AI Compute RadarElastic 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.
Radar Rent Index
What rental prices did
One line per GPU class from the same hourly marketplace snapshots as the table: this week's median against last week's, and against the first week we collected. Described, never forecast.
| GPU | This week | vs. last week | Since start | Offers | Day band $/h |
|---|---|---|---|---|---|
| RTX 4090latest day Sep 18 | $0.54median of 6 days | +35 %up sharply | +44 %since Aug 26 · 24 days on record | 64supply steady | $0.27–$0.61min–p75 of the latest day |
| RTX 5090latest day Sep 18 | $0.68median of 6 days | +26 %up sharply | +40 %since Aug 26 · 24 days on record | 64supply steady | $0.40–$0.77min–p75 of the latest day |
| A100latest day Sep 18 | $0.82median of 7 days | −10 %down | −12 %since Aug 26 · 24 days on record | 21fewer offers than last week | $0.33–$0.94min–p75 of the latest day |
| H100latest day Sep 18 | $2.80median of 7 days | 0 %flat | 0 %since Aug 26 · 24 days on record | 11supply steady | $1.90–$5.37min–p75 of the latest day |
| H200latest day Sep 18 | $4.61median of 7 days | 0 %flat | +9 %since Aug 26 · 24 days on record | 5fewer offers than last week | $3.29–$4.74min–p75 of the latest day |
| B200latest day Sep 18 | $8.14median of 7 days | +8 %up | +35 %since Aug 26 · 24 days on record | 1fewer offers than last week | $5.63–$8.75min–p75 of the latest day |
Excluded as marketplace glitches (a day more than twice or less than half the surrounding days): RTX 4090 Sep 12, RTX 5090 Sep 12. A 30-day change appears once thirty days are on record, from Sep 24.
A day is the median of its hourly medians, counting only hours with at least five verified offers. A week is the median of its days and needs at least four of them. Words follow fixed thresholds: up or down from ±8 %, sharply from ±20 %; offers up or down from ±15 %. How the index is made →
Price history
Robust daily medians per GPU class: the median of each day's hourly medians, counting only hours with at least five verified offers. Days the Rent Index excludes as marketplace glitches are left out. 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 | $0.53 | 8,792 h≈ 21 years at 8 h/week |
| RTX 5090 | $0.65 | 7,258 h≈ 17 years at 8 h/week |
| A100 | $0.71 | 6,601 h≈ 16 years at 8 h/week |
| H100 | $3.14 | 1,497 h≈ 4 years at 8 h/week |
| H200 | $4.74 | 991 h≈ 2 years at 8 h/week |
| B200 | $7.38 | 637 h≈ 18 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.
The guide: buy or rent a GPU? →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.