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.