Modelling GPU rental prices
GPU rental prices go up and down. There are small day-to-day changes, longer trends and periods when prices move more sharply. I’m interested in these patterns and what they might tell me about demand picking up or capacity becoming harder to find, and the reasons behind price changes.
For this article, there is a good example to look at in May 2026. RTX 3090 asking prices on Vast.ai rose sharply, then fell back in June. And there was a dataset on Hugging Face with data on these listings. So I wanted to explore it. I don’t know exactly what drove the move, but there is data to work with. How can I wrangle it and play with it to get a clearer picture of what happened?
A price baseline
Before looking at the bigger price changes, it helps to get a sense of how GPU rental prices normally move. I compare each listing’s daily median asking price with the day before: a move from $0.20 to $0.22 an hour is a 10% rise. I combine these price ratios using a geometric mean, giving each listing equal weight. Linking the daily changes produces the chained Jevons index shown in black, with April 30 set to 100.
I then compare this index with the slowly adjusting baseline shown as the dashed line. The shaded band is based on April’s variation around it, roughly 2.8% either side, providing a reference for judging the later sharp increase in price.
Calculating the baseline
I use March and April as the reference for usual price variation. The question is whether later prices move substantially beyond that range and stay there.
The baseline uses exponential smoothing on the log index, giving more weight to recent prices. I limit its adjustment to about 0.26% a day, so a sudden increase stays visible against the earlier price level. The index records the increase in full.
It is today’s index; Bt is today’s baseline. The adjustment uses their ratio on a log scale. α sets the speed, and clip limits the gap to ±c.
March and April set the adjustment parameters (α ≈ 0.094, c ≈ 0.027) and the band’s width. These parameters stay fixed from May onward. The baseline can move, while the band keeps the same percentage width.
The May price rise
In mid-May, prices begin to leave that range. On May 30, the index reaches 240.1, while the baseline is 105.4: 128% above its baseline.
The main period above the range runs from May 14 to June 13, 31 days, using three consecutive days outside the band to identify a sustained change. Prices then fall back towards the baseline.
Repricing across sellers
I measure seller participation by comparing each listing’s price with its April median, then taking the median change for each seller. A listing needs at least five days of April quotes; each seller counts once.
On May 30, 39 of 40 sellers were more than 10% above their April prices. Raising the threshold to 100% still leaves 34 of 40. For most of these sellers, prices had more than doubled.
The price rise lasted several weeks and involved most of the sellers that were tracked. Prices increased on individual listings, so cheaper offers disappearing cannot explain the rise on its own. Each search returns at most 64 listings, so this describes the sellers in the sample, but should still be indicative of the wider price movement.
Pearl mining
Miners were renting GPUs to earn Pearl’s PRL token during the price rise. Tom’s Hardware reported on May 31 that miners were using RTX 4090 and 5090 instances on Vast.ai and RunPod, while a June study also reported people earning PRL on rented RTX 3090s. Alongside these reports, network hash rate shows the increase in mining activity, which is a big tell that mining PRL was directly linked to renting listings on Vast.ai.
So in a situation where PRL earnings can cover the rental cost and fees, miners have a reason to rent available listings, creating a strong incentive and big demand for the GPUs listed. As more listings are rented, buyers are left competing for less available capacity. In turn, that gives sellers an incentive to raise their asking prices. I think it’s clear that this demand from Pearl miners very likely contributed to the broad repricing in May. And while this was a sharper surge, we now also see more sustained demand affecting overall prices since the summer. So this is not a one-off event, in my opinion, and needs to be understood and studied more.
In addition, other GPU rental prices were rising too. AIMC’s report for May 25–31 recorded average on-demand prices up 4.9% for H100 SXM and 8.8% for H200 SXM. Very interesting.
What comes next
Thanks to MarcusLammers for the dataset behind this study. A next step would be to monitor live GPU rental listings myself, collecting prices and availability over time. This would make it possible to look for unusual patterns, anomalies and so on. The RTX 3090 is a good starting point, but more interesting would be to expand the scope to H100s, H200s and other accelerators, and add more marketplaces. With time, we could understand typical price movements, infer where and when participants rent GPUs, and map out the usual and unusual patterns to better understand the GPU rental market and, more broadly, the compute market itself.
Data: MarcusLammers, CC BY 4.0.