GPUDECK · SUPPLY SIDE · FOR OPERATORS
Most “how much can my GPU earn” pages multiply a listed hourly price by an assumed utilization and call it income. The record says that arithmetic is wrong on both factors — and it is the only marketing this page will do.
The listed price is an ask, not your revenue. The price context publishes medians and floors per segment: the spread between them is real, and where you land in it depends on your hardware, host quality and reliability tier. Utilization is not a constant you choose. And even an occupied hour is not fully paid: on the measured fleet, cancelled attempts consume 0–34% of occupied time by card — work that holds your GPU and earns nothing (the goodput record). Completion yield runs 0.68–0.95 by card on the reference fleet.
The daily census shows models entering and leaving a public GPU network continuously. Presence is not earnings — but churn is real, and a card that was competitive at listing time can stop being rented without notice. Persistent absence in the census is consistent with operators withdrawing hardware; the census does not establish why.
No earnings projections, no “best marketplace to list on”, no payback calculators. Those numbers depend on utilization and yield figures nobody publishes honestly for your exact card in your exact deployment. What the record offers instead: dated asks by segment, measured goodput and yield on one disclosed fleet, and the raw files to run your own arithmetic with stated assumptions.