Hacker News Reader

Comment by hedora | Hacker News Reader

Yes, but, I can run qwen 3.8 flash next (> 100b parameters; ~ opus 4.6) on a 200W tdp desktop.

So, assume 8 of those, and it’s a space heater per house. I have never been disturbed by my neighbor’s space heater.

The problem is centralization, not the absolute energy usage. (And also that LLMs are trending to 100x more efficient than the data center sizing assumed).

View on HN

Centralization is vastly more energy efficient than decentralization, by orders of magnitude. LLMS love batching, to an insane degree. Running a single conversation through a GPU is about the same cost as doing ~100 in parallel.

Centralization is a huge huge environmental win, far far more than even cloud computing was compared to tons of inefficient, under utilized racks spread though our office buildings.

Huge environmental win if considering a fixed supply & demand, but Jevons Paradox kicks in with tokens in the cloud

This is very true. I'm more careful with what I ask my local model, since it's about 1/3 as fast as the cloud models I'm used to.

> Centralization is a huge huge environmental win

In a vacuum, yes. Assuming everyone is a good actor.

Totally true. But then it introduces a much greater degree of control, and it creates the aforementioned local concentration of energy usage, noise, etc. So it's a tradeoff, more than a strict win.

Decentralization can also cause havoc because of back-EMF for instance, in any case the public infrastructure needs time to adapt to demand.

Reply on news.ycombinator.com