NovaPodThe RTX 5090 rental price is the query I keep seeing, because it is the largest cloud-GPU market and the street card is still a mess. I checked this marketplace on November 4, 2026, then set those asks next to public indexes from the same two days.
A rental hour and a boxed card are not the same product. The numbers below are asks, not a contract rate, and they move. If you need the card on your desk, read yesterday's street-price note. If you need 32 GB for a job, start here.
What the RTX 5090 rental price is today
On NovaPod the cheapest RTX 5090 ask was $0.26 an hour. The median ask was $0.46. There were 648 offers, each with 32 GB. That is the live list, billed per second from a crypto balance, not a monthly seat.
GridCompute's read of Vast.ai at 05:00 UTC on November 4 put the median of 227 on-demand listings at $0.71 an hour. The cheapest listing there was $0.45, and the middle half sat between $0.60 and $0.87. They had RunPod at $1.19, and a seven-day average median of $0.70.
A separate index, updated November 3, had the on-demand median at $0.72 across providers, with a cheapest in-stock on-demand rate of $0.35 and a 90-day rise of about 20% in that median. Spot on that same page was not cheaper: the spot median was $0.85. Our $0.26 floor is a marketplace low, not the price every host charges.
| Source | Low | Median | What it counts |
|---|---|---|---|
| NovaPod, Nov 4 | $0.26/hr | $0.46/hr | 648 offers, 32 GB |
| Vast.ai via GridCompute, Nov 4 | $0.45/hr | $0.71/hr | 227 on-demand listings |
| Provider index, Nov 3 | $0.35/hr | $0.72/hr | On-demand across providers |
H100 rental price, for the jobs 32 GB cannot hold
People searching an RTX 5090 rental price are often one step from an H100 rental price. On this list the H100 SXM started at $1.21 an hour, median $1.68, with 65 offers and 80 GB. The H200 started at $2.76, median $2.76, 82 offers, 141 GB. The B200 started at $4.38, median $6.30, 51 offers, 192 GB.
Public posted medians are higher than our floor. One November index had H100 on-demand from about $1.80 and a provider median near $3.43. A September cloud index had the H100 on-demand median at $3.25 and the H200 at $4.40. Use those as the neighborhood, then read the card you can actually launch.
32 GB on a 5090 is enough for a 7B–14B model, a short fine-tune, and a lot of image work. It is the wrong box for a 70B that needs the weights resident. That is when the H100 rental price, or the H200, is the relevant number.
RTX 4090 rental price, and the rest of the cheap tape
The RTX 4090 on this marketplace started at $0.09 an hour, median $0.35, with 419 offers and 24 GB. Public on-demand rates for a 4090 were still being quoted from about $0.35, with a provider median near $0.57. The 4090 is the volume card under the 5090. It is not a substitute when the job wants 32 GB.
The A100 SXM4 here started at $0.30, median $0.46, 158 offers, 80 GB. That is close to the 5090 median and buys more than twice the memory. If the model is memory-bound and the interconnect does not matter, look at the A100 before you pay 5090 scarcity on the street.
Why the RTX 5090 price in shops is a different story
Retail samples do not agree with each other, and none of them agree with the rental tape. TechSpot's October survey, which we wrote up on November 3, put the cheapest in-stock US RTX 5090 at $9,400 against a $1,999 MSRP. A November 2 report had the Founders Edition showing $1,999 and out of stock, with partner boards far above sticker. One writeup put those partner boards near $3,340.
I am not going to average $3,340 and $9,400 into a fake street price. The useful fact is that a new 5090 is either missing or expensive, while an hour of one is still under a dollar on every index I checked this week. Buy the card only if you need it on the desk for months. Rent it if the job ends.
How to read an RTX 5090 rental price
Compare memory first, then the ask, then the number of offers. A $0.26 floor with 648 offers is a real market. A single $0.35 host on another site is a quote. Storage, bandwidth, and a minimum GPU count are often extra. GridCompute's cheapest Vast row wanted two GPUs.
Stop the pod when the batch is done. Per-second billing only helps if the clock actually stops. For the launch path, open the RTX 5090 on the live list, or the H100 SXM if 32 GB is the thing that will fail.