A Look at AI's Carbon Cost: User Impact vs. Sector Reality

#AI #AICarbonFootprint #AIEnergy #ChatGPT #EnvironmentalImpact #TechTransparency #SustainableAI #AIInfrastructure #DataCenters #GreenTech #AIethics #LLMenergy #xAI #Memphis #AirPollution

https://winbuzzer.com/2025/05/07/a-look-at-ais-carbon-cost-user-impact-vs-sector-reality-xcxwbn/

So, although Google uses some renewable energy, it revealed yesterday that its greenhouse gas emissions have risen by 48% over the past five years.

https://www.theguardian.com/technology/article/2024/jul/02/google-ai-emissions

#DeGoogle #LLM #LLMEnergy

On an impulse, I finally did a bit of research on this.

2024-06-20 Researchers run high-performing large language model on the energy needed to power a lightbulb - as with many things, improvements are already underway.

2024-02-16 How much electricity does AI consume?

...which led me to this study (dated after the article, so I don't know when it was actually done): Power Hungry Processing:
Watts Driving the Cost of AI Deployment?

We can see that classification tasks for both images and text are on the lower end of the spectrum in terms of emissions (ranging between 0.002 and 0.007 kWh for 1,000 in-
ferences), whereas generative tasks such as text generation and summarization use, on average, over 10 times more energy for the same number of inferences (around 0.05 kWh for 1,000 inferences), and multimodal tasks such as image captioning and image generation are on the highest end of the spectrum (0.06-2.9 kWh for 1,000 inferences).

Putting this in more domestic units, (Wh = Watt-hours) per query (average per each), and comparing it to more familiar power-uses:

• 2 -7 Wh for classification tasks
• 50 Wh for text-generation
• 60 - 2900 Wh for image captioning and generation

My impression has been when people talk about how much energy GenAI uses, they're usually talking about text-generation queries -- "getting an answer from ChatGPT".

If this study is accurate, one text answer uses about as much power as a non-gaming desktop computer uses in 30-60 minutes. That's not trivial. (Another comparison: the Verge article cites "streaming an hour of Netflix" as using about 800 Wh -- equivalent to about 16 text-queries.)

I think it's important, though, to consider the context. If I'm sitting at my computer asking an LLM a question about how to do something, it's generally because I've already spent a considerable amount of time trying to figure it out on my own.

If the LLM gets the answer even partly right (which has been my experience maybe 80% of the time?), that can save me many hours of further research -- so we need to look at the hours of computer-power I don't use because of that LLM energy-usage, as well as the additional resources I don't personally consume (our home office seems to burn about 300-500 Watts at any given moment, and that's not including the HVAC) while trying to solve that problem.

There are ethical and sustainability issues here, yes (how is it right for someone's image request to be able to command that much energy??) -- but I'm seeing a lot of one-sided takes around this, and I think we need to be more analytical.

Another thing: the enormous chunks of energy often cited in these discussions are often so huge that they must be referring to the energy used to train the models, not for usage of a trained model. From the Verge article:

Training a large language model like GPT-3, for example, is estimated to use just under 1,300 megawatt hours (MWh) of electricity; about as much power as consumed annually by 130 US homes.

If we're going to be concerned about LLMs and power-usage, it's important to understand details like this so we're not fighting the wrong thing.

So, for example, maybe what we should focus on is regulating the training of models -- require them to document their energy usage and release their training information to the public. (...and until this regulation is in place, we could boycott GenAI companies that don't conform to reasonable guidelines.)

We could also require that when a datacenter (of any kind) says it's running on sustainable energy, that has to be capacity which they built, or are connected to directly -- so they're not just draining that energy off the grid so someone else can't use it.

TLDR: Any new tool is going to create new opportunities and new problems. We can work to make the best of those opportunities and solve the problems by understanding what we're dealing with -- or we can misunderstand it and make things worse.

#LLMenergy #GenAI