🤦 Here we go again: A new, very shaky claim about AI and heat is spreading fast

Last week I wrote about the false claim that AI has an extremely large thirst and water consumption. Now a new claim is spreading, which appears to be at least as shaky.

In a paper, a group of researchers claim that AI data centers (large server facilities) create a local temperature increase of on average 2°C — a “data heat island effect” — up to 10 kilometers from a data center. This would affect nearly 350 million people.

Sounds bad, right?

Several major news outlets jumped on the study:

CNN:
Scientists have found an alarming environmental impact of vast data centers
They are creating “heat islands,” warming the land around them by up to 16 degrees Fahrenheit, and making life hotter for more than 340 million people.

Independent:
Data centers ‘warming the land around them by up to 16 degrees’
New research from the University of Cambridge reveals that the rapid expansion of AI data centers is creating "data heat islands" that significantly warm surrounding environments.

India Today:
AI Data centres creating heat islands, raising temp by 9°C, scientists confirm
In extreme cases, the jump can be as high as the difference between a pleasant Delhi winter morning and a scorching May afternoon.

Sounds really bad, right?

You can turn off the fan and relax, because the study is very likely off track.

Once again, it’s Andy Masley who has worked hard to review the contents of the paper, but I have also examined it closely myself.

The data does not measure what the researchers claim

First of all, it has not been peer-reviewed. That in itself does not mean it is wrong, but for the media above to present it with certainty is directly irresponsible.

The study’s biggest weakness is that the satellite data they use measures surface temperature, not air temperature. That means the temperature of ground, snow, grass, roofs, or tree canopies. Despite this, they write about “atmospheric heat,” microclimates, and effects on human health and welfare — conclusions that require air temperature data, not surface temperature data.

They claim that the temperature increase is caused by waste heat from data centers. That heat is released into the air and would have to be extreme both in temperature and volume to warm the ground surface up to 10 kilometers away.

Instead, there is a much more reasonable explanation for the temperature increase: a data center means new surfaces — buildings with roofs, parking lots, roads, and other surrounding infrastructure.

Imagine walking barefoot on grass on a warm, sunny summer day and then stepping onto an asphalt road.

Since the study excludes data centers in or near cities — to avoid mixing in urban effects — these data centers have likely replaced grass, forest, shrubs, or other natural land. In other words, new surfaces that are warmer.

This is an incredibly obvious objection, which the study does not even attempt to test. They could have compared with other types of buildings that do not produce waste heat, and seen whether the same temperature effect occurs there (for example, a large logistics warehouse, which, like a data center, leads to surrounding infrastructure and buildings). If yes, then the surface temperature increase has nothing to do with it being a data center. If no, then it might indicate that the study is correct.

Collapses under its own implausibility

Another weakness is the sheer magnitude of the effect. Even without calculating, alarm bells should have gone off when considering what would be required for waste heat to warm such a large area.

Imagine stepping out of a restaurant and, as you pass the kitchen, feeling warm air blowing from a fan. That is the kitchen’s “waste heat.” These kinds of fans are supposed to heat an area of 314 square kilometers! Of course, thousands of servers generate far more heat than a kitchen — but 314 square kilometers!

I asked ChatGPT and Gemini to estimate the amount of energy required to heat an area as much as the study claims. Gemini arrives at 4.5 GW. ChatGPT estimated 2–4 GW.

A modern, massive hyperscale data center has a maximum electricity consumption (and thus maximum heat output) of roughly 0.1 to 0.5 GW (and since the study’s data ends in 2024, there are likely none or very few data centers that large in the dataset). The 4.5 GW required by physics to sustain the measured temperature gradient across the landscape is therefore about four to 40 times higher than what the world’s largest data centers can produce.

Stop the spread

There are additional small and large problems with this paper, but for those I refer you to Andy Masley’s review.

Now we need to go into pandemic mode. The vaccine against this study gaining the same traction as the baseless claim about AI and water consumption is facts and scrutiny.

If you see this study continuing to spread uncritically in news media or social media, share this and Andy Masley’s article with them.

Mathias Sundin
Angry Optimist