Mastodawn

RE: https://cosocial.ca/@evan/115580076628853324

This was posted 4 months ago, i.e. forever in LLM time. I would really like to see a fully-worked through analysis of the actual GHG cost of #GenAI in general and for coding applications specifically. Including, obviously, training, data centre infrastructure, silicon fabrication, etc.

The reason: I have trouble reconciling these numbers with the insane volumes of investment capital going into the space.

Tim Hergert Mar 13

@timbray while maybe not complete, Hank Green goes into a good amount of detail on water use when it comes to #AI datacentre rollouts

https://infosec.exchange/@cjust/116059849580662612

Raphael Lullis Mar 13

But then you are moving the goal posts. Even if we could stop all construction for data centers and investment into new models, there are already inference models available on the existing infrastructure which do not take a lot of power.

@timbray @evan @timbray @evan yes, and if only one person was using LLMs for 30 minutes here and there, the usage wouldn’t be a problem. But that’s not the scale, it’s tens of millions 24x7.

And even if no one used them, the datacenters consume that power and water constantly. The servers don’t sleep sans queries. They are always running. They require massive amounts of energy and water simply to exist.

John Ripley Mar 13

@timbray That number for LLM energy is bullshit, and the source in that thread appears to be something Sam Altman blogged, who as we know is someone who never just makes stuff up. (His post reads like ChatGPT voice, which of course it is) There is some severe weaseling be off by at least an order of magnitude—maybe the average query is "What day is it". Non-trivial queries light up 5-10kW of rack and that runs for significantly longer than seconds. This is roughly the same as a moving EV car.

@timbray Is it just me, or do the guys doing the calculations on "entire energy consumption for an LLM programmer versus a human programmer" seem worryingly close to declaring that entire subsections of the population aren't worth the energy consumption, could be replaced by "AI", and would happily suggest culling them as if the only value of a person is lines of code generated 😐

@ibboard yes, that’s the underlying vibe I’m getting too. @timbray

Ed Wiebe Mar 13

@timbray When a new data centre is built it will take resources from the utility or add its own new generation. In the Bad Place (and many other places) this means it's likely adding new fossil fuel consumption infrastructure and increasing carbon emissions. Until recently that country was matching growth in utility generation with renewables: growth was 'green'. That has changed due to their stupidity and due to unregulated new generation like Musk's data centre(s) etc.

It's a bit dumb to say cars are worse because while yes they are bad, in sane countries we are starting to manage that problem well with electrification, changes to the fleet.

New data centres are generally bad, because they are designed to generate slop for the enrichment of the already very harmfully wealthy class and because they require new power generation that is almost certainly not coming from renewables.

Evan Prodromou Mar 13

@timbray my previous job was building greenhouse gas inventories.

Data centres are responsible for about 1% of global greenhouse gas emissions. AI is responsible for about 15% of the data centre emissions, or about 0.15% of global emissions.

People who talk about AI burning up the planet don't spend a lot of time thinking about what's really burning up the planet: fossil fuels for transportation and heating, deforestation, and cattle.

Evan Prodromou Mar 13

@timbray computer usage just isn't as carbon intensive a process as driving a car. Even with hundreds of billions being spent on new data centres, it's not a big part of the global emissions profile.

Evan Prodromou Mar 13

@timbray I think it's a good thing to try to reduce the emissions from data centres and AI. We need to do that with every single human activity that exists.

But I don't think it's fair to say that using AI is a singularly irresponsible activity and morally indefensible on ecological grounds.

Nick Doty Mar 13

@evan @timbray right, we need to reduce emissions in every sector of society. In tech, where we work, AI has led most companies to abandon their promises and increase emissions where they had been on track to reducing them. It's not as significant as transportation or heating or food, but for our industry, it's a huge backwards step, so it makes sense that we would be upset about it.

Evan Prodromou Mar 13

@npdoty @timbray that's a good way to put it!

Cassandra is only carbon now Mar 13

@evan @timbray If we need to do that with every single human activity, why *exempt* AI, when it's badly exacerbating an existing problem?

