In a study published on Monday by the peer-reviewed journal Patterns, data scientist Alex de Vries-Gao estimated the carbon emissions from electricity used by AI at between 33 million and 80 million metric tons.
That higher figure would put it above last year’s totals for Chile (78m tons), Czechia (78m tons), Romania (71m tons), and New York City (48m tons, including both CO2 and other greenhouse gases).
They use the majority of water during the training phase, but present only usage numbers for people to fall for like you are doing right here.
That is like only counting the time spent by a delivery driver walking packages to a house and ignoring all of the time spent getting it to the delivery company, sorting it, driving it to the airport, flying it to another city, driving it to the distribution center, sorting it again, and then driving it to your house. Sure, if you only count the time delivery people spent walking to houses it isn’t that much time at all!
For sane models, that’s way overstated. Stuff like GLM 4.6 or Kimi K2 is trained on peanuts, and their inference GPU time blows it away.
I have not checked on the latest OpenAI/Grok training cost claims. But if any company is spending tens of millions (or hundreds?) on a single training run… that’s just stupid. It means they’re burning GPUs ridiculously inefficiently, for the sake of keeping up appearances. Llama 4 rather definitively proved that scaling up doesn’t work.
The hype about ever increasing training costs is a grift to get people to give Sam Altman money. He doesn’t need that for the architectures they’re using, and it won’t be long before everyone figures it out and switches to cheaper models for most usage.