Thank you for spending the time to write this when most sensible people would just go "It's obviously stupid" and ignore it. If it weren't for people like you, it would go unchallenged.
When I was originally reading the paper, I was sketched out that they didn't include any specific example satellite photo example comparisons or anything.
For the pixel areas, it would be better to write 0.5km × 0.5km or 250,000m² etc. instead of .5km² and 500m². (You can get away with 1km²!) This doesn't affect your conclusions.
A fine example of which would be Google Mayes County, Oklahoma. It is pretty big but built inside an old Gatorade bottling plant, which was also big, and in the middle of nowhere with bare ground on all sides for miles.
Tried to analyze the heating effects of the data center there by downloading similar NASA data but it seems mostly random without any visible change in temp at all
I wasn’t around for the initial water panic arguments, but I hope to use the knowledge I gained from this to challenge my anti ai friends. Hopefully this disinfo doesn’t spread as easily, thanks for writing this.
Hah, I saw the headline this morning, read that the study hadn't been peer reviewed yet, and filed it under "check back in when it's been peer reviewed". Thanks for your service.
That's not so, but I can see how that may have been misleading. I thought about fact checking the article but didn't bother until it had at least been peer reviewed (which usually means reviewed prior to appearing in a journal, just to define terminology - and some journals have higher standards than others) - Andy fact checked it instead and satisfied my curiosity.
My only take is that the cover images should be different.* I happily pay ≥100 for my Claude subscription, but there is something in me that found the cover image displeasing, and I think that feeling is stronger amongst people who should be reading your essays the most.
*Different = whatever Seb Krier does to generate images or good old non-generated pics
The more I look at that graph, the more I'm confused. I can't figure out how they got the results they did *regardless* of what you assume is causing the warming.
Note that the "temperature difference" is actually the difference between that month and a 5-year moving average (as explained in the text). So if the change suddenly occurred in one month and the temp was flat after that, then you would see a jump, and then a slow decay of the difference as more of the post-change section was in the moving average.
And I don't understand how the top and bottom lines are so flat either - they are the *single* highest and lowest data points, so one would think they would jump around a lot just due to natural variation (I guess I'm not that familiar with LST data - I wonder how much LST just varies randomly day-to-day or month-to-month? Obviously air temperature varies a lot, that's how we have warm and cold weather, but maybe surface temperature varies a lot less?
Just taking a glance at the MODIS data here: https://neo.gsfc.nasa.gov/view.php?datasetId=MOD_LSTD_M&year=2025 and visually comparing is with e.g. the same month in 2024, it does seem remarkably consistent, so maybe that isn't *that* surprising? But I would really like to see their actual data. I wonder if some part of their preprocessing is doing something weird to it.
(And also, the fact that the "date when the data center became operational" might be erroneous/inconsistent just makes it an even bigger mystery. It seems like that would just cause the trend to be more smeared out in the x-axis direction, so it would be less of a sharp jump?)
The main weakness in this paper is that it lacks a control or comparator. Without it, we cannot distinguish between the heating caused by normal urbanization vs. the heating caused by data centers specifically. That much is clear.
Coming from academia, I would bet that they actually did do it (i.e., include a control / comparator) but that it didn't fit neatly into their manuscript. I suspect that is why they put in on arxiv first instead of submitting it for peer review. The strategy surrounding publication and scientific communication is a separate conversation, however.
Their approach is kinda neat, and I think it would be cool to apply it more broadly: I'd be interested in a study that includes i) areas that remain rural overtime, ii) areas that begin rural and are then urbanized, and iii) areas that begin rural but receive a datacenter. That way, you could tease out the actual effect size of datacenters on LST heating, which would be far more compelling.
Since their methods are pretty straightforward, and since they're using publicly available data, I don't think it would be hard for someone to try to replicate their results and include additional comparators to verify.
Part of data centers' heat effects are, in some cases, because they're using on-site power sources. This NBC story on the Memphis-area AI data center shows what I mean.
The heat exhaust from power generation doesn’t show up on the type of satellite data this study collects, though I agree it’s definitely a reasonable concern otherwise. imo the air pollution from turbines like the ones at xAI is the real danger here way more than the heat.
Waste heat warming land measured by satellites is a weird proxy. Just measure how many watts the data centres pull! I can't think of any good reason one would prefer anything other than direct power use
i dont think i understand how the building construction itself could result in such a sudden (<1 month) change in temperature? a large building takes more than a month to construct, right?
I suspect this might actually be because the main thing the satellite is picking up, the roof, is one of the last things to go up, and they're setting the baseline a few months before the DC went up when it's still being constructed, so the baseline is already where there's building present, but the roof causes the jump?
We're already discussing this in the other thread - but I want to be clear, this is completely wrong from a construction timing and building science standpoint. The baseline already exists before the roof goes on. If anything, the roof will reduce the heat signature when it goes on.
