Cloud Providers Price GPUs at Nearly Double Rates, Fluence Finds

The AI Compute Race report prices GPU compute outside the biggest buyers: a three-year commitment cuts the median price by 48%, while 17 of 22 tracked GPU models rose in price over the past year

DALLAS, Texas – November 13, 2026 – Fluence has released The AI Compute Race, detailing GPU cloud compute costs for organizations that are neither frontier labs nor hyperscalers. Created alongside research partners NEAR AI, io.net, Boundless, Phala, Gensyn, Fortytwo, Theta, Compute Labs, and GAIB, the study reveals that purchasing location dictates cost. For H100, H200, and B200 models, hyperscalers list hourly rates at 1.98 times the GPU-cloud median, ranging from 1.4 times for the B200 to 2.8 times for the H100. Furthermore, an H200 ranges from $4.08 to $10.60 per hour depending on the listing.

Provider choice is nearly matched in importance by commitment duration. Compared to individual on-demand rates, a one-year reservation decreases the median price by 21%, while a three-year commitment reduces it by 48%. Meanwhile, spot capacity for interruptible workloads lists at 53% of the standard on-demand price.

Budgets may require adjustment for buyers anticipating declining GPU costs. GetDeploying tracked 22 GPU models and found that 17 of them became more expensive over the course of a year, with the median H100 renting for 13% to 30% higher depending on the timeframe measured.

The study outlines several drivers behind the escalating prices:

• Demand shifted toward open models. CNBC reports that the share of OpenRouter tokens held by Chinese models among companies across the US, Europe, and the Global South jumped from 6–13% to 57–67% in less than eight months. Open weights enable any enterprise possessing GPUs to deploy a near-frontier model on its own hardware.
• Newest hardware was secured by major buyers. Projected 2026 capital expenditures by the four largest hyperscalers total $720–740 billion, whereas the four largest disclosed deals between AI labs/Meta and neoclouds amount to roughly $233 billion—with three representing “up to” figures and a significant portion allocated to undelivered capacity.
• GPU-hours are transforming into financial instruments. CME, ICE, and Nodal Exchange have introduced GPU futures tied to distinct price indices. Concurrently, lenders are issuing loans backed by GPUs, exemplified by Lambda’s investment-grade-rated loan at 6.78% and Sharon AI’s unrated loan at 9.95%.

“Open-weight models are close behind the best closed ones, and companies are switching to them. Any company with GPUs can now serve a near-frontier model on hardware it controls. The GPUs are the hard part,” stated Evgeny Ponomarev, Co-founder of Fluence.

To reduce compute expenses, the report outlines four strategies: evaluate identical chips across various provider categories, secure long-term commitments for consistent workloads, align the specific GPU generation with the corresponding model, and look beyond hyperscalers because much recent neocloud capacity is booked by labs years in advance.

The complete study can be accessed at fluence.ai/compute-race-2026.

About Fluence

Fluence Cloud is a GPU cloud provider sourcing capacity from independent data centers (fluence.ai). Established in 2017 by Dmitry Kurinskiy, Evgeny Ponomarev, and Tom Trowbridge, the organization has secured $14 million in funding from backers including 1KX, Multicoin, Distributed Global, Blockchange, Tiger Global, and Protocol Labs.

Short FAQs

What is the main finding of The AI Compute Race report?
Where a company purchases a GPU-hour determines its price, with hyperscalers listing GPUs at 1.98 times the GPU-cloud median across H100, H200, and B200 models.

How do commitment lengths affect GPU cloud prices?
Compared to on-demand rates, a one-year reservation decreases the median price by 21%, a three-year term reduces it by 48%, and spot capacity lists at 53% of on-demand pricing.

Are GPU prices falling or rising?
Prices are largely rising; 17 out of 22 tracked GPU models cost more than they did a year prior, with the median H100 renting for 13% to 30% more.