The expansion of AI data centers is entering an awkward phase. The discussion is no longer just about how much electricity they consume, how much water they need, or where they should connect to the grid. Now, another question has arisen: whether these facilities can raise the local temperature of the areas where they are built.
A study published on arXiv in April 2026, authored by researchers associated with the University of Cambridge and other institutions, addresses exactly that. The work analyzes land surface temperatures via satellite data from 2004 to 2024, cross-referencing these with data center locations, and concludes that after the operation of these facilities begins, there is an observed average increase of 2.07 ºC in the surface temperature over the analyzed area. The range varies from 0.3 ºC to 9.1 ºC, with 95% of cases concentrated between 1.5 ºC and 2.4 ºC.
This figure is striking, but it’s important to interpret it carefully without turning it into a slogan. The study refers to land surface temperature, not necessarily the air temperature perceived at street level by pedestrians. Additionally, it groups data centers built over two decades, from 2004 to 2024. This detail is significant: many of these facilities predate the recent boom in generative AI, so labeling the entire “AI heat island” phenomenon as solely AI-related may be an oversimplification.
What does the study really say
The paper uses MODIS land surface temperature data, reconstructed globally at a higher resolution of 500 meters, combined with the locations of data centers. The authors start with over 11,000 locations, filter out those in densely urbanized areas, and work with 8,472 data centers outside these zones. After cleaning problematic data and outliers, the analysis relies on 6,733 points.
The methodology compares land surface temperature before and after each data center’s commissioning. According to the authors, the increase is not limited to the immediate vicinity of the installation; the effect remains measurable up to 10 kilometers away, though it diminishes with distance. The study indicates a mean monthly increase of 1 ºC could be detected up to about 4.5 kilometers, with the intensity clearly dropping around 7 kilometers.
It also estimates that more than 340 million people could live within the influence radius of these so-called “data heat islands.” This figure is calculated by overlaying 10-kilometer radius zones around data centers with population maps from WorldPop.
| Study Element | Main Data |
|---|---|
| Analysis Period | 2004-2024 |
| Thermal Source | Land surface temperature via satellite |
| Initial Locations | Over 11,000 data centers |
| Centers outside dense urban areas | 8,472 |
| Points after cleaning | 6,733 |
| Observed average increase | 2.07 ºC |
| Range observed | 0.3-9.1 ºC |
| 95th percentile | 1.5-2.4 ºC |
| Estimated reach | Up to 10 km |
| Potentially exposed population | Over 340 million |
These numbers deserve attention. However, not everyone should interpret them straightforwardly.
Methodological nuance: AI heat or land use change?
The first nuance lies in the terminology. The paper refers to “AI data centers” and “AI hyperscalers,” but the time series begins in 2004. At that time, the current deployment of foundational models did not exist, nor was there the massive demand for GPUs for training and inference, or the race to build data centers specifically designed for generative AI. Back then, there were data centers, cloud services, hosting, colocation, networks, and digital services—but not the consumption patterns today associated with AI.
This does not invalidate the study but suggests the headline should be more precise. What the work appears to measure more solidly is the local thermal impact associated with the deployment of data centers in specific locations. Attributing this effect solely to AI may be more debatable.
A second nuance is physical. Part of the temperature increase could come from the energy dissipation of the hardware, but another part may be linked to land use change—asphalt, roofs, open areas, parking lots, reduced vegetation, roads, and materials with low reflectivity. This is the classic urban heat island effect, manifesting through infrastructure.
The authors attempt to minimize this methodological confounding by excluding centers located in dense urban zones, aiming to avoid interference from factories, heating, traffic, residential areas, or industrial activities. Still, the doubt remains: a rural parcel converted into a technical installation doesn’t change only because it is counted; it changes because it becomes urbanized.
It’s also worth remembering that not all data centers are the same. A 15-year-old server room is very different from a modern hyperscale campus; an installation with air cooling differs significantly from one with liquid cooling; and a data center in a dry climate is different from one connected to urban heat networks.
What is important not to overstate
There is a temptation to portray data centers as giant radiators heating entire neighborhoods by themselves. The reality is more complex. All the electricity consumed in a data center ultimately turns into heat, but the local impact depends on how that heat is dissipated, the surface area involved, the facility design, climate, wind, vegetation, insulation, cooling methods, and the surrounding environment.
In other words: heat does exist, but not every increase in thermal readings around a data center can be attributed solely to the servers. In many cases, the dominant factor could be the physical transformation of the land itself.
| Quick-read risk | Necessary nuance |
|---|---|
| “AI raises neighborhood temperatures by 2 ºC” | The study measures land surface temperature, not always air temperature |
| “All are AI data centers” | The data covers 2004-2024, before the boom in generative AI |
| “Heat comes solely from servers” | Asphalt, roofs, vegetation, albedo, and urbanization also influence |
| “Impact extends equally up to 10 km” | The impact diminishes with distance |
| “No regulation exists” | Europe already has increasing reporting and efficiency obligations |
Therefore, methodological critique should not be used to deny the problem but to focus it better.
