The Invisible Heat Machine: How Much Warmth Does an AI Data Centre Actually Produce?

 Author: Muhammad Waqar Khan

Drive through Ashburn, Virginia, on a grey winter morning, and you will notice something odd. Tucked between suburban shopping centres and empty industrial lots sit vast, windowless concrete buildings — no signage, no visitors, no obvious purpose. Look closely, and you will see massive cooling units bolted to the exterior walls, humming constantly as they push warm air into the cold morning sky. These are data centres, and they are the physical engine rooms behind every AI chatbot response, every cloud backup, every streamed film. They never sleep, and they never cool down.

As artificial intelligence has accelerated from a niche research topic into a commercial arms race, the data centres that power it have grown dramatically in size, number, and appetite. The heat they generate is no longer just an engineering inconvenience — it has become an environmental, political, and economic challenge that entire cities and electricity grids are now scrambling to manage.

The Invisible Heat Machine

Why Heat Is the Unavoidable Byproduct of Computing

Before examining the scale of the problem, it is helpful to understand why heat and computing are inextricably linked.

Every time a processor — whether a traditional CPU or a modern AI-focused GPU — performs a calculation, it draws electricity. Most of that electrical energy does not go into useful computation; physics dictates that it is converted into heat. A large portion of what a data centre pays for in electricity ends up radiating off circuit boards and server racks.

Traditional data centre chips, the standard CPUs of a decade ago, ran at roughly 150 to 200 watts per chip. When AI became serious business, the industry shifted to GPUs, which are far better suited to the parallel calculations that AI models require. Early AI GPUs ran at around 400 watts. By 2023, state-of-the-art AI GPUs were running at 700 watts per chip. The next generation of chips arriving in 2024 and 2025 pushes that figure to around 1,200 watts per chip — roughly the power draw of a small space heater, except these chips are packed eight to a blade, ten blades to a rack, and hundreds of racks to a building.

The result is staggering power density. As of the mid-2020s, individual server racks in AI data centres commonly support power requirements of 20 kilowatts or more. The average density is projected to climb from around 36 kilowatts per rack in 2023 to 50 kilowatts per rack by 2027. All of that electricity becomes heat that must be actively removed from the building, or the hardware will throttle, fail, or catch fire.

How Much Heat Does a Modern AI Data Centre Actually Produce?

The short answer: an enormous amount, and it scales with the size of the facility.

A standard hyperscale data centre — the kind built by Amazon, Google, Meta, or Microsoft to support cloud computing and AI at scale — typically requires between 100 and 300 megawatts of electricity to operate continuously. According to IBM, these facilities house at least 5,000 servers and occupy a minimum of around 10,000 square feet, though most modern hyperscale sites are dramatically larger.

To put that electricity consumption in context: 100 megawatts is enough to power roughly 80,000 to 100,000 average homes. A 300-megawatt facility exceeds the peak power draw of some mid-sized cities. Every watt drawn eventually becomes waste heat that must be expelled.

Roughly half or more of a data centre's total electricity demand goes directly to running IT equipment. Much of the remainder goes toward cooling that equipment, making cooling one of the single largest operational costs in the industry.

The heat output is measurable not just at the building level but at a landscape level. A peer-reviewed study using NASA satellite data tracked land surface temperatures globally from 2004 to 2024 and cross-referenced the measurements with over 11,000 AI data centre locations worldwide. The findings were striking: temperature increases near data centres ranged from 0.3 degrees Celsius to 9.1 degrees Celsius compared to the surrounding baseline. The effect was observable from orbit.

At the scale of the entire industry, the numbers are even more sobering. The International Energy Agency estimated that data centres consumed approximately 415 terawatt-hours of electricity in 2024 — about 1.5 percent of the entire world's electricity supply. That consumption has grown at roughly 15 percent annually over the past five years and is projected to nearly double, reaching around 945 terawatt-hours by 2030. The IEA's analysis further notes that AI-specific server electricity consumption is growing at roughly 30 percent per year in its base case scenario.

Where Are AI Data Centres Built — And Why?

Location is not accidental. The sites chosen for major data centres reflect a careful calculation of four factors: affordable and reliable electricity, high-bandwidth fibre network access, available land, and — increasingly — access to water for cooling.

Northern Virginia: The World's Data Centre Capital

No single location concentrates more data centre infrastructure than Loudoun County, Virginia, particularly the town of Ashburn. The area is so dense with facilities that it carries the unofficial nickname "Data Center Alley." Northern Virginia accounts for nearly 7 gigawatts of installed data centre capacity and is home to more than 600 facilities, with around 5 percent of the world's total capacity currently under construction in the region.

