Dynamic Pricing Brings Price Discovery to GPU Spot Rental Markets
More than half of GPU spot price series did not change once in nine days. Nebius's new dynamic spot pricing shows what a price that moves with capacity reveals.
Data window: October 8 to 16, 2026. Every figure in this post is a spot or preemptible price, in USD per GPU-hour.
On October 8, Nebius moved its preemptible GPU capacity to dynamic pricing: a spot price that resets with available capacity and demand, updated as often as every 15 minutes. Current prices and recent history are public: Nebius's website shows seven days and its API, which needs no account, shows 30 (Nebius docs). Hyperscalers have published spot prices for years, but dynamic spot pricing is still rare among neoclouds. We wanted to understand what a dynamic price adds, and Nebius shared its quote data with us. Those figures come from Nebius's pricing API, sampled every 15 minutes and averaged by day, not from the daily collection we use for every other seller; the method note at the end gives details.
Silicon Data measures the GPU rental market every day across on-demand, reserved, and spot or preemptible capacity. Our spot coverage spans hyperscalers, neoclouds and GPU marketplaces. Spot is the thinnest layer: the major hyperscalers all offer it, but most neoclouds sell only on-demand and reserved capacity. In principle it is also the layer most responsive to supply and demand. Over the nine days from October 8 to 16, we asked two questions: how often spot prices actually change, and what a dynamic price shows that a posted one cannot.
Key takeaways
- Most spot prices rarely change. More than half of the spot price series we track did not change once in nine days, and three quarters changed on two days or fewer.
- Dynamic prices can show where capacity is loose or tight. A price that resets with capacity reveals conditions a posted price hides. In Nebius's first nine days, its H100 price held steady while its H200, B200 and B300 prices moved on most days.
- Better prices make better benchmarks. Silicon Data combines posted prices with transaction data from neocloud data partners. Frequently updated prices from any provider add to the price discovery that indices, forward curves and futures rely on.
Terms used in this post
- Spot (or preemptible): discounted capacity the provider can interrupt and reclaim at short notice. Providers use either name, and so do we. Spot describes the terms of service, not how the price is set: a spot price can be posted or dynamic.
- On-demand: capacity rented by the hour at a listed rate, with no commitment, that the provider will not interrupt.
- Reserved (or term contract): capacity committed for a fixed period, usually months to years, at a negotiated price. Most GPU capacity is sold this way.
- Posted price: a listed price that the provider changes occasionally, by hand or on a schedule.
- Dynamic price: a price set by an algorithm and updated frequently as available capacity and demand change. Nebius's new spot pricing is an example.
- Transaction price: what a buyer actually paid, which can differ from the listed price.
- Neocloud: a GPU cloud provider other than the hyperscalers (AWS, Microsoft Azure, Google Cloud and Oracle). We use provider and seller interchangeably.
- Market median: the middle spot price across the sellers in our coverage on a given day, excluding Nebius.
How often spot prices change
We counted, for every seller and GPU in our spot coverage, how many of the eight day-to-day steps in the window brought a price change of more than 0.5%.
The two rows are not measured the same way. Each seller in the top row is one price observation per day. Nebius's row is the daily average of its 15-minute quotes, so any intraday move shows up as a change in the daily figure.
More than half of the spot price series we track held the same price for all nine days, and three quarters changed on two days or fewer. Prices like these are set and revisited occasionally rather than reset as capacity changes; for many sellers a stable posted price can be a rational choice given operational cost, customer preference for predictability, or simply thin spot volume. The series that moved on most days came from two sellers. One is a marketplace, where prices are set by many hosts. The other rose by about 2% each day on every GPU. Nebius's dynamic price changed on most days for H200, B200 and B300. Its H100 price sat at $0.79, the floor of Nebius's H100 spot range, for most of the window after the first day or two; Nebius offers H100 spot capacity in a single region, eu-north1.
What dynamic prices add
Nebius replaced a fixed preemptible price with one set per GPU and region from available capacity and real-time demand, updated as often as every 15 minutes and bounded by a price floor and a ceiling one cent below on-demand (Nebius announcement). Customers can follow the spot price or set a maximum price; if the spot price rises above it, the instance is stopped and can be restarted later.
AWS, Azure and Google Cloud already publish spot prices, but they are smoothed: AWS replaced auction-based spot pricing with gradual adjustments in 2017, and the hyperscaler spot prices in our coverage changed at most once over the window. The difference is design: hyperscalers smooth their spot prices, while Nebius's can update as often as every 15 minutes.
