Explainer · utilization · with video · Updated October 2026
95% of AI GPUs Are Sitting Idle
Most GPUs bought for AI, at big companies and in people's homes, spend most of the day doing nothing. That idle time is the hidden cost of owning one, and it's the reason renting is so cheap right now. Here is what an idle GPU really costs per hour of actual work.
Watch: 95% of AI GPUs Are Sitting Idle
One GPU, one day. The number I use in the video is 5%: the average company that went out and bought GPUs for AI is getting something like five cents of work out of every dollar of hardware. Five percent of a day is 1.2 hours. For the other 22.8 hours the card sits there, like a collection of Pokémon cards.
That matters because a GPU's price is paid up front, whether it works or not. Spread a $30,000 H100 over three years and it costs about $1.14 an hour if it runs flat out. Run it 1.2 hours a day and every hour of real work costs you over $22. The cheapest H100 we can rent today is about $2.
Owned H100: cost per hour of real work
Rent: $2.14/hr
$30,000 purchase spread over three years of use; power, cooling and hosting left out, and all of them make owning cost more.
Owned RTX 5090: cost per hour of real work
Rent: $0.39/hr
$1,999 launch price spread over three years; street prices above launch push every bar up.
What owning costs per hour of real work
| Hours of work a day | Utilization | H100 ($30,000) | RTX 5090 ($1,999) |
|---|---|---|---|
| 1.2 | 5% | $22.83 | $1.52 |
| 2 | 8% | $13.70 | $0.91 |
| 4 | 17% | $6.85 | $0.46 |
| 8 | 33% | $3.42 | $0.23 |
| 12 | 50% | $2.28 | $0.15 |
| 24 | 100% | $1.14 | $0.08 |
| Cheapest rental | you pay only while it works | $2.14 | $0.39 |
Three-year life, launch prices, no electricity or resale. The rental row is the cheapest hourly rate we tracked on RunPod or Vast.ai in October 2026.
Owning only wins if it stays busy. At today's rates an owned H100 has to work about 13 hours a day, every day for three years, before it beats renting, and that's before you pay to power and cool it. For a company that's a hard bar to clear. For a person it's almost impossible, because you sleep, you have a job, and AI work comes in bursts.
The RTX 5090 is closer: about 4.7 hours a day at its launch price. If you really are running something on it for five hours every day, owning one can make sense. Most people aren't, and at the street prices people are actually paying, the bar moves higher.
Why companies bought before they knew what to do with it. A lot of companies bought GPUs during the shortage because they felt they had to 'use AI', not because they had a job for the hardware. Some had more money than they knew what to do with, or could get it on credit. That's how you end up with racks of cards that work an hour a day.
It also explains where old cards end up. When a new architecture lands, the old ones get offloaded in bulk. That's what happened with the Tesla P40: not a bad GPU, just too old an architecture for serious work today. For anything serious you're looking at 80GB and up, like the 141GB H200 or the 288GB B300, and those you rent.
Pay for the work, not the waiting. The way I run GPU Battle's own tests is the way I'd suggest for anyone: set up batch jobs with an API key on a service like RunPod or Modal, and the GPU spins up only when there's work to do. Image generation, video, image-to-3D, data jobs: rent by the hour, then shut it off. Five dollars for an hour of serious compute is nothing next to five thousand for a card that sits idle.
Before buying anything, ask yourself one honest question: is this for a return on the money, or is it a hobby? Both are fine. Just don't pay the hobby price and call it an investment.
Our verdict
Utilization decides everything. An owned H100 working 1.2 hours a day costs about $22.83 per useful hour against $2.14 to rent one, and it needs about 13 hours of work a day to beat renting. Unless your card stays busy most of the day, rent bursts and batch jobs by the hour and pay only for the work.
FAQ
- What does GPU utilization mean?
- The share of the day a GPU is actually doing work. At 5% it works 1.2 hours out of 24; the purchase price is paid either way, so low utilization makes every useful hour expensive.
- Is it worth buying an H100?
- Only if it stays busy. Spread over three years, a $30,000 H100 beats the cheapest rental ($2.14/hr in October 2026) only above about 13 hours of work a day, before power and cooling.
- When does owning an RTX 5090 beat renting one?
- At its $1,999 launch price and a $0.39/hr rental, at about 4.7 hours of real work a day over three years. Higher street prices or electricity raise that number.
- Why are GPU rentals so cheap?
- Platforms like RunPod, Vast.ai and Modal pool thousands of GPUs and rent them to whoever needs them, so their cards stay busy far more than one owner's card would. That high utilization is what lets them charge so little per hour.
How we test
Cost per useful hour is the purchase price divided by the hours of work over a three-year life (hours a day x 365 x 3). Prices are launch prices: $30,000 for the H100 and $1,999 for the RTX 5090. Rental rates are the cheapest hourly prices we tracked on RunPod and Vast.ai on October 5, 2026; the rentals page has live numbers. Electricity, cooling, hosting and resale value are left out. Written from the GPU Battle video '95% of AI GPUs Are Sitting Idle'.