Mentioned in This Article
1. GPU rental prices rise across every tracked chip
GPU rental prices climbed for every Nvidia chip generation tracked by Ornn's Compute Price Index in the 30 days to November 11, 2026. The Blackwell-based B200 settled at $9.84 per GPU-hour, up 32.1%, while the Hopper-based H200 jumped 48.8% to $7.75. The older H100 rose 8.3% to $3.25 and the A100 gained 11.6% to $1.06.
The readings point to a market in which the cost of renting AI computing power is moving higher across the board, with the sharpest increases concentrated in the most advanced hardware. For developers, startups and enterprises that rent rather than own their chips, the shift translates directly into higher bills for training and running AI models.
The figures come from the Ornn Compute Price Index, known as OCPI, a set of benchmarks created by Ornn AI Inc. to measure what customers actually pay to rent graphics processing units by the hour.
2. What are GPU rental prices right now?
As of the November 11, 2026 settlement, renting one Nvidia B200 cost about $9.84 per hour, an H200 about $7.75, an H100 about $3.25 and an A100 about $1.06, based on Ornn's index of executed transactions. Those are hourly prices for a single GPU, and they are the benchmark figures for each chip model, not quotes from any one provider.
Translated into longer periods, the numbers add up quickly. At $9.84 per hour, a single B200 running around the clock costs roughly $236 per day and more than $7,000 over a 30-day month. Large AI workloads rarely run on one chip; training runs and inference clusters often use hundreds or thousands of GPUs at once, so a change of a few dollars per hour can move total costs by millions over the life of a project.
The 30-day changes were uneven. The H200's 48.8% rise was the largest move in the group, followed by the B200's 32.1% gain. The A100, the oldest chip in the index, rose 11.6%, slightly more than the H100's 8.3% increase.
3. The B200 and H200 lead the price surge
The steepest gains belong to the two newest chips in the index. The B200 is part of Nvidia's Blackwell generation, its latest data center architecture, while the H200 is an upgraded member of the Hopper family that preceded Blackwell. Both are in demand because they offer more memory and faster processing for the large models that dominate current AI development.
The H200's move stands out. A nearly 49% increase in a single month is a sharp repricing for any commodity, and it lifted the H200 to more than double the hourly price of the H100, the chip it was designed to succeed. Both belong to the Hopper line, but the H200 carries substantially more high-bandwidth memory, which makes it better suited to running very large models.
The B200's 32.1% gain kept it firmly at the top of the price ladder. At $9.84 per hour, it costs roughly nine times as much to rent as an A100, a gap that reflects how much faster and more capable each new generation of AI hardware has become.
4. Older Nvidia chips get more expensive too
Prices for previous-generation hardware also moved higher, though by smaller margins. The H100, which became the defining AI chip of the past several years, rose 8.3% to $3.25 per GPU-hour. The A100, an even older design from the Ampere generation, climbed 11.6% to $1.06.
The fact that every tracked chip became more expensive at the same time is significant. If the increases were limited to the newest hardware, they could be explained by a scramble for scarce cutting-edge supply. Rising prices for older chips suggest that demand for rented compute is broad enough to lift the whole market, including hardware that is no longer at the frontier.
Older chips still have real uses. Many inference tasks, smaller models, fine-tuning jobs and research projects run well on H100 and A100 hardware, and customers who cannot secure or afford newer chips often turn to previous generations instead. Higher prices at the low end of the ladder suggest that this fallback option is getting more expensive as well.
5. How does the Ornn Compute Price Index work?
The OCPI measures GPU rental costs using volume-weighted averages of executed transactions, rather than surveys of providers or the list prices they advertise. Ornn's methodology counts only completed rentals between verified counterparties, so each index reflects prices that customers actually paid.
That approach addresses a long-standing problem in the market for AI computing power: a lack of transparent pricing. Cloud providers publish rate cards, but large customers often negotiate private discounts, and smaller providers may change prices frequently or not publish them at all. A benchmark built from completed deals aims to show where the market is actually clearing.
The data draws heavily on the so-called neocloud market, the newer generation of cloud companies that specialize in renting out GPU capacity for AI workloads rather than offering the broad range of services sold by the largest cloud platforms. Ornn's indices cover on-demand rentals, while reserved and long-term agreements are tracked separately. Several Ornn tickers are available on Bloomberg, giving financial professionals a way to follow the readings alongside other market data.
6. Renting versus owning AI hardware
A large share of AI work runs on rented hardware rather than chips that a company buys outright. Purchasing data center GPUs requires heavy upfront spending, along with the power, cooling and networking needed to run them, and buyers can wait a long time for delivery of the newest models. Renting by the hour lets customers start work immediately and pay only for what they use.
That flexibility comes with exposure to market prices. A company that owns its hardware has a largely fixed cost once the purchase is made, while a company that rents on demand pays whatever the market charges at the moment it needs capacity. When rental prices rise as quickly as they did over the past 30 days, on-demand users feel the change right away.
The unit used in Ornn's index, the GPU-hour, captures this directly. It represents one graphics processing unit running for one hour, so a cluster of 1,000 B200 chips running for a full day consumes 24,000 GPU-hours, and every dollar added to the hourly benchmark adds tens of thousands of dollars to that single day's bill.
7. Why GPU rental prices matter beyond the AI industry
Computing power has become one of the most important inputs in the technology economy, and its price affects who can build AI products and how quickly. When rental costs climb, well-funded companies can absorb the increase more easily, while smaller developers and research groups may have to scale back experiments or wait for capacity.
Rising prices also carry signals for investors. GPU rental rates offer a real-time read on demand for AI infrastructure, a theme that has driven large swings in technology stocks and has drawn the attention of companies across the digital asset sector, including bitcoin miners that have repurposed data center capacity for AI hosting. A benchmark that shows rental prices moving up across every chip generation provides a data point on how strong that demand remains.
For buyers of compute, the readings underline the value of planning ahead. Customers relying on short-term, on-demand rentals are the most exposed to sudden repricing, while those with longer contracts have more predictable costs. A transparent price index makes it easier to compare offers and decide when locking in capacity makes sense.
8. Who is Ornn, the company behind the index?
Ornn is a young company, founded in 2025 by a team of MIT graduates with the stated aim of bringing transparent pricing to the market for GPU rentals that power AI workloads. In addition to its indices, the company operates a marketplace where GPU capacity can be bought and sold.
The startup raised $33 million in a seed round in June 2026, a large sum for a company at such an early stage and a sign of investor interest in treating computing power as a tradable commodity. Ornn's pitch is that as adoption of Blackwell hardware spreads, buyers and sellers increasingly struggle to see where supply and demand actually sit, and that a trusted reference price can fill that gap.
The company positions its data as a benchmark for pricing, hedging and financing in compute markets, and says neocloud operators and financial firms are increasingly using it for those purposes. The comparison to established commodity markets is deliberate: in oil, metals or electricity, widely followed benchmark prices allow producers and consumers to set contracts, manage risk and raise financing against future output.
9. What to watch in the GPU rental market
The next question is whether the latest increases mark a lasting shift or a short-term spike. Because Ornn's indices are updated continuously from new transactions, the coming weeks of readings will show whether B200 and H200 prices hold near their current levels or ease back as more capacity comes online.
The relationship between generations will also be worth following. The gap between the B200 and the A100, and between the H200 and the H100, shows how strongly customers are paying up for newer hardware. Any narrowing would suggest that supply of advanced chips is catching up with demand, while a widening gap would point to continued scarcity at the top of the market. For now, the data shows a market in which every tier of AI computing power has become more expensive to rent.