needs in its AI cooling review. Cold plates sit directly on GPUs and other hot components. Liquid carries the heat into pumps and heat exchangers, then the building rejects that heat outdoors.

The cooling construction invoice may include:

• Cold plates, manifolds, hoses, and rack connections
• Coolant distribution units and redundant pumps
• Facility pipework and heat exchangers
• Chillers, cooling towers, or dry coolers
• Water treatment and filtration
• Leak sensors and automatic shutdown controls
• Backup cooling for pump or power failures
• Installation, testing, and commissioning

A low equipment quote may cover only the rack-side hardware. It may leave out the large pipes, building changes, outdoor heat-rejection equipment, electrical work, and backup capacity. This is why a cold-plate price should not be compared with a complete cooling construction budget.

How Much Freshwater Can an AI Data Center Use?

A large data center can consume millions of gallons of water per day, and one large facility can use as much water as a city of 50,000 people, according to the Environmental and Energy Study Institute’s data center resource review.

At 5 million gallons per day, a water-cooled campus would consume:

• 1.825 billion gallons per year
• About 7.6 Olympic swimming pools every day
• About 2,765 Olympic pools per year
• Water on the scale of a small city rather than an office building

Not every AI data center uses that much. A closed-loop design may take a large first fill and reuse the same liquid. An evaporative system continually loses water into the atmosphere. A dry-cooling system can consume very little water on-site but may use more electricity, especially during hot weather.

Water use also exists outside the data center fence. Power stations may consume water while generating the electricity that runs the GPUs and cooling plant. A company can report low direct water use while its electric supply carries a much larger indirect water footprint.

Closed-Loop Cooling With Millions of Gallons

The Stargate campus in Abilene offers a useful edge case. The facility’s closed-loop cooling system was expected to need about 8 million gallons of city water for its first fill. That is less than the water consumed by a large evaporative campus over a full year, but it is still a substantial one-time withdrawal.

The loop will need top-ups, maintenance water, and periodic draining and refilling. Google has said an initial closed-loop fill for one Texas data center building can range from 1.5 million to 2 million gallons. Those water disclosures were reported in the Texas Tribune’s Texas data center guide.

Eight million gallons would fill about 12 Olympic swimming pools. The comparison also shows why “closed loop” and “zero water” are not interchangeable. The water remains in circulation for long periods, but the system still requires an initial supply and later service.

The choice creates a water-and-electricity trade. Evaporative cooling can reduce compressor and fan power but consumes water. Dry cooling protects freshwater supplies but can raise electric demand on the hottest days, when the grid may already be strained.

The Water Project Costs Millions

Municipal water can look cheap beside a billion-dollar GPU order. At an illustrative $5 per 1,000 gallons, consuming 100 million gallons costs only $500,000 for the water itself. Consuming 1 billion gallons costs $5 million.

That calculation leaves out the expensive part. A town may need new mains, pumps, storage, treatment capacity, wastewater systems, reclaimed-water pipes, and emergency supply. The company, utility, taxpayers, or some combination of the three must pay for those assets.

Quincy, Washington, built a water-reuse system serving data center cooling with an identified project cost of $31 million. The system offsets about 138 million gallons per year of demand, according to the EPA’s Quincy reuse case study.

That project works out to roughly $225 of construction cost for each gallon of annual reuse capacity, although the infrastructure lasts many years. It also proves that “using reclaimed water” can require a separate public works project rather than a cheap pipe connected to the nearest treatment plant.

How Much of a Giant AI Campus Is Spent on Cooling?

Epoch AI estimates that a one-gigawatt AI data center requires about $38 billion in upfront capital and around $900 million per year in operating expenses. Servers dominate the budget, but land, buildings, utility work, electrical equipment, and cooling still form a multibillion-dollar support system. Its calculations appear in the one-gigawatt cost model.

Using the cooling construction range in this article, a one-gigawatt cooling plant could cost $1 billion to $1.5 billion. That equals about 2.6% to 3.9% of the modeled $38 billion campus price.

The percentage sounds small only because the AI chips are extraordinarily expensive. The cooling system alone could cost more than many hospitals, skyscrapers, stadiums, or airport terminals.

OpenAI says Stargate is pursuing $500 billion of investment and 10 gigawatts of U.S. AI infrastructure. OpenAI and its partners reported that Stargate had secured a path toward the full commitment through multiple campus developments in its Stargate expansion announcement.

If 10 gigawatts were built using the same broad cooling model, the cooling construction alone could represent $10 billion to $15 billion. That is a scale estimate, not a disclosed Stargate cooling budget.

