Monetize idle GPUs, honestly: marketplace, colocation — or sell the card

The three routes, without the funnel

Every “net” figure here is worked step by step below — gross at three utilisation levels, minus power — where a platform takes a host-side cut, that depends on the model and is handled in the fee section. Nothing here is an earnings promise; utilisation is the renter’s decision, not yours. Route 1 — list on a marketplace that takes hosts. You install the platform’s software, pass its verification, set or accept a price, and renters find you. The platform handles the customer, the billing and the metering, and takes its side of the money — how much, and taken from whom, differs by platform and gets its own section below. Your costs are power, wear, and keeping the machine online. This is the only route that makes sense below a rack of hardware. Route 2 — colocation. You rent rack space, power, cooling and bandwidth in a data centre, put your machines in it, and list them yourself (or through a marketplace on top). The facility bill is fixed and monthly: it arrives at 100% whether your machines rent at 100% or at zero. Colocation buys the density and uptime a home connection cannot offer — an 8-GPU node wants kilowatts, not a wall socket — and only pays once the fleet is big enough that the fixed bill is a small slice of the gross. Route 3 — sell the hardware. No funnel will suggest this, because nobody earns a commission when you sell your own card. It belongs in the comparison anyway: a used RTX 4090 sells today for roughly what 16–34 months of realistic renting nets (worked below), and the card at the end of two years of renting is worth less than the card today. Renting wins when you want the card kept, or genuinely believe utilisation will run high for years. Selling wins more often than the hosting guides admit.

The marketplace maths: one RTX 4090

Worked at the cheapest verified renter price for a 4090 hour on our board — $0.34/hr (RunPod Community, checked Nov 10, 2026). That is what a renter actually pays for this card today, so it is the honest rate for a host listing to compete at. A month is 730 hours. Power is the card alone at its rated 450W TDP, at an assumed $0.12/kWh — home rates vary widely, and the rest of the PC draws more, so treat the power line as a floor.

Two honest footnotes. First, the platform’s cut: on nafnaf’s model the 5% fee is added on the renter’s side, so a host asking $0.34 receives $0.34 and the renter pays $0.357 — the table stands as written there. On a platform that deducts from the host’s side, take that percentage off the gross line before power; the section below says which models we could verify. Second, a PC left on around the clock draws idle power too — at an assumed 50–100W (labelled assumption), roughly $4–$9 a month, owed whether or not a job ever lands.

The same maths at node scale: 8× H100

Datacenter cards rent as nodes. The best verified 8-GPU H100 rate on our board is TensorDock’s $2.25 per GPU per hour at an 8-GPU minimum — $18.00 per node-hour (checked Nov 10, 2026). Power is the eight cards alone at the H100 SXM’s rated 700W TDP each — 5.6 kW, before CPUs, networking and cooling, so again a floor.

The node’s gross is just over fifty times the single 4090’s — and unlike the home card, the node has a landlord. The line missing from this table is the facility bill, and it is missing on purpose: there is no honest universal figure for it. It gets its own section, as a formula, below.

What the platforms take — verified, and what is not published

One structural point matters more than any percentage: whether the fee comes off your side or is added to the renter’s. A fee added on top (nafnaf’s model, and Vast’s since 2024) leaves your rate intact but makes your listing look pricier to a comparing renter. What decides your month is never the fee alone — it is the fee model multiplied by the utilisation that platform actually delivers to a card like yours.

Colocation: the fixed bill, as a formula

Colocation is priced per kilowatt or per rack per month, negotiated per facility, per region, per power contract — Spheron’s own colocation guide concedes the same point, in its words that a colocation quote “is closer to a real estate lease than a cloud price list.” So this guide will not quote you a colo price. It gives you the only number that matters, computable the moment a facility quotes you:

That is the share of the month your machines must rent just to cover the facility, before power and before any platform cut. Worked example with a clearly illustrative bill: at a facility bill of $1,500 a month (illustrative — get quotes; your market will differ) and the node’s $18.00/hr, break-even is $1,500 ÷ $13,140 ≈ 11% of hours to cover the facility alone, and roughly 12% once the node’s power margin is counted. Double the bill to $3,000 and break-even doubles with it, to about 23%. The formula is the deliverable: a quote you cannot run through it is a quote you do not understand yet. Two anchors, both labelled. Spheron’s own cost model for a 1,024-GPU H100 cluster (their post, figures stated as of early Oct 2026) itemises colocation and networking at roughly $75,000 a month — about $73 per GPU per month — with power billed separately. That is a thousand-GPU buyer’s number; a single-rack tenant has none of that leverage and should expect a far higher per-GPU cost, which is why route 2 starts at “8+ datacenter cards” above and not at two. And density is the gate before price: Spheron states GPU colocation runs 10–15 kW per rack as the 2026 baseline against 3–5 kW for standard server colocation — so a standard enterprise colo contract may not be able to power your rack at any price.

The 70% question, tested

Hosting content routinely models 70% utilisation or better, because the numbers look good there. Nothing about a marketplace promises it: utilisation is set by how many renters want your exact card, in your location, at your price, against every other identical listing — for a consumer card, thousands of them. Look at the same 4090 across the range, from the table above: $208.78 a month rented every hour, $146.15 at 70%, $83.51 at 40%. The card did not change; the assumption did, and it moved the answer two and a half times. Any earnings figure quoted without its utilisation assumption printed next to it is marketing, not maths.

Now the comparison the funnel leaves out — renting against selling. A used RTX 4090 sold for roughly $2,400–$2,800 in recent completed eBay sales (sold-listing trackers read Nov 10, 2026; condition and model move the figure, and trackers disagree with each other — the method works with whatever your card would actually fetch). At the worked rates above, matching even the low end within a year means netting about $200 a month, against a ceiling of $208.78 if the card rents literally every hour: the card would need to rent about 96% of all hours for a year straight. Against the top of the range ($2,800, or $233 a month), no utilisation level gets there at all. Over two years the picture flips, but only just, and only at sustained high utilisation: 24 months at 70% nets about $3,500 before any host-side cut, against $2,400–$2,800 today — while the rented card ends those two years older, more worn, and worth less, and the $0.34 rate has to survive a market that has repriced everything else. The honest summary: renting a single 4090 is a reasonable way to keep a card you want anyway and offset its cost. As a pure investment against selling, it needs years and luck.

The demand is real — and it still is not a promise

The renter side of this market is genuinely tight, and our own record shows it. Providers do not raise prices twelve times into weak demand:

Hold the counterweight in the same hand. A rising list price at a datacenter provider is evidence renters pay more for capacity in facilities — it is not evidence that your specific card, in your specific postcode, beside thousands of identical marketplace listings, rents one hour more. Datacenter demand concentrates on datacenter cards in datacenter buildings; the consumer-card end clears at $0.34 an hour precisely because supply there is deep. Both facts are in our record. Quote either one alone and you are selling something.

When not to bother