Re: Environmental impact of Debian's AI usage

Matthias Urlichs <[email protected]> Wed, 19 Aug 2026 16:34:14 +0200

Newsgroups	gmane.linux.debian.devel.vote
Message-ID	<[email protected]>

On 19.08.26 10:27, Matthias Urlichs wrote:
> On 18.08.26 23:21, Zlatan Todoric wrote:
>> (some cities reported already an average increase of 4C
>> because of AI datacenters),
> Source? IMHO that number is highly suspect. 

I asked Claude how fast the air above a 100-MW data center rises, 
assuming sea level, 20 °C and no cross wind. The answer is (roughly) 5…8 
m/s, or Beaufort 4, and it'd be less than 21°C warm at 500m (~1500ft).

If the data center emits a whole whooping gigawatt, we can multiply the 
plume's speed by two, roughly. Bft 5. The temperature is higher and ends 
up at ~23°C at 1km. If we then assume a stiff cross wind of 10 m/s the 
center of the plume would be >300m up at 1 km distance.

It is thus highly unlikely that the warm air ends up on ground level in 
the next-closest city, even if we assume a strong wind going in the 
right (or wrong …) direction. You simply can't keep that amount of hot 
air on the ground. Somewhat similar to a hot-air balloon.

NB: Yeah I asked an AI. So sue me, but I don't have the time, or the 
motivation, or the physics background, to calculate that kind of thing 
myself. Also, if I did it'd most likely require more power to do it than 
Claude Sonnet.

NB2: IMHO this kind of AI answer is rather unlikely to be diluted by 
cross-training. While I agree that self-reinforcing AI sloppiness is a 
major concern, esp. as far as textual output is concerned, that should 
be a fixable problem. If nothing else succeeds (Pangram comes to mind 
…), well, we can simply not train on prose written after 2022. Or 2026, 
given that right now Pangram is still reasonably reliable.

-- 
-- regards
-- 
-- Matthias Urlichs

OpenPGP_signature.asc (application/pgp-signature, 840 B)