Best Way to Monetize Idle GPUs: Colocation vs GPU Marketplace (2026)

The best way to monetize idle GPUs splits two structurally different ways, and the gap between them is bigger than most owners assume. An NVIDIA DGX H100 system draws up to 10.2 kW at full system load, and that kilowatt figure is what both monetization paths for an idle GPU fleet actually price against. Lease that power to a tenant under a colocation deal and you get a fixed rate every month, whether the rack runs at 95% or sits mostly dark. Keep the GPUs yourself and sell GPU-hours through a marketplace instead, and the same kilowatt can earn several times more, or come up short of covering its own power bill, depending entirely on how often someone actually rents it. This isn't colocation feeding a marketplace listing as sequential steps on one deal, which is the usual framing; it's two separate payouts on the same power.

TL;DR: Colocation vs GPU Marketplace, Which Pays More?

• Colocation's unit: a DGX H100 draws up to 10.2 kW at full load, the figure colocation leases bill against, per NVIDIA's design guide.
• Marketplace's breakeven: a self-funded operator needs roughly 70% utilization to break even, per Hashrate Index; a debt-financed build on a signed offtake contract can run near 100%.
• Rate swings: H100 rental fell 64-75% from its 2023 peak, then rose 20% year over year through 2026, per Hashrate Index.
• Where they meet: Spheron's GPU Supplier Program lets an already-colocated owner keep their own price floor. See the live H100 GPU rental rate.

The Best Way to Monetize Idle GPUs: Lease the Floor or List the Capacity

Most writing on this topic, including our own GPU colocation explainer, treats colocation and a marketplace listing as two layers of the same deal: colocation gets the facility, the marketplace gets the renters, and you need both. That's true for an owner who already owns the GPUs and just needs demand. But step back one level and there's a genuinely separate decision sitting underneath it, about the same kilowatt of power.

If you control data center capacity, specifically power, cooling, and floor space, you have two structurally different ways to turn that capacity into revenue:

1. Lease it to a tenant. Sign a colocation agreement with a GPU owner who brings their own hardware. You get paid for the power, cooling, and space you provide, at a rate set when the lease is signed. What the tenant's GPUs do after that, whether they run at 90% utilization or sit idle for a month, doesn't change what you're paid.
2. Operate GPUs on that power yourself and sell compute. Buy or lease the GPUs, run them on your own power and floor space, and list the capacity on a marketplace so renters can book GPU-hours directly from you. Now your revenue is a function of how many hours actually sell and at what rate, not a fixed lease payment.

Both paths monetize the same kW. The difference is who carries the utilization risk and who gets the upside when demand is strong. Our neocloud explainer covers the related question of how a GPU-first operator builds and sells capacity once that second path is chosen; this post stays one level up, on the choice between the two paths themselves, before either one is built out further.

What Colocation Actually Pays: Rent Per kW, Not Per GPU-Hour

Colocation income is set by the lease, not by what runs on the rack. That's the entire appeal of the model and the entire cap on its upside.

The unit the lease actually prices against is power, usually quoted per kilowatt committed, not per GPU or per hour of compute. A DGX H100 system is a useful reference point here because NVIDIA publishes the number directly: the system's maximum power draw is 10.2 kW at full load. Whatever a facility's per-kW colocation rate is, that's the figure it gets multiplied against for a single DGX H100 node, and it doesn't matter to the lease whether that node is training a model around the clock or powered on and idle.

There's no single published rate here the way there is for GPU rental. Our colocation cost breakdown goes through why: colocation pricing moves with utility power rates in the region, whether the facility is a retrofit or purpose-built for GPU density, the redundancy tier (Tier III vs Tier IV power and cooling), and whether cooling is metered separately or bundled into the base rate. A quote from one facility tells you very little about what another facility would charge for the same rack, because colocation prices like a real estate lease, negotiated per deal, not like a rate card.

What that negotiated, fixed structure buys the lessor is certainty. The rent doesn't fall if the tenant's workload dries up, and it doesn't rise if the tenant's GPUs suddenly get fully booked. For an owner deciding how to monetize spare power capacity, that's the honest tradeoff on this side of the choice: a lower, flatter ceiling in exchange for income that doesn't depend on anyone actually renting the compute.

What a GPU Marketplace Actually Pays: Utilization-Driven, Not Fixed

Selling GPU-hours yourself removes the fixed-rent ceiling and replaces it with a number that moves every day: how many hours actually sell, and at what rate.

The first variable is utilization, and the breakeven line is higher than most owners assume. Hashrate Index's own neocloud unit-economics model puts the breakeven point for a self-funded operator, one taking demand risk directly by selling into a marketplace rather than against a signed offtake contract, at roughly 70% utilization. Below that line, the GPU-hours sold don't cover the capex, power, and operating cost of running the fleet. A debt-financed build underwritten against a contracted buyer can instead run close to 100% effective utilization, because the revenue is locked in before the GPUs ever power on; that's a fundamentally different risk profile from selling spot capacity to whoever shows up.

