An idle GPU is not free. Depreciation, colocation and operations keep running at zero percent utilization. Here is what one 8-GPU server really costs per month.
An idle GPU is not free. Depreciation, colocation and operations keep running even when the GPU is doing nothing, and gpu depreciation is usually the largest of them. This guide builds a simple monthly cost model for one 8-GPU server, allocates its fixed cost across the GPUs, shows the idle gpu cost under three depreciation assumptions with a Blackwell comparison, and ends with a calculator for your own numbers. Every input is labelled, sourced and dated, and none of it is a promise about what any GPU will earn.
Key takeaways
The instinct is to price idle time by the electricity meter. For a GPU server that is the wrong lens, because the largest costs are fixed. Four of them keep running whether the GPUs are busy or not.
Depreciation. The hardware loses value every month it exists, used or not. This is the biggest line and the least visible, because no invoice arrives for it. Colocation or power. GPU colocation is commonly priced per committed kilowatt, with committed kW at $130 to $250 per kW per month described as standard, so the bill does not shrink when the machines are idle. What a per-kW rate includes varies by provider, and metered billing exists but is less common, so check what your contract covers. Operations. Monitoring, spares, driver and firmware upkeep, and someone on call. Capital. If the purchase was financed, the interest runs on schedule, and if it was paid in cash, the money is not earning anything elsewhere.
Idle gpu waste is what those four lines add up to across the hours nothing useful runs. The rest of this guide puts numbers on it.
How long a GPU lasts depends on which clock you mean. Its physical life is how long it runs, its accounting life is whatever depreciation schedule a company chooses, and its economic life is how long it keeps earning the kind of margin that justified buying it. A card does not have to fail to lose value. That is why gpu lifespan, gpu useful life and gpu residual value are three separate questions, and why a gpu depreciation schedule that looks reasonable on one balance sheet can look aggressive on another.
The accounting choices are public for the largest buyers. According to its FY2025 filing, Amazon uses five to six years, having cut a subset of its servers and networking equipment from six years to five effective January 1, 2025. Meta's FY2025 filing lists servers and network assets at five to 5.5 years, and Alphabet is reported at six years. Meta's extension to 5.5 years reportedly cut its depreciation expense by roughly $2.3 billion in the first nine months of 2025 alone. These are accounting useful lives, not guarantees of how long a GPU stays economically productive. On the other side, investor Michael Burry argues that real replacement cycles run closer to 2 to 3 years and that ai chip depreciation is understated across the industry, while Nvidia has argued that 4 to 6 years is the real window and that 2 to 3 is simply wrong. Those are accounting policies for hyperscaler fleets, not rules for yours, but they show how much server depreciation depends on one assumption.
The standard formula also includes residual value. Depreciation under GAAP is (cost minus salvage value) divided by useful life, not cost divided by life. The model below uses zero residual value to keep the arithmetic visible, and a note after the results shows what a residual value does to it.
This model covers depreciation, committed colocation or power capacity and operations. Metered electricity, networking, financing and other costs vary by deployment and are not included unless you enter them. For scale, interest at an illustrative 8% a year on $285,000 would add about $1,900 a month.
Fixed monthly cost of one 8-GPU server, by hardware and depreciation life
Depreciation Colocation Operations
Two readings matter. First, the depreciation assumption moves the answer by about 60% across the three H100 scenarios, which is more than any other single input. Second, residual value pulls the other way: with a 20% residual value at the end of a 5-year life, the depreciation line falls from $4,750 to $3,800 a month. The gpu resale value you can realistically expect is itself an assumption, so it is worth testing both ways.
The same arithmetic on newer hardware shows the same shape. An 8-GPU B200 server at an illustrative $450,000 (reported 2026 quotes range from $400,000 to $500,000, and one April 2026 source puts DGX B200 systems lower, so sources disagree), provisioned at the roughly 14.3 kW maximum system power NVIDIA lists for the DGX B200, with the same $200 per kW and $600 operations, works out to about $10,960 a month on a 5-year life, or roughly $1,370 per GPU. Higher hardware cost and higher power draw both raise the floor.