Placed against the rest of our spot coverage in the charts below, Nebius's daily quote sat below every other seller's on most of the nine days for each GPU: every day for H100, eight days for B200, six for B300 and five for H200. H100 held at its $0.79 floor for most of the window, about two thirds below the market median. H200 and B200 moved day to day, mostly below the median. B300, the newest GPU, started below the market and rose through the week, finishing above the median on the last three days.
| GPU | Market spot median | Market spot average | Nebius spot avg | Nebius vs median | Nebius days with a price change | Days below median | Days below all other sellers |
|---|---|---|---|---|---|---|---|
| H100 | $2.33 | $2.65 | $0.80 | −66% | 2 of 8 | 9 of 9 | 9 of 9 |
| H200 | $2.92 | $4.47 | $2.00 | −32% | 7 of 8 | 9 of 9 | 5 of 9 |
| B200 | $4.01 | $4.33 | $2.09 | −48% | 8 of 8 | 9 of 9 | 8 of 9 |
| B300 | $4.95 | $5.82 | $4.58 | −7% | 8 of 8 | 6 of 9 | 6 of 9 |
Market figures exclude Nebius. Medians and averages are across the nine days.
A posted price cannot show that pattern. A price that resets with capacity can, and here it is consistent with older GPUs being in ample supply and the newest in strong demand. Nine days from one dynamic seller is enough to describe price levels and how often prices change. It is not enough to measure volatility or how these prices respond to the rest of the market, and the B300 figures are the most sensitive to any single seller given thinner coverage.
Nebius against the spot market, GPU by GPU
Each bar spans that day's lowest and highest spot quote in Silicon Data's coverage, excluding Nebius. The blue line is the market spot median; the white line is Nebius's daily spot average, labeled with its gap to the median. A label of −66% means Nebius's daily average was 66% below that day's market median; a positive label means it was above. Our rental price indices draw on every rental type, converted to a common on-demand basis; this post looks at the spot layer on its own.
H100
Market median $2.33 per GPU-hour over the nine days; quotes ranged from $1.38 to $4.49.
H200
Market median $2.92 per GPU-hour over the nine days; quotes ranged from $2.20 to $12.05.
B200
Market median $4.01 per GPU-hour over the nine days; quotes ranged from $2.70 to $7.02.
B300
Market median $4.95 per GPU-hour over the nine days; quotes ranged from $4.53 to $11.06.
The trade-offs of dynamic pricing
Dynamic pricing also has costs. Buyers give up budget certainty: the hourly rate can change several times a day, and a workload with a maximum price is stopped when the spot price rises above it. Managing that takes tooling and attention a fixed rate does not. Sellers take on less predictable revenue and more pricing infrastructure to run. These trade-offs are part of why posted prices remain the norm.
Why this matters for benchmarks
The charts above also show what an independent benchmark makes possible: any provider, posted or dynamic, can see where its price sits against the market each day, and so can buyers. A floor, a ceiling or a list price is easier to set well when the market median is visible.
Silicon Data's indices combine the prices providers post with transaction data from our neocloud data partners: posted prices show what buyers are quoted, and transactions show what capacity actually rents for. Dynamic prices offer a third view: a frequently updated price that responds to capacity as it changes. As they become more common, they give anyone building benchmarks a cross-check and give forward curves and futures more of the price discovery they depend on.
Dynamic, transparent spot pricing is a key ingredient in price discovery for compute, even as most capacity continues to be sold on term contracts and spot remains the residual balancing layer. Nebius is among the first neoclouds to price spot capacity this way, and broader adoption across providers would add to price discovery across the market.
Track GPU spot prices with Silicon Data
This post is built on Silicon Data's daily collection: the GPU Pricing Dataset for spot, on-demand and reserved quotes across GPU models, and SiliconIndex for daily benchmarks such as the H100, H200, B200 and B300 rental price indices, all available via API.
Method and disclosure
Method. Market figures are Silicon Data's daily spot and preemptible prices, in USD per GPU-hour; no on-demand or reserved prices enter any figure in this post. Spot sellers are a small subset of the 80-plus platforms in our daily coverage and include two hyperscalers as well as neoclouds and GPU marketplaces; the number varies by GPU and is smallest for B300, where the market statistics are most sensitive to any one seller. We lead with the median because a few sellers post spot prices far above the rest, which pulls the average up; both are shown in the table. Nebius figures are the spot estimate from its Calculator API for a 1-GPU preemptible preset, sampled every 15 minutes to match Nebius's update frequency and averaged per day across the regions that offer the GPU (H100: eu-north1; H200: eu-north1, eu-west1, us-central1; B200: me-west1, us-central1; B300: eu-west2, uk-south1, us-north1). A price change is a day-to-day move of more than 0.5% in the daily price; for Nebius this is the daily average of its intraday quotes. All figures are offered prices, not realized rentals. Nebius sets its prices with its own algorithm; Silicon Data has no role in that pricing. Nebius confirmed to us that $0.79 is the floor of its H100 spot range. Public price access is described in Nebius's documentation. Program details are from Nebius's announcement of September 22, 2026. Our own collection of Nebius prices is excluded from the market figures so Nebius is not counted twice.
Disclosure. Nebius provided the quote data used in this post. Silicon Data and Nebius are in discussions about broader data sharing. Nebius did not review or approve the analysis.