From the article:

> Training GPT-4 used 50 GWh of energy. Like the 20,000 households point, this number looks ridiculously large if you don’t consider how many people are using ChatGPT.

> Since GPT-4 was trained, it has answered (at minimum) about 200,000,000 prompts per day for about 700 days. Dividing 50GWh by the total prompts, this gives us 0.3Wh per prompt. This means that, at most, including the cost of training raises the energy cost per prompt by 10%, from 10 Google searches to 11. Training doesn’t add much to ChatGPT’s energy cost.

https://andymasley.substack.com/i/162196004/training-an-ai-m...

How does people using it offset the amount of energy used to train it? If I use three hundred pounds of flour learning to make pizza, the subsequent three hundred pounds of flour I use making delicious pizzas doesn't make the first 300 go away. Am I misunderstanding the numbers?

Its not offset, its amortized. Your effective flour / pizza is (300 + 300) / num_pizzas. The total marginal flour expended will go up as you make more pizzas, but the effective cost will actually go down as the upfront cost is amortized over lifetime usage.

You don’t misunderstand the numbers, you misunderstand the point. If you flush your pizzas down the toilet, it’s a waste. If you feed 300 people with it, it’s not, even if you end up using the same amount of ingredients.

Sure, it's a value calculation.

If you're able to serve delicious pizzas afterwards, it was worth wasting the first kg (you might call it an investment).

If you're able to bring value to millions of users, it was worth to invest a few GWh into training.

You might disagree on the usefulness. I think, you shouldn't have wasted a kg of flour because I won't ever eat your pizzas anyway. But many (you, your guests, ChatGPT users) might think it was worth it.

It doesn't make it go away. Using your analogy - if you used 300lb to learn and then only made 10 lb of pizza after that, it would be a pretty poor use of resources.

If you instead went on to produce millions of pizzas for people and 30,000lb of flour, that 300lb you used to learn looks like a pretty reasonable investment.

The cost of training has to be normalized by the number of users (or queries) that rely on that training.

If you use 300 lbs of flour to learn, and 300 lbs of flour to make 300 pizzas, then the total flour cost is 2 lbs of flour per pizza.

For context thats's about equivalent to 100 transatlantic flights

To give you an idea of how many models are being trained, and how the energy costs continue to increase. https://epoch.ai/data/notable-ai-models

I mean, I guess advances could plateau and we stop spending exponentially more energy year after year...

I'm not opining on whether it's a good idea (I doubt we ever voluntarily consume less as a species), but data centres use a lot of energy and billions are being spent building them. https://www.technologyreview.com/2024/09/26/1104516/three-mi...

See my other comment here. One AI training run does not exist in a vacuum. Do you think they built billions of dollars in datacenters full of computer power just to let it sit idle?

That argument is just as bad. The training runs are equivilant to a few days to a couple of weeks of flights from New York to London. No one thinks we are going to save the environment by stopping a single plane route for 2 weeks.

One AI training run does not exist in a vacuum.

https://washingtonstatestandard.com/2025/04/11/as-demand-for...

> instead of the power usage of training models.

To make it worse, the model training cost only refers to the cost of the training itself. The externalities - everyone else being forced to drastically upscale their compute power because scraper blocking isn't foolproof - are, as usual for hypercapitalism, conveniently ignored.

AI training in its current form is unsustainable, I'd go as far to say it's a threat for the open and decentralized Internet as you have all but zero chance of standing alone against the flood of training scraper bots and more and more control gets ceded to Cloudflare et al.