I would think far less. HFT isn’t intentionally computationally expensive.

And the HFT industry is incentivized to use as little processor time as possible, and thus as little power as possible.

That's not true.

They'll take a 400W 10GHz CPU over a 5mW 1 GHz CPU every time, if it generates better financial returns.

Parent was referring to latency optimization, I believe.

But I'd assume most of those folks have ML / other slower models constantly churning away in the background, even if their HFT reactions are highly tuned and simpler.

My impression is that to the extent that HFT is about relative speed, competition can't reduce power usage.

And how much by machinery to run advertising?

Not to mention all the extra CPU load the tracking and JS spam causes on everyone's phones and computers.

There's a large overlap between serve farms and energy use backed-by/for marketing.

you're not going to hear the media complaining about that, though

Not very much. A handful of machines per exchange colo, for each of a few hundred orgs? Probably about as much power as a small-midsize college, if I were to guess?