The story of how this happened is partly historical. In 1998, Equinix opened one of the earliest commercial data centres in the area to serve companies like America Online. The region had deep fibre infrastructure, favourable tax incentives, and a political environment friendly to large commercial investments. Amazon opened its first data centre there in 2006. The momentum built from there. Virginia data centres served by Dominion Energy had a combined electricity demand of 3,583 megawatts in 2024 — nearly seven times higher than in 2013.

That concentration is now creating problems. Primary North American data centre markets hit a record-low vacancy rate of 1.6 percent in the first half of 2025, with Northern Virginia among the tightest. Pre-leasing sits at around 73 percent, meaning most capacity is spoken for years before it is even built.

The American Midwest and Beyond

With Virginia effectively at capacity, developers are looking elsewhere. Other significant US hubs include Chicago, Dallas, Phoenix, Silicon Valley, and — increasingly — Omaha, Nebraska. A Cornell University study published in late 2025 argued that the concentration of data centres in Virginia, California, and the southwest is creating unnecessary strain on water supplies and local grids, and that the Midwest and Plains states — Nebraska, South Dakota, Montana, and parts of Texas — offer better combinations of renewable energy, cooler ambient temperatures, and less-stressed infrastructure. Whether the industry follows that guidance remains an open question, but the economic logic is real.

Europe's Strategic Push Northward

European data centre development has increasingly looked toward Scandinavia, driven by two compelling advantages: cold ambient temperatures that naturally reduce cooling costs, and abundant, cheap renewable energy from hydropower.

Norway is emerging as a particularly significant destination. The region north of the Arctic Circle has surplus hydropower capacity and consistently cold air temperatures, both of which drastically reduce the cost and energy required to cool servers. One AI infrastructure company, Nscale, secured electricity in northern Norway at three to four cents per unit — far below the European average of ten cents. OpenAI announced plans for Stargate Norway, its first European data centre, near Narvik in northern Norway, designed to deliver 230 megawatts of capacity running on 100,000 Nvidia GPUs and powered entirely by renewable energy.

In Denmark, Facebook's two large data centres in Odense are connected to the local district heating network, channelling waste heat into homes. Norway-based atNorth has reached an agreement to use excess heat from its Danish facility to warm over 8,000 homes by 2028.

Other major European data centre hubs include Amsterdam, Frankfurt, Dublin, and London, though each faces growing pressure from regulators and grid operators concerned about the speed of expansion.

The Invisible Heat Machine

The Cooling Challenge: How Data Centres Manage the Heat

Given the scale of heat production, cooling is one of the most technically demanding and expensive aspects of running a modern data centre.

Traditional air cooling — essentially, very large and sophisticated air conditioning systems — has been the industry standard for decades. Cold air is pumped in to absorb heat from server racks, warmed air is expelled, and refrigeration equipment removes the heat from the expelled air before cycling it back. This works reasonably well for lower-density installations.

AI workloads, with their dramatically higher rack densities and per-chip heat output, are pushing air cooling toward its limits. The industry has responded with a shift toward liquid cooling — circulating coolant directly to the chip or through liquid-cooled plates attached to server components. Direct-to-chip and immersion cooling (where hardware is submerged in thermally conductive but electrically inert liquid) can remove heat far more efficiently than air. The liquid cooling market is growing rapidly, with analysts sizing it at approximately $4.68 billion in 2025 and projecting a compound annual growth rate above 19 percent through 2034.

Liquid cooling also enables something that air cooling does not: practical waste heat recovery. When coolant circulates through a facility and absorbs heat from servers, the warmed liquid can be piped to a heat exchanger and used to warm buildings or water rather than simply dumped into the atmosphere.

Turning Waste Heat Into Something Useful

The idea of harvesting data centre heat for productive use has moved from a theoretical curiosity to a live industry conversation.

Approximately 40 percent of the energy a data centre consumes, and nearly all of the water, goes directly to server cooling, with most of that energy ultimately converting to heat. Some of that heat is at temperatures high enough to be genuinely useful for district heating — the systems of underground pipes that carry hot water to warm homes and commercial buildings common across Northern Europe.

Research published in a peer-reviewed study found that integrating data centre waste heat into district heating systems could reduce centralized heat production by 32 percent and cut carbon emissions by up to 50 percent annually, even if waste heat supplied only 20 percent of total demand. Microsoft has estimated that heat energy recovery factors of up to 69 percent in winter and 86 percent in summer are achievable, depending on the cooling system and ambient temperatures.

The World Economic Forum has estimated the data centre heating market could eventually be worth $2.5 billion. Scientists have proposed even more ambitious applications: a 2025 study in the journal Energy and Environmental Science suggested that, with advanced cooling and heat management, data centre waste heat could theoretically power direct air capture of carbon dioxide, potentially removing 50 to 1,000 megatonnes of CO2 annually, while also producing clean water as a byproduct. That vision is still far from commercial reality, but it illustrates how the industry's relationship with heat is beginning to shift from pure liability toward potential asset.