The second variable is the rate itself, and it's far less stable than a colocation lease. H100 rental rates fell 64-75% from their 2023 peak before rising roughly 20% year over year through 2026, with one-year contract rates up nearly 40% from their late-2025 low. A separate tracker tells a similar story from a different angle: H100 on-demand rental fell roughly 64% from about $8/hour at late-2024 peak levels to $2.85-3.50/hour, while the Silicon Data H100 Rental Index fell 23%, from $3.06 in September 2024 to $2.36 in June 2025. For comparison, Spheron's own H100 SXM5 on-demand rate currently runs $2.39/hr, with spot at $2.21/hr, as of 08 Nov 2026. None of that volatility shows up in a colocation lease. It shows up directly in a marketplace seller's monthly revenue.

Capex matters too, because it sets where the breakeven utilization actually lands. All-in capex for a 1,000-GPU H100 SXM cluster runs roughly $50,000/GPU on a greenfield build, against roughly $42,000/GPU on a facility converted from an existing operation that already has power and shell infrastructure in place, an $8,000/GPU gap driven almost entirely by the facility line rather than the GPUs themselves. A lower capex base needs fewer GPU-hours sold to cover it, which is one reason the same ~70% figure isn't a universal constant; it moves with how expensive the build underneath it was.

This is the specific gap a demand-aggregating marketplace is built to close, since the harder part of clearing 70% utilization for most owners isn't the hardware, it's finding enough renters. Spheron's GPU Supplier Program lets an owner list H100, H200, B200, or B300 capacity, set their own on-demand and reserved price floors, and approve every deal, while Spheron routes qualified demand by SKU, region, and contract length and handles metering, billing, and monthly settlement. Vetting on uptime history, networking, and reference deployments typically closes in under a week, with full onboarding averaging one to two weeks. What it doesn't do is change the underlying math: it narrows the demand-finding half of the utilization problem, not the capex or the rate volatility, and it's a routing and billing layer, not a colocation broker, so the power and facility side of the decision has to already be settled before a listing is the relevant question. An owner already running near-full utilization through existing direct contracts also gets less marginal upside from adding a listing on top of what's already booked.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 08 Nov 2026; other providers reflect their most recent published rates and may have changed. Check current GPU pricing → for live rates.

The Best Way to Monetize Idle GPUs, by Fleet Size

The honest answer to "which pays more" depends on how confidently you can clear that roughly 70% utilization line, and that confidence changes with scale.

Take a concrete case to see why the line matters more than the sticker rate: ten DGX H100 systems, each drawing up to 10.2 kW at full load, is roughly 100 kW of committed power. Run that fleet at 70%+ utilization most months and selling GPU-hours through a marketplace outearns a flat colocation lease on that same power, by Hashrate Index's breakeven math. Run it at 50%, and the flat colocation rent would have paid more, because the marketplace revenue never cleared the capex and power cost of running the GPUs at all. Same power, same hardware, opposite answer, and utilization is the only input that flips it.

A handful of nodes, no existing customer base. At a few DGX H100-class systems, maybe 80-160 kW of power total, finding enough renters on your own to clear 70% utilization month after month is genuinely hard; there's no sales team, no brand, and not enough inventory to absorb a slow month. Colocation's fixed rent is the safer monetization path here, because the alternative is real downside risk: GPUs that cost real capex to acquire, sitting below breakeven, with nothing to fall back on. Listing on a marketplace can still help close the demand gap at this scale, but it's a supplement to an uncertain pipeline, not a guarantee of hitting the utilization line.

A mid-size fleet with some demand already in hand. Once an owner has enough inventory, and enough existing relationships or marketplace-routed demand, to make 70%+ utilization a realistic target rather than a hope, the marketplace path starts to win on expected value. The downside is capped by the same capex and power cost either way; the upside from selling GPU-hours at a marketplace rate, even a volatile one, is structurally higher than a flat lease once utilization clears that line consistently. This is also the scale where the shape of demand starts to matter: a fleet selling single-node, short-duration capacity looks for a different kind of marketplace than one selling in multi-node blocks. Spheron's own comparison against SF Compute is a useful illustration of how differently two marketplace-adjacent platforms can be built, one around instant per-GPU self-serve access, the other around cluster-scale CLI purchasing, and which shape fits depends on whether your fleet sells in single GPUs or 8-GPU nodes.

A large fleet with signed offtake contracts. At the scale where an owner can underwrite a build against a contracted buyer rather than selling spot into open demand, the math changes again: near-100% effective utilization is achievable through the contract itself, not through finding enough renters, which is the debt-financed model Hashrate Index describes as the alternative to the self-funded, marketplace-exposed path. At that scale, the $8,000/GPU capex gap between a greenfield build and a facility with power already in place also compounds across enough GPUs to materially change the return, which is its own argument for why large operators lean toward converted facilities when one is available. Our GPU cost optimization playbook covers the utilization math from the renter's side of this same market, useful context for understanding who's actually bidding on the capacity a large fleet lists.

The decision rule that falls out of all three cases: if you can't yet prove demand will clear roughly 70% utilization, take the fixed colocation rent. Once demand, whether direct or marketplace-routed, reliably clears that line, operating the GPUs and selling GPU-hours earns more per kW than leasing the power out ever would.

Once the facility side of this decision is settled, the question that's left is finding renters for whatever utilization gap colocation's fixed rent doesn't cover. Spheron's GPU Supplier Program routes vetted buyer demand to H100, H200, B200, and B300 capacity that's already colocated, while you keep your own price floor and sign your own contracts.

Apply to Spheron's GPU Supplier Program →