Research

The Cost of an Idle GPU: What an Unused H100 Costs You Per Day

Back to BlogWritten by Published Nov 12, 2026
Idle GPUCost of Idle GPUIdle GPU Opportunity CostGPU Depreciation Per DayH100 Idle CostAverage GPU UtilizationGPU Supplier ProgramAI Infrastructure
The Cost of an Idle GPU: What an Unused H100 Costs You Per Day

An idle GPU doesn't just fail to make money, it loses money twice over. Once as depreciation, a straight-line hit to book value that runs whether the card does a single job or not, and once as opportunity cost, the rental revenue it could have earned instead of sitting dark in a rack. Most of what currently ranks for "idle GPU" is a thermal-forum thread about fan curves and nvidia-smi output. None of it prices what an idle H100 actually costs the person who owns it, in dollars, per day. That number is bigger than most fleet owners assume, and it compounds every single day the card sits unlisted.

TL;DR: The Cost of an Idle GPU, What an Unused H100 Costs You Per Day

  • Price: Spheron H100 SXM5 runs $2.65/hr on-demand ($2.20/hr spot) as of 17 Nov 2026; check Spheron's H100 rental page for today's rate before running this math on your own fleet.
  • Depreciation: A $25,000-35,000 H100 card (channel listings, 20 Jul 2026) depreciated straight-line over 5 years books roughly $14-19 a day, whether it's used or not.
  • Utilization: Average GPU utilization across tens of thousands of Kubernetes clusters is 5%, per Cast AI's 2026 report, about 20x more capacity provisioned than used.
  • Breakeven: A self-funded GPU marketplace operator needs roughly 70% utilization to break even, per Hashrate Index's neocloud unit-economics model.
  • Fix: Spheron's GPU Supplier Program lets an owner list idle H100 through B300 capacity and set their own price floor, narrowing the demand-finding half of the utilization problem.

Depreciation vs Opportunity Cost: Two Different Numbers

These two numbers get conflated constantly, and they shouldn't be, because only one of them changes based on what you actually do with the GPU today.

Depreciation is fixed and scheduled. An H100 card runs $25,000-$35,000 on channel listings, and once it's bought, its book value falls on a straight-line schedule regardless of whether it trains a model around the clock or sits powered off for a quarter. The accounting clock doesn't track utilization, it tracks the calendar. Our own depreciation breakdown goes further into the gap between that book schedule and what the card is actually worth on the resale market by generation, which matters for refresh timing but doesn't change this calculation: book depreciation runs on its own clock no matter what the GPU does.

Opportunity cost is the opposite. It only exists because demand existed somewhere and the owner didn't capture it. If nobody wanted to rent that GPU this hour, there's no opportunity cost, just the depreciation sitting underneath it. But GPU rental demand is real, which is exactly why an unlisted or unrented H100 isn't a neutral asset quietly aging. It's actively forgoing revenue every hour it sits there with nobody renting it.

There's a third cost that rarely gets modeled at all: facility power. A system that's powered on but idle still pulls a meaningful fraction of the full-load power just keeping memory, cooling, and networking alive. Being "idle" in a server rack rarely means the power bill drops to zero. An owner financing the buildout feels this acutely: a GPU-backed loan still has debt service due every month whether or not the collateral underneath it is earning anything that month.

The Daily Math at Today's Rental Rates

Put real numbers against an idle H100 and the daily total is higher than it looks from either side alone.

Cost componentPer day (per GPU)Basis
Book depreciation$14-19$25,000-35,000 buy price, straight-line over a 5-year schedule
Forgone revenue, on-demand$2.65/hr x 24 hoursSpheron H100 SXM5 on-demand rate, full 24-hour day, as of 17 Nov 2026
Forgone revenue, spot$2.20/hr x 24 hoursSpheron H100 SXM5 spot rate, full 24-hour day, as of 17 Nov 2026

As a frozen worked example, not a live figure: on 12 Nov 2026, Spheron's H100 SXM5 ran $2.39/hr on-demand and $2.20/hr spot, which put depreciation plus forgone on-demand revenue at roughly $71-76 a day combined, and depreciation plus forgone spot revenue at roughly $67-72 a day. Those input rates move with the market, so rerun the math with the live rows above or check current GPU pricing rather than treating that snapshot total as current.

The combined figure assumes an owner who could sell every idle hour at the on-demand rate, which is itself optimistic. Most fleets can't fill 100% of their idle hours even once they start listing. But it's the ceiling the math is working against, and that's exactly why even a modest climb in utilization moves real money: going from 20% to 50% utilization on a 10-card fleet is the difference between collecting a small fraction of the on-demand opportunity cost row above and collecting most of it, without buying a single additional GPU. Depreciation keeps accruing at $14-19/day per card no matter what; utilization only moves the forgone-revenue half of that math.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 17 Nov 2026; other providers reflect their most recent published rates and may have changed. Check current GPU pricing → for live rates.