The real issue: more infrastructure for software that isn’t frugal
Data centers, especially in Europe, are operating within a regulatory framework. The EU’s Energy Efficiency Directive has introduced reporting obligations for data centers with a power demand over 500 kW, and the European Commission is working on a common database and minimum performance standards for the sector.
This does not mean the impact is solved. Instead, the conversation should shift from “the building pollutes” to a more uncomfortable question: what software are we running and with what energy discipline?
Generative AI has normalized a culture of computational abundance. Longer prompts, unnecessarily verbose responses, generalist models used for small tasks, uncleaned storage, automatic retries, agents querying too many tools, and systems compensating for inefficiency with more hardware. The visible container is the data center, but many consumption decisions are made during software design.
The International Energy Agency predicts global data center electricity consumption will double to around 945 TWh by 2030, accounting for nearly 3% of worldwide electricity use. They also indicate that between 2024 and 2030, data center electricity demand will grow at about 15% annually—much faster than other sectors.
This growth isn’t controlled solely by better cooling solutions. It requires developing smaller models, more efficient inference, reuse of computational resources, setting reasonable context limits, carbon-aware scheduling, turning off unnecessary loads, leaner architectures, and sustainability metrics integrated into daily operations.
Hyperscalers also have a transparency debt
Another often-overlooked point is that many organizations use AI and cloud services without precisely knowing the environmental footprint of each client, workload, or service. Major providers publish sustainability reports, but granular data necessary for companies to include this consumption in their own reports remains limited.
To measure their real impact, a company needs data broken down by region, service type, energy consumption, water use, associated emissions, heat reuse, PUE, WUE, and temporal trends. Without this granularity, sustainability remains a rough estimate.
The study itself proposes measures in both software and hardware. It mentions more efficient data selection, compression, model pruning, inference strategies with lower footprints, as well as hardware improvements like more efficient circuitry, dynamic power response, coordinated cooling, and passive techniques like radiative coatings to reduce thermal loads.
What should IT teams do
The practical takeaway for systems administrators, cloud architects, DevOps teams, and infrastructure managers is clear: sustainability must become part of technical KPIs. It can no longer be confined to an annual report.
For years, metrics such as availability, latency, cost, throughput, CPU usage, memory, IOPS, or errors have been monitored. Now, additional KPIs should include transaction efficiency, query energy consumption, inference cost, actual resource utilization, unused stored data, reuse ratios, environmental footprint per environment, and hardware lifespan.
| Area | Traditional KPI | Additional sustainability KPI |
|---|---|---|
| Infrastructure | CPU, RAM, IOPS, availability | Consumption per service and actual utilization |
| AI | Tokens, latency, response quality | Energy per inference and cost per useful task |
| Storage | TB used, performance, backups | Cold data, duplicates, unnecessary retention |
| Software | Response time, errors | Efficiency per operation |
| Cloud | Monthly cost | Cost, emissions, and location per workload |
| Operation | Incidents and SLA adherence | Powered-down capacity, consolidation, reuse |
The biggest mistake would be to use the heat island debate as a simple battle between technophilia and anti-AI sentiment. Digital infrastructure is necessary, and data centers are part of that ecosystem. But that doesn’t justify designing software as if energy, water, land, and hardware were unlimited resources.
Cambridge’s study opens a valuable conversation, even if imperfect. Its findings point to a phenomenon worth monitoring, but its interpretation requires caution: there may be local thermal effects associated with data centers, though not all of that effect is exclusive to AI or solely due to computational heat.
The key question isn’t whether we should halt data center expansion. It’s how to prevent software inefficiency from being masked by adding more land, chips, cooling, and power.
Frequently Asked Questions
Does the study prove that AI heats entire neighborhoods?
It shows an association between data centers and increased land surface temperature in their surroundings. Attributing this solely to AI is more debatable, especially since the period begins in 2004.
What’s the difference between surface temperature and air temperature?
Surface temperature measures the heat of ground, roofs, and visible materials via satellite. It doesn’t always match the temperature humans perceive, but it can influence local microclimates.
Is the effect only from server heat?
Not necessarily. Asphalt, roofs, vegetation loss, land design, and other land use changes play a role.
Are data centers regulated in Europe?
Yes. The EU has introduced reporting obligations for data centers with demand over 500 kW and is working on energy efficiency standards for the sector.
What can technical teams do?
Design more efficient software, measure workload-specific consumption, reduce unnecessary storage, optimize inference, use appropriate models, and incorporate sustainability KPIs into daily operations.
Sources:
