I reckon I’m less persuaded than you about the value of LLMs – though maybe still a hair above average on these forums? I think they have uses, just not ones that are likely to be cost-effective in the foreseeable future. But I found this blog post helpful in putting the energy costs in perspective, and on that point I might be a little more sanguine.
The internet as a whole is energy-hungry, and getting more so. YouTube and TikTok and Netflix and the (non-generative) AI backbone of analytics by which tech companies try to monetize our online time… all those things fill up data centers and greatly outweigh current GenAI energy consumption. And of course all of them are put in the shade by flying in airplanes and driving cars and living in Canada and other ridculously energy-intensive stuff we’ve normalized in modern life.
(Also by the terawatt-hour annual energy consumption of Bitcoin and other proof-of-work blockchain enterprises, all of which is even less useful than ChatGPT. I suspect a lot of people, including me, have a tendency to mentally blur articles about crypto energy use into ones about AI energy use… when the reality is that energy-wise we’re comparing the equivalent of Poland to the equivalent of a couple of Brainerds.)
Of course, if the fever-dreams of companies like OpenAI hold true, then 50Gwh to train its last model (equivalent to a few hundred transatlantic flights) is only the barest beginning, and over the coming decade truly eye-watering amounts of additional compute and energy will be poured into the AGI quest, as they have been into blockchain. But maybe here’s where my greater skepticism about the value of LLMs affects my perspective: I don’t think that’s really going to happen.
I think Ed Zitron (the pleonastic and increasingly splenetic blogger I’ve linked to a few times upthread) is onto something when he points out that Microsoft is semi-quietly cancelling new data center builds and Amazon may be doing the same. OpenAI spent $9 billion to earn $4 billion last year, and the emperor’s nakedness can only be covered up for so long. We’re not in ZIRP-land any more, even before Trump started setting investors’ traditional safe havens on fire. A startup that consistently loses eye-watering amounts of money just doesn’t have the length of runway it did ten years ago.
So my bet is still on the bubble popping. That doesn’t mean the amount of energy we spend on data centers won’t continue to grow, but it’ll keep growing for reasons we mostly can’t hang on GenAI’s door. Not because we’re setting the world on fire in pursuit of the Singularity.