> the red line is the average, and the error bars represent the 95th percentile bounds, meaning the vast majority of measurements fell within those much smaller bounds
the paper itself seems a bit unclear about what the range represents, but if those are error bars for the mean, then all datapoints could in theory be outside of that range.
From the paper: "The shaded areas show the interval between the maximum and minimum value of LST increase that has been recorded across the considered AI hyperscalers. Finally, the bar across the average line identifies the limit of the 95th percentile of the distribution we compute."
You and Claude mention "dark-roofed" buildings, but *are* they dark? When I see large commercial buildings, the roofs are generally white.
Given how good the LLMs can be at arguing a case (any case), it might be useful to prompt it the other way, or have it critique its critique. Only prompting it one way (to find flaws in the paper) may leave you with some confirmation bias.
It would be interesting to try to reproduce but substitute large warehouses or something for data centers and see how large the effect is. If you get 98% of the effect (or whatever it is) from a warehouse, that's a simple story to tell to substantiate the "it's just construction" explanation.
This whole thread is a very interesting read as someone who is interested in AI and tech but has a professional background in buildings and building science.
I think the bigger effect is removal of vegetation - plants move nutrients by forcing water evaporation (which is why they need so much water) and this process converts a lot of heat into vapor.
Thank you for spending the time to write this when most sensible people would just go "It's obviously stupid" and ignore it. If it weren't for people like you, it would go unchallenged.
The before/after pics of the Bajio industrial park from Google Earth are crazy and like 2% of it is data centers and the other 98% of it is car parts.
When I was originally reading the paper, I was sketched out that they didn't include any specific example satellite photo example comparisons or anything.
For the pixel areas, it would be better to write 0.5km × 0.5km or 250,000m² etc. instead of .5km² and 500m². (You can get away with 1km²!) This doesn't affect your conclusions.
Was debating this, thanks!
Yeah, as written it is confusing
This is confusing in the paper itself. It says
"We used a reconstructed MODIS LST dataset (produced
by NASA) acquired worldwide from 2004 to 2024 over an enhanced 500m resolution grid"
The lowest resolution public MODIS LST is 1 km. (Table 1 here lpdaac.usgs.gov/documents/715/MOD11_User_Guide_V61.pdf)
They link to this paper (https://www.researchgate.net/publication/262974152_Surface_Temperatures_at_the_Continental_Scale_Tracking_Changes_with_Remote_Sensing_at_Unprecedented_Detail) which talks about a spatial resolution of 250m. I'm guessing they are using a statistical modeling technique that was created for that paper, but they wanted smaller error bars so they went bigger?
it's like they have never heard of the urban heat island!
An interesting control would have been to find a brownfield data center that has been built on land that wasn't farmland/woodland before hand.
A fine example of which would be Google Mayes County, Oklahoma. It is pretty big but built inside an old Gatorade bottling plant, which was also big, and in the middle of nowhere with bare ground on all sides for miles.
Tried to analyze the heating effects of the data center there by downloading similar NASA data but it seems mostly random without any visible change in temp at all
Which seems to confirm your initial hypothesis
They specifically mention the urban heat island in the first line of the intro to their paper. The data heat island effect is named after it.
I wasn’t around for the initial water panic arguments, but I hope to use the knowledge I gained from this to challenge my anti ai friends. Hopefully this disinfo doesn’t spread as easily, thanks for writing this.
Hah, I saw the headline this morning, read that the study hadn't been peer reviewed yet, and filed it under "check back in when it's been peer reviewed". Thanks for your service.
This isn't what peer reviewed means. This is an opinion piece propped up by chatbots themselves.
He means the original study?
He seems to consider your blog post here a peer review.
That's not so, but I can see how that may have been misleading. I thought about fact checking the article but didn't bother until it had at least been peer reviewed (which usually means reviewed prior to appearing in a journal, just to define terminology - and some journals have higher standards than others) - Andy fact checked it instead and satisfied my curiosity.
Andy wrote an opinion piece on it and showed pretty basic misunderstanding of building science and construction in general.
Fantastic as always!
My only take is that the cover images should be different.* I happily pay ≥100 for my Claude subscription, but there is something in me that found the cover image displeasing, and I think that feeling is stronger amongst people who should be reading your essays the most.
*Different = whatever Seb Krier does to generate images or good old non-generated pics
Made a temporary art change, will find something better soon
Appreciate you writing this down!
(nitpick) I would not conclude "don't build", but "build smart".
I did some research on urban heating in college. Our results back then
- painting the roof white, to reflect sunlight more, helps a ton already
- green roofs help even more. They store rain water in the soil which then evaporates later and cools down the building.
PS: White roofs also lower air condition / cooling costs, making them worthwhile from a purely economic perspective too.