Three Real AI Campus Cost Cases

Abilene Stargate: Crusoe’s campus is planned for 1.2 gigawatts and about 4 million square feet. Applying the planning model gives a possible cooling construction range of $1.2 billion to $1.8 billion. The disclosed closed-loop system reduces routine freshwater demand compared with a large evaporative design, but the initial fill still requires millions of gallons.

Meta Richland Parish: Meta identifies its Louisiana data center as an investment exceeding $10 billion. The company is also committing more than $300 million to surrounding roads, water, and wastewater infrastructure, according to its Richland Parish project page. That $300 million is not a cooling-system quote. It shows how an AI campus can create infrastructure spending far beyond the server hall.

Google’s global operations: Google reported replenishing 4.5 billion gallons of water in 2024, equal to 64% of its freshwater consumption, through projects intended to restore water to communities and ecosystems. The disclosure appears in Google’s 2025 environmental report. Replenishment does not erase local timing and watershed effects, but it puts a price and project obligation around water use that is missing from a basic utility bill.

A 100 MW AI Campus Cooling Invoice

Consider a 100-megawatt AI campus using liquid-cooled GPU racks.

Cooling construction at $1 million to $1.5 million per megawatt produces a capital budget of $100 million to $150 million.

Assume cooling machinery draws another 20 to 40 megawatts. Running 24 hours per day at 8.62 cents per kilowatt-hour produces an annual cooling-electricity bill of $15.1 million to $30.2 million.

Add an annual maintenance and replacement reserve equal to 4% of cooling construction, or $4 million to $6 million. Add $500,000 to $3 million for water, wastewater, treatment chemicals, and local utility charges, depending on the cooling method and location.

• Cooling construction: $100 million to $150 million
• First-year cooling electricity: $15.1 million to $30.2 million
• Maintenance reserve: $4 million to $6 million
• Water and treatment allowance: $500,000 to $3 million
• First-year cooling total: $119.6 million to $189.2 million
• Cooling cost after construction: $19.6 million to $39.2 million per year
• Daily operating cost: $53,700 to $107,400

The range shows why the advertised hardware price is not the real total. The rack equipment may be only the first layer. Electricity can add tens of millions of dollars every year, while water access and municipal infrastructure can become separate construction projects.

Environmental Costs

The International Energy Agency expects global data center electricity consumption to reach about 945 terawatt-hours by 2030, slightly more than Japan uses today. AI is the main driver of that growth, according to the IEA’s Energy and AI outlook.

The effects extend beyond electricity and freshwater:

• Gas or diesel generation can add carbon dioxide and local air pollution.
• Chillers and refrigeration equipment can leak high-impact refrigerants.
• Cooling towers use treatment chemicals and produce wastewater.
• Concrete, steel, pipes, pumps, and servers carry manufacturing emissions.
• Fans, chillers, generators, and substations can create constant neighborhood noise.
• New transmission, generation, roads, water plants, and sewer systems may shift costs toward utilities and ratepayers.

The site matters. A gallon consumed in a wet region does not carry the same scarcity burden as a gallon withdrawn during drought. A megawatt-hour supplied by renewable or nuclear generation does not carry the same emissions as one supplied by coal or gas. A low-water cooling system may use more electricity, while a low-electricity evaporative system may consume more freshwater.

What We Verified

• Checked the liquid-cooling construction model against KPMG’s published per-megawatt design and retrofit bands.
• Confirmed the electricity calculations using the EIA’s 8.62-cent-per-kilowatt-hour U.S. industrial average.
• Cross-referenced the one-billion-query scenario with the published AI inference energy research linked above.
• Verified the 20% to 40% cooling-energy range against ASHRAE’s 2026 AI data center framework.
• Confirmed the Abilene closed-loop water disclosures through Texas reporting and developer project information.
• Cross-referenced municipal infrastructure costs with EPA, Meta, Google, EESI, and company project disclosures.

Related AI Infrastructure Costs

Cooling is one part of the pressure AI campuses place on utilities. The potential effect on household bills, capacity markets, and substations is covered in AI data center grid costs.

A commercial walk-in cooler installation gives a smaller-scale comparison for refrigeration equipment, electrical upgrades, and heat rejection. Critical facilities also need specialized protection, making the cost of a commercial fire sprinkler system part of the broader building budget.

Municipal water and wastewater expansion can become a separate expense from the cooling equipment. The cost to connect to public sewer shows how excavation, permits, capacity, and utility charges raise the real total. Water-treatment operators may also compare the cost of commercial nanobubbler systems for oxygenation and treatment applications.