God I hate these fake science calculation posts.

His prompts were each using 2.9 million token. That's a massive amount, he was basically purposefully using complex tasks that needed massive amounts of data parsing. We're talking about like 1 percent of users that are using 2.9 million per prompt.

The method used to calculate doesn't include batching. These companies aren't running one request per gpu here. Even the authors of the method admit it over estimates by 4 to 20x.

I find this analysis is a useful starting point for discussion, although there are plenty of details one might adjust.
Personally I'm using (inter alia) claude to help me refine an interactive climate model (example here - although that's last year's version pre-ai-help ). So I care about these things.

As my own life also has an energy cost - even just sitting at a desk with computers and some heat light and food. I reckoned by my own crude calculations that my 'human' energy cost per hour was considerably higher than that of my AI assistant, which certainly helps me progress faster, so the net effect was less energy per 'task done', meanwhile we don't have infinite time to solve such problems. I'm only using claude within the limits of a pro subscription, and achieve that with tough claude_md instructions - not to go digging rabbit holes without consulting me. Sometimes it analyses and fixes autonomously and efficiently, but you have to keep alert - sometimes I interrupt and say no there's a simpler way, and draft better algorithms / structures myself. Also I use scala whose sophisticated (non-ai) tooling constrains mistakes and its mcp/lsp makes searching and refactoring across a large codebase much more efficient than claude's normal grok by subagents. Combine tools carefully, not brute force.

Evidently a big unclarified issue is the energy cost of training these things. But we don't need so much more training - for my purposes they are already good enough. The frequent new releases are about scary headlines to pump the IPOs. If this race could slow down, we could just learn to use what we've got more efficiently. In the general public discussion, I'd also appreciate clarification about how much of 'AI' energy-use is going into creating images and videos, rather than text and code, my hunch is it’s much worse for videos most of which are about trivial stuff. Also loads of datacenter energy is wasted transmitting talking-head videos around the world - that's really inefficient. So well designed code, part-aided by ai, might help find more efficient ways to run needed global dialogue.

Seems like an evil plot to deploy energy sucking machines when we are in a global energy crisis. Maybe evil did win after all?

Disclaimer: not promoting or supporting AI usage, but adding some observations after reading.

Two points that needs more exploration.

One is briefly mentioned by the author that AI usage of a software engineer is at the very end of usage distribution. So this cannot be taken as a generalised estimate as the tiltle suggest.

Second, I don't see any mention of energy usage for research and development of the models. This is about what happens once the model is built. Unless AI companies disclose this, these are not going to put things into perspective.

Worth also considering all the extra times a kettle is boiled to make tea while waiting for the agents to complete

Ok. Is using AI compatible in the slightest with the very idea of Solarpunk? I'd say no - either 80% of commenters here are migrants from other communities or you should really rethink your life. (Really, defending AI power usage!? It's a waste regardless if it uses 2 million tokens or 1)

Grimy already pointed out that this article is non-scientific nonsense but there's something else it's missing: The efficiency of AI over time.

Qwen-3.8-27B is due out next week and it's supposed to be on par with Claude (Sonnet, latest). That runs on a regular PC GPU.

Big AI isn't the only game in town. They're just what makes the non-tech news.

The open weights stuff gets drastic improvements every three months or so and there's breakthroughs in efficiency every six months (ish). We're only at the baby steps of AI tech and the advancement is behaving like a chaotic Moore's Law that leapfrogs itself regularly, then stagnates for a bit, then there's another big shift.

If this keeps up, the efficiency of AI will become something that makes using huge data centers for anything but training a waste.