LLMS and the Environment

Of all of the legitimate objections to LLM use, I find the environmental costs the weakest.

Depending upon the website where you read, a ChatGPT query uses up to 0.003 kwh of power and 2 ml of water, with similar values for other LLMs. A power user averaging 100 queries per day will consume 0.3 kwh of power and 0.2l of water - probably less. That much power is roughly equivalent to 0.1kg (0.22lb) of CO2 emissions.

Compare that to eating a hamburger. The environmental cost of a single hamburger is 3.1kg (6.8lb) of CO2 emissions and 1600l of water. A portion of tofu with equivalent protein uses less than 5% as much CO2 and water.

So in a 30 day month a ChatGPT user will use 9kwh of power, 6L of water, and 3kg (6.6lb) of CO2 emissions. In the same 30 day month if they eat 15 beef hamburgers instead of 15 soy / pea / bean burgers, they will use 22,800 liters of extra water and 44kg (97lb) of additional C02 emissions.

Now, in fairness, I suspect most people decrying the environmental costs of LLMs aren't aware of the severity of the environmental impact of meat consumption.

But if you've read this post, now you know. Don't trust my numbers, check elsewhere.

Note: a common objection to vegetarian sources of protein is that they aren't as efficiently absorbed in the human body as protein from meat. There are two counters to that argument:

1. The research indicating that protein and more specifically the amino acids in protein from meat is superior to those in vegetable sources has come into question in recent years.
2. The difference in environmental efficiency of growing plant-based proteins and meat-based proteins is profound. Even if protein from beef was three times as good as protein from soy, or nuts, or grains, it would still be environmentally more efficient to grow three times as many plants.

I'm against expanded data center construction anywhere. But if we want to save our environment, reducing meat consumption is the most efficient starting option.