I think your figures are probably wild underestimates given the corporate secrecy around actual usage, but even so, can we really afford to go 0.15% in the wrong direction on the climate crisis?

lucindacatchlove.bsky.social Mar 13

Yeah, the water usage is a far bigger problem with data centers on the environmental front. It is almost like the energy use drama is intended as a diversion from the real issue, our very real global water crisis. But there's also stuff like Elon illegally using gas generators.

blturner Mar 13

@evan @timbray numbers are pointless without sources?

dr2chase Mar 13

@ben @evan @timbray you can start here to make estimates on bounds: https://www.epa.gov/ghgemissions/inventory-us-greenhouse-gas-emissions-and-sinks

Inventory of U.S. Greenhouse Gas Emissions and Sinks | US EPA

The national greenhouse gas inventory is developed each year to track trends in U.S. emissions and removals. Find emissions by source, economic sector and greenhouse gas.

US EPA

Evan Prodromou Mar 13

https://www.scientificamerican.com/article/ai-will-drive-doubling-of-data-center-energy-demand-by-2030/

"By comparison, data centres consumed 415 TWh in 2024, roughly 1.5% of the world’s total electricity consumption (see ‘Global electricity growth’)."

Electricity accounts for about 1/3 of global GHGs, so data centres are <1% of global GHGs.

"They found that servers for AI accounted for 24% of server electricity demand and 15% of total data centre energy demand in 2024."

AI Will Drive Doubling of Data Center Energy Demand by 2030

Data centers accounted for about 1.5 percent of global electricity consumption in 2024, an amount expected to double by 2030 because of AI use

Scientific American

dr2chase Mar 13

@evan @timbray I’ve started to suspect that every eco-moral-panic around something that isn’t cars/trucks or beef is quietly promoted by big fossil or big ag as a deflection. Remember plastic straws?

Evan Prodromou Mar 13

@dr2chase @timbray we don't have very good public education about GHG emissions, especially from different sectors. The press make a lot of apples-to-oranges comparisons that are hard for people to grasp in isolation, like saying that the fashion industry is responsible for more emissions than the EU.

Evan Prodromou Mar 13

@dr2chase @timbray is that a lot? A little? Does that count clothing made in the EU?

Evan Prodromou Mar 13

@dr2chase @timbray when we slice the pie geographically, but also by activity, and also by economic sector, it can get confusing.

For example, livestock account for about 10-15% of GHG emissions. Part of that is because they fart and burp and shit everywhere, creating a lot of methane, one of the worst greenhouse gases. But another part is because we destroy a lot of forests to make pastureland, taking out an important sink for CO2. Yet another part is processing and transportation.

Evan Prodromou Mar 13

@dr2chase @timbray when you break that down by activity type, you'd put the different contributions in different buckets: livestock cultivation, deforestation, industrial processes, transportation. It's only when you look at the full economic sector that you get the combined numbers.

Evan Prodromou Mar 13

@dr2chase @timbray if you look at the US GHG inventory, it's broken down by activity. So it's hard to see what contribution comes from furniture, from movies, from the wine industry. Their contributions are split across different activities.

dr2chase Mar 13

@evan @timbray the GHG footprint of the (nitrogen) fertilizer used to grow the corn/soy to feed the cattle is also a factor. There is some reuse, e.g. use brewers grains from ethanol (for cars!) manufacture, as cattle feed. And also in the-o-ry we can use electricity to make ammonia, instead of natural gas.

Erin Dalzell (He/Him) 🇨🇦Mar 14

@evan @dr2chase @timbray The Atlantic: https://www.theatlantic.com/magazine/2026/04/ai-data-centers-energy-demands/686064/

Inside the Dirty, Dystopian World of AI Data Centers

The race to power AI is already remaking the physical world.

The Atlantic

Evan Prodromou Mar 14

@emd @dr2chase @timbray hey, did you see my latest poll?

https://cosocial.ca/@evan/116223800384155825

Evan Prodromou (@[email protected])

Have you ever read a greenhouse gas inventory? https://en.wikipedia.org/wiki/Greenhouse_gas_inventory #EvanPoll #poll [ ] Yes [ ] Yes, but... [ ] No, but... [ ] No

CoSocial

dr2chase Mar 15

@evan @emd @timbray too late for me, but yes.

Paul Turnbull has moved.🇨🇦Mar 13

@timbray What is missed in that analysis is that all the emissions from GenAI are new emissions. They aren’t replacing anything. Given the size of the power plants (mostly methane powered) being built to power them they will result in a massive increase in overall carbon emissions.

Evan Prodromou Mar 13

@Chigaze @timbray worldwide? Sure. They might double or triple. As a percentage of total emissions? Still very tiny.

Paul Turnbull has moved.🇨🇦Mar 13

@evan I would question the premise that doubling or tripling our emissions right now is insignificant or even would represent a tiny amount of total emissions over the last 150 years. I'm curious what your numbers are for that.