Great piece. I really hope this becomes the common understanding as opposed to the nonsense in the paper. I think it will
The more I look at that graph, the more I'm confused. I can't figure out how they got the results they did *regardless* of what you assume is causing the warming.
Note that the "temperature difference" is actually the difference between that month and a 5-year moving average (as explained in the text). So if the change suddenly occurred in one month and the temp was flat after that, then you would see a jump, and then a slow decay of the difference as more of the post-change section was in the moving average.
And I don't understand how the top and bottom lines are so flat either - they are the *single* highest and lowest data points, so one would think they would jump around a lot just due to natural variation (I guess I'm not that familiar with LST data - I wonder how much LST just varies randomly day-to-day or month-to-month? Obviously air temperature varies a lot, that's how we have warm and cold weather, but maybe surface temperature varies a lot less?
Just taking a glance at the MODIS data here: https://neo.gsfc.nasa.gov/view.php?datasetId=MOD_LSTD_M&year=2025 and visually comparing is with e.g. the same month in 2024, it does seem remarkably consistent, so maybe that isn't *that* surprising? But I would really like to see their actual data. I wonder if some part of their preprocessing is doing something weird to it.
(And also, the fact that the "date when the data center became operational" might be erroneous/inconsistent just makes it an even bigger mystery. It seems like that would just cause the trend to be more smeared out in the x-axis direction, so it would be less of a sharp jump?)
The main weakness in this paper is that it lacks a control or comparator. Without it, we cannot distinguish between the heating caused by normal urbanization vs. the heating caused by data centers specifically. That much is clear.
Coming from academia, I would bet that they actually did do it (i.e., include a control / comparator) but that it didn't fit neatly into their manuscript. I suspect that is why they put in on arxiv first instead of submitting it for peer review. The strategy surrounding publication and scientific communication is a separate conversation, however.
Their approach is kinda neat, and I think it would be cool to apply it more broadly: I'd be interested in a study that includes i) areas that remain rural overtime, ii) areas that begin rural and are then urbanized, and iii) areas that begin rural but receive a datacenter. That way, you could tease out the actual effect size of datacenters on LST heating, which would be far more compelling.
Since their methods are pretty straightforward, and since they're using publicly available data, I don't think it would be hard for someone to try to replicate their results and include additional comparators to verify.
Part of data centers' heat effects are, in some cases, because they're using on-site power sources. This NBC story on the Memphis-area AI data center shows what I mean.
https://youtu.be/C8rU4dv2w8Q?si=FOXKM__xEbtufde7
The heat exhaust from power generation doesn’t show up on the type of satellite data this study collects, though I agree it’s definitely a reasonable concern otherwise. imo the air pollution from turbines like the ones at xAI is the real danger here way more than the heat.
Waste heat warming land measured by satellites is a weird proxy. Just measure how many watts the data centres pull! I can't think of any good reason one would prefer anything other than direct power use
i dont think i understand how the building construction itself could result in such a sudden (<1 month) change in temperature? a large building takes more than a month to construct, right?
I suspect this might actually be because the main thing the satellite is picking up, the roof, is one of the last things to go up, and they're setting the baseline a few months before the DC went up when it's still being constructed, so the baseline is already where there's building present, but the roof causes the jump?
We're already discussing this in the other thread - but I want to be clear, this is completely wrong from a construction timing and building science standpoint. The baseline already exists before the roof goes on. If anything, the roof will reduce the heat signature when it goes on.
nitpick: i dont think this sentence is true:
> the red line is the average, and the error bars represent the 95th percentile bounds, meaning the vast majority of measurements fell within those much smaller bounds
the paper itself seems a bit unclear about what the range represents, but if those are error bars for the mean, then all datapoints could in theory be outside of that range.
From the paper: "The shaded areas show the interval between the maximum and minimum value of LST increase that has been recorded across the considered AI hyperscalers. Finally, the bar across the average line identifies the limit of the 95th percentile of the distribution we compute."
You and Claude mention "dark-roofed" buildings, but *are* they dark? When I see large commercial buildings, the roofs are generally white.
Given how good the LLMs can be at arguing a case (any case), it might be useful to prompt it the other way, or have it critique its critique. Only prompting it one way (to find flaws in the paper) may leave you with some confirmation bias.
It would be interesting to try to reproduce but substitute large warehouses or something for data centers and see how large the effect is. If you get 98% of the effect (or whatever it is) from a warehouse, that's a simple story to tell to substantiate the "it's just construction" explanation.
This whole thread is a very interesting read as someone who is interested in AI and tech but has a professional background in buildings and building science.
This is the exact intersection I need!
I think the bigger effect is removal of vegetation - plants move nutrients by forcing water evaporation (which is why they need so much water) and this process converts a lot of heat into vapor.