The capacity squeeze that has been pushing up VPS prices all year sits at the other end of the market too, on the GPU servers that AI workloads run on. One of the larger listed GPU clouds, Nebius, has just moved its on-demand prices again — the second time it has done so since the spring — and the new rates are a useful marker for anyone pricing an AI server rather than a plain VPS.
What changed
From October 1, 2026, Nebius raised its on-demand, pay-as-you-go rates across its four current Nvidia platforms. On its own published price list the per-GPU-hour numbers now read: H100 NVLink at $4.50, up from $3.85 (about 17%); H200 NVLink at $5.40, up from $4.50 (20%); B200 NVLink at $8.50, up from $7.15 (about 19%); and B300 NVLink at $9.50, up from $7.85 (21%). CPU-only servers were caught in the same round — the AMD EPYC Genoa vCPU now lists at $0.015 per CPU-hour, roughly a quarter more than before. Separately, from October 8 the provider moved preemptible capacity for these platforms onto dynamic spot pricing, so the cheap interruptible tier no longer has a fixed rate.
These are list prices for on-demand use. Committed and reserved terms are quoted separately and run well below the hourly rate; the increase lands hardest on buyers who pay by the hour with no commitment.
What it costs on the bill
The hourly figures are small, but GPU servers run continuously and in multiples, so they add up fast. A single H100 left running around the clock for a 730-hour month now costs about $3,285 on demand, against roughly $2,810 before October 1 — close to $475 a month more for the same card. Nvidia's NVLink parts are usually taken eight to a node, and an 8×H100 box now runs $36 an hour, or about $26,280 a month flat out, up from near $22,480 — a jump of roughly $3,800 a month per node. On the newest silicon the numbers are larger again: a B300 is $9.50 an hour, so an eight-GPU B300 node is $76 an hour, more than $55,000 a month.
The longer trend is the sharper one. In early May the same H100 listed at $2.95 an hour; at $4.50 it is about 53% higher in roughly five months, and the B300 has moved from $6.10 to $9.50 over the same stretch. For a team that budgeted an H100 at spring rates, the single-card line has grown by more than $1,100 a month.
Why it is going up
Nebius has not tied the increase to the cost of any one component. Its own executives have described a market where demand outruns supply: the company has said it sold capacity as fast as it could bring it online, and that it could sell its entire 2027 capacity on current deal terms. That is the same story driving the memory and storage shortages we have covered from the VPS side — capacity, power and advanced-packaging supply are the bottleneck, and when a provider cannot build fast enough, the on-demand price is the lever it pulls first. A hardware generation that is nominally faster per dollar does not help if you cannot get time on it at the old rate.
What to do if you are choosing a GPU server
On-demand is now the most expensive way to run a steady AI load, and the gap to committed pricing is wide enough to change the decision. Treat the hourly rate as the price of flexibility: right for experiments, short fine-tuning runs, and bursty inference where idle time would otherwise be wasted. For anything that runs most of the month — production inference, long training jobs — a reserved or committed term, or spot capacity with checkpointing, is usually far cheaper per useful hour, and the premium for on-demand has only grown with this round.
The other move is to compare before committing. Per-GPU-hour rates across the GPU clouds still vary by a factor of two or more for the same card, and the headline price is only part of it — egress, storage and the minimum node size all land in the bill. Our AI workload page sets out how to size the server to the model rather than to the marketing, and the catalog lets you line providers up on price per resource instead of per press release.
Figures in this piece are from Nebius's published compute pricing page and from reporting by The Motley Fool on the October 1 increase, cross-checked against the provider's own effective-date notice.