The greenest GPU is the one already racked.
Every offer this site routes toward is a machine that exists today: bought, shipped, racked, powered — and waiting. Using it displaces demand for another machine, not for another data center.
Utilization before expansion.
An idle GPU is manufacturing footprint already spent, producing nothing. The industry's default answer to demand is more hardware; the cheaper and cleaner answer, wherever it fits, is more work per machine that already exists. That is the entire product: a market that makes the idle machine the easiest one to find.
The arithmetic of an already-built card.
A GPU's footprint is paid twice: once in the factory, once at the wall. The factory share — fabrication, packaging, the data-center shell built around the rack — is spent the day the machine lands, whether it ever works or not. Put a job on an idle card and the only new cost is the marginal wall-power of the work itself; the host was already drawing standby power in a cooled room. Meet the same demand with new hardware instead and the factory cost is paid all over again — plus the standby draw of one more machine that will, in turn, sit idle between jobs. We publish no number for the factory share, because we cannot meter it from here; published server life-cycle assessments put it at a material double-digit fraction of lifetime emissions, and the displacement argument holds at any value in that range.
For the Scope 3 reader.
Plainly: renting through this market gives you no certificate and no line-item for your inventory — we do attributional carbon accounting for nobody, ourselves included. The claim on this page is consequential, not attributional: hours served on existing slack are hours that never help justify the next rack. If your framework counts certificates, this page offers none — and that refusal is exactly what the number above is worth.
Idle capacity is demand-shaped, not wasted.
Interruptible listings are slack in the grid of compute — priced low precisely because they are spare. Batch jobs, training runs, render queues and agent workloads can soak that slack instead of pulling new capacity online. The market does the routing; the price does the persuading.
What we refuse to publish.
No carbon calculators, no offsets, no green badges, no wattage we didn't meter. Claims like that need measurements we don't have — so we don't make them. We publish exactly one environmental number: the share of the listed market sitting idle, measured from live listings and recorded every half hour. It's charted on the idle-capacity index, and it comes with its caveats attached.