Post

This is great - it's about time someone updated the discourse on LLM energy usage to reflect that coding agents use massively more prompts than occasional questions to ChatGPT Simon estimates that a day of coding agent usage comes out close to the energy needed to run a dishwasher

Simon P. Couch@simonpcouch.com

Whenever I read discourse on AI energy/water use that focuses on the "median query," I can't help but feel misled. Coding agents like Claude Code send hundreds of longer-than-median queries every session, and I run dozens of sessions a day. On my blog: www.simonpcouch.com/blog/2026-01...

89 reposts399 likes

I'm wondering if the end goal should not be to scale down the LLMs to the extend of them running competently on a standard dev laptop and having the inference go there. This way, you still sell a product license (cursor and the likes), model upgrades, but then defer the energy cost on the end user

I'm saying this from a mac that has trouble answering large context questions in ollama with a qwen 14b and similar. But the results are good-ish. Go small, I say.

The problem there is that we've seen the capability difference between the best available models (Opus 4.5, GPT-5.2) and the best ones you can realistically run on a laptop - the ~30B Qwens, GLM-4.7 Flash etc The competitive open weight models - full GLM-4.7, Kimi K2 - need >600GB of RAM

Maybe in a few years time 1TB of GPU-accessible RAM will become affordable for developer laptops, at which point we may get local models that compete with today's frontier models

This may never happen. I think it makes so little sense to have this resource locally available. The analogy here is flying in a plane across the Atlantic. That does make (commercial) sense for one person alone (rare exceptions exist).

I think it’s just as likely if not more that 100B param models will arrive that perform as well as today 1T param models.

I agree that as of today, the difference is massive. It's more of a general phlosophical musingon my end, I wonder if the industry tries to go beyond the transformers or earnestly looks at their relative efficiency. I fear it's conceptually easier to spend that VC money on data centers though.

What I’m really curious about is the impact of fleets of frontier-model bots. People are going to be Hello World-ing that approach a lot soon.

All signs very much do point towards orchestration of a large N of agents 24/7 with durable state so context rot doesn’t happen. A serious acceleration but one that is significantly hard capped on power capacity

Different take:

Ed Zitron@edzitron.com

This piece is full of egregious assumptions about the cost of running Anthropic’s models based on absolutely nothing, but this may be the worst - estimating that the cost of serving Claude opus 4.5 is comparable to GPT-4o. Sometimes I wonder if techmeme has a gas leak

This is the kind of honest accounting the AI discourse desperately needs. Most energy takes are either "it's nothing" or "it's boiling the oceans". Dishwasher is usefully specific.

Based on estimated costs per query for some of these companies to even break even, I have to admit I expected it to be several orders of magnitude larger. (I guess the costs are mostly in training and hardware?)

Training is expensive and they're all spending like crazy in a desperate race to build the best models on the basis that they think that will win them the market. Which I'm not entirely convinced is well-founded.

That's a great analysis. I guess the comparison with a dishwasher it's just for us to use as reference, right? Because if not, it's a weird comparison. Cloud code is a professional tool. We don't compare an industrial oven with the home air fryer.