The fixed cost above is paid in full at any utilization. What utilization changes is how many hours carry it. Using the 5-year scenario, one fully idle day costs the server about $235, or about $29 per GPU. Here is what that idle gpu cost and gpu utilization cost look like across a month at three levels, where utilization means the share of hours the GPUs do paid or otherwise useful work.
The last column is the useful one. Your cost per productive GPU-hour is the base cost per GPU-hour divided by utilization, so halving utilization doubles it. That is the number to hold against any rate or internal charge-back you can verify. A GPU that is busy a quarter of the time has to earn or save four times its base hourly cost in the hours it works just to cover itself. This is fixed-cost allocation, not an all-in cost per GPU-hour: metered electricity, networking and financing sit outside it unless you add them.
One caution on the word utilization. A dashboard reading is not the same as productive work, because the standard GPU utilization percentage only says that some kernel was running. The packet.ai guide on why GPU utilization dashboards mislead covers the difference, and it applies here: use hours of real, useful work, not the dashboard percentage.
Idle hours rarely look like a decision. They accumulate in predictable places.
Nights and weekends. A team that uses a cluster from 9 to 6 on weekdays touches 45 of the 168 hours in a week, which is about 27%. A two-shift team reaches about 54%. Neither figure allows for waiting inside those hours. Project gaps. A training run ends, the next dataset is not ready, and the machines wait. Hardware bought for a peak sits through the valleys, and the cost of idle gpu hours in those valleys is the same as in the peak.
Schedule. Move work into the quiet hours. Batch jobs, retraining and evaluation can run overnight and at weekends if they are queued rather than launched by hand. This costs engineering time and nothing else, which makes it the cheapest first step.
Share. Put the hardware behind a scheduler so several teams draw on one pool instead of each holding its own idle machines. The aim is fewer machines sitting empty because one group booked them for a peak that has passed.
List. If hours stay idle after scheduling and sharing, selling them is the third option. Platforms and networks differ widely in what they require, what they pay and how they treat your own use of the hardware, so compare the routes before applying. Some marketplaces require dedicated machines: Vast.ai's verification documentation says personal workloads automatically fail verification, so check how a platform treats your own jobs before you rely on listing the gaps. The packet.ai guide to the five ways to turn idle GPUs into revenue sets out the options, and the provider guide covers packet.ai's own program. Revenue from a spare hour only reduces the cost of an idle gpu if it exceeds what that hour really costs you to sell, so run the model again with the platform's terms before committing.
Replace the example inputs with yours to see the real cost of owning a gpu month by month. The calculation takes five lines.
The calculator below runs the same five lines in your browser and sends nothing anywhere. Its default values reproduce the 5-year H100 scenario at 50% utilization.
Idle GPU cost calculator
Enter your numbers and press the button.
The last line of the output, cost per productive GPU-hour, is the number to hold against any rate or internal charge-back you can verify yourself. This guide does not quote a rental rate, since rates move.
Run it three times, once for each useful life you can defend, and keep the range instead of a single answer. If the range is wide, the depreciation assumption is doing most of the work, which is the same conclusion the hyperscaler accounting debate reaches. If you decide the idle share is worth selling, the packet.ai provider page describes what its program involves and the requirements change over time, so read the current version.
Retrieved October 7, 2026. Hardware prices and colocation rates move quickly; verify current figures before you use them.
Last reviewed: October 7, 2026. This is an illustrative cost model, not accounting or financial advice, and it contains no earnings figures and no guarantee of income. Hardware prices, colocation rates and depreciation practice change frequently, and the example's inputs are assumptions you should replace with your own.
Same models. Same API. Fraction of the cost. Start free — no credit card required.
Start Building →