Elastic compute
Rent the spike.
A heavy job once a month does not justify a permanent purchase. Rent an accelerator for the afternoon, hand it back, pay for the hours.
Price history
Median verified $/h per GPU class, one reading per day from the same marketplace snapshot the table above uses. Lines appear once a class has three days of readings.
Break-even
How many rented hours equal the price of buying a machine outright. Rented accelerators are not like-for-like substitutes for a personal AI system — they are faster but temporary, and your data leaves your building — so read this as a budget comparison, not a spec comparison. Reference purchase: NVIDIA DGX Spark at its published MSRP of $4,699.
| GPU | Median $/h | Rented hours to match a purchase |
|---|---|---|
| RTX 4090 | $0.40 | 11,726 h≈ 28 years at 8 h/week |
| RTX 5090 | $0.51 | 9,206 h≈ 22 years at 8 h/week |
| A100 | $0.91 | 5,140 h≈ 12 years at 8 h/week |
| H100 | $3.07 | 1,532 h≈ 4 years at 8 h/week |
| H200 | $4.61 | 1,020 h≈ 2 years at 8 h/week |
| B200 | $6.00 | 783 h≈ 23 months at 8 h/week |
The purchase price is NVIDIA's published MSRP. It excludes electricity and, on the other side of the ledger, any resale value you would recover — owning is therefore somewhat cheaper than this table implies, renting somewhat less attractive at high hours. Hourly rates come from the collected price bands above.
Rent when
- + The job is occasional — a fine-tune, a batch run, one big evaluation.
- + You need far more memory than you own, briefly.
- + You want to try a model class before committing to hardware.
Buy when
- + You run models daily — owned hardware has no marginal cost.
- + Your data must not leave your machine.
- + Latency matters and you do not want to upload anything.