Why Average GPU Utilization Sits at 5%

Cast AI's 2026 State of Kubernetes Optimization Report puts average GPU utilization at just 5%. Laurent Gil, Cast AI's co-founder and president, put the gap bluntly: "A GPU sitting idle costs dollars per hour. A CPU sitting idle costs cents. And 95% of GPU capacity is doing nothing."

That 5% figure means organizations are provisioning roughly 20x more GPU capacity than they're actually using. Run the per-day math from the section above across even a modest fleet sitting at 5% utilization, and the forgone revenue and depreciation stack up fast, mostly invisibly, because nobody's P&L has a line item called "idle GPU cost." It shows up instead as a vague sense that compute spend is high relative to output.

So is 90% utilization too aggressive a target? No, it's the opposite of the industry's actual problem. A well-run fleet or a marketplace listing clearing 90% utilization is close to ideal, not overloaded; the real failure mode the Cast AI numbers describe is fleets running at 5%. The gap between "owns a GPU" and "profits from a GPU" is almost entirely this number.

Turning Idle GPU Capacity Into Listed Revenue

The fix for the daily math above isn't a cheaper GPU or a different depreciation schedule, it's finding renters for the hours that would otherwise sit unused. That's a demand problem, not a hardware problem, and it's the specific gap a marketplace is built to close.

Spheron's GPU Supplier Program isn't a reverse auction and it doesn't resell capacity under its own brand, so a supplier keeps the pricing control that a fixed colocation lease trades away.

It's worth being plain about what this doesn't solve. It doesn't negotiate power, cooling, or the facility lease, so the colocation or ownership decision has to already be settled before a listing is the relevant next question; our colocation vs marketplace breakdown covers that earlier decision in detail. And it's a demand-routing and billing layer, not a guarantee of hitting a given utilization number. An owner with no track record still has to pass manual vetting, and a listing that's missing documented networking or SLA language gets passed over by renters no matter which marketplace it's on, as the specifics of what gets a submission skipped make clear.

None of that changes the core arithmetic from earlier in this post. An H100 sitting unlisted still books $14-19 a day in depreciation and forgoes a full day of on-demand revenue, 24 hours at $2.65/hr as of 17 Nov 2026, in rental revenue it could otherwise have collected. The only input that moves that number is whether someone is renting it, and that's the half of the equation a listing is built to fix.

An idle GPU's daily cost doesn't go away because nobody's renting it. Listing the capacity is what actually closes the gap between the 5% utilization most fleets run and the roughly 70% a self-funded marketplace seller needs to break even.

Apply to Spheron's GPU Supplier Program →

FAQ / 04

Frequently Asked Questions

Two numbers, not one. Straight-line book depreciation on a $25,000-35,000 H100 card runs roughly $14-19 a day over a 5-year schedule, whether the card ever powers on. On top of that, every idle hour is rental revenue the owner never collects: Spheron's H100 on-demand rate was $2.65/hr as of 17 Nov 2026, so 24 idle hours add up to roughly 24 times that hourly rate in forgone revenue alone. Add the two together and an idle H100 costs its owner well over $50 a day, depreciation plus whatever that day's on-demand rate works out to.

5% is the average, not a healthy target. Cast AI's 2026 State of Kubernetes Optimization Report found average GPU utilization sits at just 5%, meaning most fleets provision roughly 20x the GPU capacity they actually use. A fleet or marketplace listing running closer to 90% is a sign of a well-matched supply-to-demand setup, not an overloaded one; the real problem the industry has is nowhere near that high.

Depreciation is fixed and scheduled: the card loses book value on a straight-line schedule whether it runs a job every hour or sits dark in a rack for a quarter. Opportunity cost is variable: it's the rental revenue the GPU could have earned that specific hour, and it only exists where there was demand to sell into. A GPU owner who never lists capacity still books the depreciation loss; they just never recover any of the opportunity cost sitting on top of it.

By finding renters for the hours the GPU would otherwise sit unused.

// Try It Yourself

Try It on Real GPUs

The GPUs behind these guides are the ones you can rent here: H100s, H200s, B200s, and more, billed per minute after a 20-minute minimum runtime, with no contracts. Pick one and you are live in under two minutes.

View Pricing
→
Or talk to our team
Spheron GPU instance going live and linking data centers across the globe