The Energy Source Question

How a data centre is powered matters as much as how much power it uses. A facility running on coal-fired electricity has a very different environmental footprint from one running on hydropower.

The leading tech companies have made commitments to clean energy, though the path to achieving those commitments is complicated. Amazon leads corporate renewable energy procurement with over 20 gigawatts of capacity across more than 500 projects globally. Google and Microsoft have made similar commitments. The technology sector accounted for around 40 percent of all corporate power purchase agreements for renewables signed in 2025.

The nuclear option has also re-entered the conversation. In 2024, Microsoft signed a 20-year agreement with Constellation Energy to restart a reactor at the Three Mile Island nuclear plant in Pennsylvania — the first restart of a closed US nuclear plant. The pipeline of agreements between data centre operators and small modular reactor projects has grown from 25 gigawatts in late 2024 to 45 gigawatts by mid-2026. These are small factory-built nuclear units that could eventually provide reliable, low-carbon power directly adjacent to major data centre campuses.

Not all expansion is green, however. Some developers, constrained by slow grid connection timelines, are building data centres adjacent to natural gas power plants that operate entirely off the main grid. The tension between ambitious sustainability pledges and the practical urgency of connecting AI infrastructure is real and ongoing.

Common Misconceptions Worth Addressing

One persistent misconception is that AI queries are essentially free from an energy standpoint — that asking a chatbot a question is no different from a Google search. The reality is more nuanced. Training large AI models is extraordinarily energy-intensive: training GPT-4 alone reportedly required around 30 megawatts of sustained power. Running finished AI models (called inference) is less intense per query but happens billions of times a day. The cumulative demand is substantial and growing.

A second misconception is that efficiency improvements will solve the problem on their own. GPU performance per watt has indeed improved dramatically — by some estimates, a 4,000-fold improvement over the past decade, according to Nvidia. But the pace of AI adoption and the sheer scale of investment ($580 billion globally in AI-focused data centre infrastructure in 2025 alone) is outpacing efficiency gains. Total electricity demand from data centres rose 17 percent in 2025, even as hardware became more efficient.

A third assumption worth challenging is that data centres are always located where they make the most environmental sense. As the Cornell study noted, many facilities are concentrated in locations where legacy infrastructure exists but where water stress and grid strain are increasingly serious concerns.

Frequently Asked Questions

How hot does the inside of a data centre get?

Without active cooling, server rooms would quickly reach temperatures that damage hardware — chips can generate enough heat to exceed 100 degrees Celsius at the component level. Operators maintain strict temperature and humidity ranges, typically keeping the inlet air to servers at around 20 to 27 degrees Celsius. The exhaust side of a rack, where hot air leaves, can exceed 45 degrees Celsius even under managed conditions.

Do data centres cause local warming?

Research using satellite temperature data found measurable surface temperature increases near data centres ranging from 0.3 to 9.1 degrees Celsius compared to baseline measurements. The effect is real but highly variable depending on facility size, cooling method, and local geography.

Which country has the most data centres?

The United States leads by a significant margin. As of 2025, the US has approximately 5,400 data centres, accounting for roughly 45 percent of the world's total. The US has around ten times more data centres than China and most European countries.

Why are so many data centres in Virginia?

Northern Virginia has historically offered a combination of deep fibre-optic infrastructure, state and local tax incentives, reliable electricity from Dominion Energy, and proximity to major internet exchange points on the US East Coast. Those advantages attracted early investments, which attracted more investment, creating a self-reinforcing concentration that now has serious capacity constraints.

Can data centre heat actually warm homes?

Yes, and this is already happening in several Nordic countries. Facebook's Odense data centres in Denmark feed waste heat into the local district heating network. The key constraints are geography — the data centre must be near enough to the buildings being heated to make pipe infrastructure practical — and temperature, since some district heating systems require higher temperatures than data centre waste heat naturally provides, necessitating heat pumps to upgrade the temperature.

The Bigger Picture

The relationship between AI and heat is one of the defining infrastructure challenges of the current decade. The electricity demand is real, the heat output is measurable from space, and the concentration of facilities in a handful of already-stressed locations creates identifiable risks to local grids, water supplies, and communities.

At the same time, the pace of innovation in cooling technology, renewable energy procurement, and waste heat recovery is genuine. The industry moving from viewing heat as a waste product to viewing it as a potential resource — for warming homes, capturing carbon, or producing water — represents a meaningful shift in thinking, even if practical implementation is still in its early stages.

What matters most is that the people making decisions about where and how to build these facilities — policymakers, utility regulators, city planners, and corporate infrastructure teams — treat the heat equation as a first-order concern rather than an afterthought. The windowless buildings humming along the Dulles Toll Road are not going away. The question is whether their heat goes up in the air or gets put to work.

Author