We broke down 30 days of rewards across the fleet by GPU class, region and uptime. The short version: utilization beats raw FLOPs.
We broke down thirty days of rewards across the fleet by GPU class, region, and uptime. The headline: a well-utilized RTX 4090 frequently out-earns a poorly-scheduled A100. Utilization beats raw FLOPs.
Spot price floats with demand, so the operators who earn most aren't the ones with the biggest cards — they're the ones online during peak scheduling windows with low latency to the regions generating jobs.
Region matters more than expected. Nodes in Tokyo, Frankfurt and São Paulo saw the highest fill rates this quarter as teams in those time zones ran overnight benchmark sweeps.
If you're considering running a node, the practical advice is boring and correct: keep it online, keep latency low, and let the scheduler do the rest.