The honest payback math on buying a GPU for AI rental in 2026

The revenue side

A GPU rents by the hour on an open marketplace. Current rates run roughly €0.10/hr for a 16 GB mid-range card up to about €0.43/hr for a 32 GB flagship. There are 720 hours in a month, but no machine rents continuously — utilization depends on demand for your specific card.

The marketplace takes 25% of gross before anything reaches you.

The cost side

Electricity is billed for hours actually worked: a 460 W machine at 55% utilization draws about 182 kWh/month, which is roughly €29 at €0.16/kWh. Managed hosting replaces that with a flat fee based on the machine's power envelope.

A management fee (ours is 15%) applies to what remains. It is charged last, on net, so it scales with your income rather than against it.

What that produces

For a 16 GB card around the €2,700 mark, hosted at home: roughly €22/mo at 40% utilization, €31 at 55%, €39 at 70%, and about €56 at a theoretical 100%.

That is a payback period measured in years — around 6 to 8 at realistic loads, not months. Anyone quoting dramatically better is either assuming 24/7 utilization or using stale rates.

When it still makes sense

The hardware remains yours throughout. Payback means the rentals returned the purchase price — not that the machine is worn out. It keeps its resale value and keeps earning after that point.

The calculation changes substantially if you use the GPU yourself. A card that rents at €0.22/hr represents roughly €159/month of compute you would otherwise pay a cloud provider for. If you would have bought that compute anyway, the rental income is offsetting a cost rather than being the whole return.

Where it does not make sense: as a short-term investment, as a guaranteed yield, or with low-VRAM cards that renters filter out.