How heat and extreme weather threaten AI data centers

Summer 2026 became for data centers what engineers call a “real-time stress test.” France recorded 44.3 °C—the hottest day since 1947—Germany broke its absolute temperature record three times in just a few days, and the Dawn supercomputer at the University of Cambridge went offline. These are not isolated incidents: the AI industry is building computing capacity faster than it can adapt to the climate that these same capacities partly drive.

The scale of the problem in numbers

The main paradox of 2026: the data center industry is experiencing the largest investment boom in history—no less than $750 billion has already been directed toward building new sites, compared to $450 billion the previous year, and Moody’s forecasts that investments will exceed $3 trillion over the next five years. At the same time, a large portion of what is built ends up in climate-risk zones.

First Street, a company specializing in climate risk analysis, estimates that 79% of global data center capacity is exposed to an elevated risk from extreme weather—including heat, floods, and hurricane-force winds. According to Maplecroft analysts, as early as 2025, 56% of large data centers had a high or very high risk rating specifically from extreme heat, and by 2080, this share could rise to 80%.

Geography exacerbates the problem:

North America—about 46% of capacity is located in regions prone to chronic heat and drought.

Europe—this figure reaches approximately 50%.

Asia-Pacific—up to 89% of capacity is under similar climate risks.

Zurich Insurance Group notes that over the past three years, extreme weather events accounted for one-third of all losses in its US data center insurance division.

Why AI specifically exacerbates the problem

A standard server rack dissipates 5–15 kW of heat. A rack with AI accelerators is a different matter entirely: platforms based on NVIDIA H100/H200 or AMD MI300X generate 40–120 kW of heat per rack, and the upcoming Rubin generation, with densities exceeding 100 kW per cabinet, makes air cooling physically unfeasible. For comparison: an NVL72 configuration using Blackwell chips, with a density of 140 kW per rack in a 100 MW data center, supports approximately 650 such racks and around 47,000 GPUs.

Cooling, even without AI workloads, is one of the largest energy consumers in a data center: it accounts for about 40% of the facility’s total electricity usage even at normal air temperatures. During heatwaves, this share rises further—precisely when home and office air conditioners create peak load on the same power grid.

According to the International Energy Agency (IEA), global electricity consumption by data centers is expected to grow from 415 TWh in 2024 to approximately 945 TWh by 2030, with AI accelerators being the main driver of this growth. Some estimates suggest that AI-specific capacity alone could increase from 4 GW in 2024 to 123 GW by 2035—more than 30 times.

Water consumption is also rising: according to AMPAC, training a single large model can consume approximately 185,000 gallons (about 700,000 liters) of water for cooling, and by 2028, global data center water consumption could reach 280 billion liters per year. However, as an analysis by Xylem and Global Water Intelligence showed, direct data center cooling accounts for only about 4% of the additional water footprint of the AI industry by 2050—the far more significant contributions come from electricity generation and chip manufacturing.

How heat has already taken data centers offline

AWS, US-EAST-1 region, May 2026. A thermal incident at the data center caused a cooling system failure, leading servers to overheat and shut down. This affected EC2 and EBS services, forcing engineers to redirect customer traffic to other sites.

Oracle, South London. The company reported that “unreasonably high temperatures” impacted cloud and networking equipment, causing service disruptions for customers throughout the day.

Dawn Supercomputer, Cambridge, June–July 2026. A cooling system failure at the West Cambridge data center, amid record-breaking heat in the UK, led to the shutdown of one of the university’s key computing resources for several days.

France, June 2026. Due to a transformer failure caused by heat, approximately 68,000 households in Finistère were left without electricity, while up to 106,000 consumers nationwide were affected. During the same period, French nuclear power plants were forced to reduce output by 4.1 GW because river water temperatures exceeded the threshold required for full reactor operation.

What experts say

Guillermo Benito, technical director of data center operator Nabiax, points out that global climate risk assessments are heavily skewed toward Asia and the US, making direct extrapolation to Europe incorrect—but the overall trend toward the need for adaptation across all countries remains unchanged, he says.

Tadani, an analyst, describes the core of the problem with maximum clarity: data centers require the most energy precisely when the power grid can supply the least. He cites the example of Turin, where May heatwaves with temperatures around 38 °C created thermal stress on underground cables, leading to repeated power outages in the city.

Christopher Tozzi, a technology analyst at CoreSite, links the rise in heat generation specifically to AI workloads: in his view, the combination of growing energy consumption and massive heat output makes the transition to liquid cooling one of the key trends of the year, as for some AI servers, it is the only cooling method capable of handling their thermal load.

Daniel Mewton, an energy infrastructure expert and partner at the law firm Slaughter and May, noted in an interview with Reuters that a complete data center outage is an “extremely rare” event, whereas cooling system issues occur significantly more often.

What the industry is doing now

Transition to liquid cooling. At densities exceeding 100 kW per rack, air cooling ceases to be physically viable, not just economically. This is why major operators are switching to cold plate cooling directly on the chips—a approach that reduces direct water consumption by 70–90% and simultaneously allows for removing significantly more heat with less air volume.

Nvidia’s water-free design. In June 2026, Nvidia unveiled a reference data center design, DSX, for the next-generation Rubin platform: a closed-loop liquid cooling system with heat rejection via dry or adiabatic coolers instead of evaporative cooling towers allows water consumption at the site to be reduced to near zero. The market reaction was telling: shares of traditional HVAC system manufacturers—Modine, Johnson Controls, Trane Technologies—dropped by 5–7.5% in the moment, which was interpreted as a signal that investors expect a structural shift in cooling approaches across the entire industry.

Redesigning new campuses with future climate in mind, not past. Reports cited by TechRadar explicitly state that investment decisions for new data centers should be based on projected temperature models rather than historical climate data, and water and energy availability for long-term cooling should become as critical a site-selection criterion as land price or grid connectivity.

Using backup generators to relieve grid strain. During the July 2026 heatwave in the US, the Department of Energy mandated that data centers switch to backup generators to free up grid capacity for residential air conditioning—a direct indication of how closely data center energy infrastructure and urban energy systems are intertwined.

New efficiency metrics. In addition to the long-established PUE (Power Usage Effectiveness) metric, the industry began implementing PCE (Power Compute Effectiveness) in 2026—it evaluates not just infrastructure overhead, but how many useful computations each consumed watt delivers, which is especially important for dense AI workloads.