#The break-even formula
A monthly server costs the same whether it runs for one hour or all month. An hourly instance costs its rate times the hours you keep it. The two cost the same at exactly one point:
break-even hours = monthly price ÷ hourly price
share of month = break-even hours ÷ 730
hours per day = break-even hours × 24 ÷ 730730 is the average number of hours in a month (8,760 hours a year ÷ 12). Take one H100 at our $1,779 a month and the market median of $3.49 per GPU-hour: 1,779 ÷ 3.49 = 509.7 hours. That is 69.8% of the month, or 16.8 hours a day. Below that, hourly billing is cheaper; above it, the monthly server is.
Both sides are per GPU. A server with n GPUs costs n times our per-GPU price, and the hourly market is quoted per GPU too, so the break-even is the same for one GPU or eight.
#Break-even for each datacenter GPU
The table uses the public on-demand list prices we collected from 40 GPU clouds and hosts on 23 September 2026: the median per-GPU hourly price, and the lowest one we found. Spot, marketplace and hyperscaler prices are left out.
| GPU | Our price per GPU / month | Median $/GPU-hour | Break-even vs median | Lowest $/GPU-hour | Break-even vs lowest |
|---|---|---|---|---|---|
| B300 | $4,019 | $7.87 | 511 h (16.8 h/day) | $6.60 | 609 h (20.0 h/day) |
| B200 | $3,439 | $6.74 | 510 h (16.8 h/day) | $5.43 | 633 h (20.8 h/day) |
| H200 | $2,279 | $4.465 | 510 h (16.8 h/day) | $3.49 | 653 h (21.5 h/day) |
| H100 | $1,779 | $3.49 | 510 h (16.8 h/day) | $2.69 | 661 h (21.7 h/day) |
| MI355X (indicative) | $1,409 | $2.77 | 509 h (16.7 h/day) | $2.59 | 544 h (17.9 h/day) |
| RTX PRO 6000 | $959 | $1.89 | 507 h (16.7 h/day) | $1.335 | 718 h (23.6 h/day) |
| L40S | $789 | $1.56 | 506 h (16.6 h/day) | $0.646 | 1,221 h (never) |
The pattern is not a coincidence. Our price is the market median × 730 hours, minus 30%, rounded down, so against the median the break-even always lands just under 70% of the month: about 510 hours, or 21 days of round-the-clock use. Against the cheapest listing, the L40S aside, it moves to between 17.9 and 23.6 hours a day.
No provider publishes an on-demand price for the MI355X, so its row uses the two public starting prices we found, $2.59 and $2.95 per GPU-hour, whose median is $2.77.
For the L40S, one dedicated host lists a lower price than ours: the equivalent of $0.646 per GPU-hour, about $472 a month. Against our $789 the break-even would be 1,221 hours, longer than a month, so that offer is cheaper at any utilisation.
GeForce cards are a different market: most hosts already rent them by the month, so the comparison is direct. Taking each provider’s cheapest server, the median RTX 5090 costs $481 a month across 15 providers, against our $329; the median RTX 4090 costs $387 across 16 providers, against our $269.
#Against hyperscalers and marketplaces
Hyperscalers are not in our medians, but many teams compare against them. Their on-demand rates are much higher, so the break-even comes early:
| On-demand reference | $/GPU-hour | Compared with | Break-even |
|---|---|---|---|
| AWS p5 (H100) | $6.88 | H100 at $1,779 | 259 h (8.5 h/day) |
| Google Cloud a3-highgpu-8g (H100) | $11.07 | H100 at $1,779 | 161 h (5.3 h/day) |
| Azure ND H100 v5 | $12.29 | H100 at $1,779 | 145 h (4.8 h/day) |
| AWS p5en (H200) | $7.91 | H200 at $2,279 | 288 h (9.5 h/day) |
| AWS p6-b200 | $14.24 | B200 at $3,439 | 242 h (7.9 h/day) |
| AWS p6-b300 | $17.80 | B300 at $4,019 | 226 h (7.4 h/day) |
Marketplaces work the other way. The median H100 on Vast.ai was $2.27 per GPU-hour, which puts the break-even at 784 hours: longer than a month, so it is cheaper at any utilisation. For the H200 ($4.34) the break-even is 525 hours, close to the on-demand figure; for the B300 ($11.24) it is 358 hours. Marketplace GPUs come from independent hosts, and an interruptible instance is stopped when someone outbids it, so this capacity suits jobs that checkpoint often.
#When hourly still wins
Hourly billing is the right choice when the GPUs are needed in bursts and released in between:
- Office-hours development. Eight hours a day on 22 working days is 176 hours. On one H100 at the $3.49 median that costs $614.24, well below $1,779.
- A one-off experiment. A 10-day run on 8× H100 is 240 hours: 240 × 8 × $3.49 = $6,700.80, against $14,232 for a month of the same server.
- Short scale-out. Dozens of GPUs for a weekend of evaluation, handed back on Monday.
- Interruptible batch work. Sweeps and batch jobs that checkpoint often can run on spot or marketplace capacity.
The longer the job, the closer it gets to the line. Three weeks on 8× H100 (504 hours) costs $14,071.68 at the median hourly rate, only $160.32 less than a month of the same server. A full month costs $20,381.60 by the hour, $6,149.60 more than the monthly server.
#The idle-time reality
The break-even assumes you pay only for the hours you need. In practice an hourly instance bills for every hour it is on, busy or not.
- Training is more than GPU time. Data preparation, debugging, failed runs, evaluation and checkpoint uploads keep the machine on while the GPUs are partly idle, and it is often simpler to leave an instance running between two experiments than to stop it and set it up again.
- Inference has to be up when requests arrive. A model served to users runs 730 hours a month, quiet nights included, so a production endpoint is past every break-even in the table.
- Stopping has a cost. A stopped on-demand instance returns its GPUs to the pool, and nothing guarantees the same configuration is free when you come back.
So the useful question is not how many hours your job computes, but how many hours you would keep the machine running. If that is more than 16.8 hours a day on average, about 510 hours a month, a monthly server beats the median hourly offer.
#Predictability
A monthly price is fixed for the month. The bill is the price times the number of servers, whatever the job does, and there is no meter to watch. Hourly list prices move: Nebius, for example, raises its on-demand prices for the H100, H200, B200 and B300 by 17 to 21% on 1 October 2026. An hourly budget has to absorb changes like that in the middle of a project.
Our servers are prepaid one month at a time and renewed month by month, with no long-term contract, so the commitment is one month, not a year. Payment is in BTC, ETH, USDT, USDC or LTC; see billing and payments.
#How to decide
- Estimate the hours per month you would keep the GPUs on, not the hours of pure compute.
- Divide our monthly price by the hourly price you would otherwise pay. If your hours are above the result, rent by the month; below it, rent by the hour.
- If you are close to the line, count the idle hours an hourly instance also bills, and remember that hourly list prices can change during your project.
Prices for every server size are on the pricing page, and the comparison page puts each GPU side by side. To size the machine first, read how much GPU memory your model needs. When you know what you need, configure a server.