Hyperscalers charge nearly double the going rate for GPU hours, Fluence finds

Where a company buys a GPU-hour decides what it pays. That's the core finding of The AI Compute Race, a Fluence report that prices cloud compute for everyone outside the frontier labs and hyperscalers. Across H100, H200 and B200 chips, hyperscalers list at 1.98 times the GPU-cloud median, from 1.4 times on the B200 to 2.8 times on the H100. The same H200 rents for anywhere between $4.08 and $10.60 an hour depending on the seller.

Committing helps almost as much as shopping around. Measured against each provider's own on-demand rate, a one-year reservation cuts the median price by 21%, and a three-year term cuts it by 48%. Spot capacity lists at 53% of on-demand, for workloads that can take an interruption.

Anyone betting on cheaper GPUs should redo their budget. Seventeen of the 22 GPU models tracked by GetDeploying cost more than a year earlier, and the median H100 rents for 13% to 30% more depending on the window measured.

Fluence co-founder Evgeny Ponomarev pins the squeeze on open-weight models closing in on the best closed ones. Companies are switching to them, he said, which means any firm with GPUs can serve a near-frontier model on hardware it controls. His conclusion: the GPUs are the hard part.

The report closes with four ways to pay less for the same compute: compare the same chip across provider types, commit when the workload is steady, match the GPU generation to the model, and look beyond the hyperscalers, since much of the newest neocloud capacity is already contracted to labs years ahead. The full report is at fluence.ai/compute-race-2026.

Fluence Cloud supplies GPU capacity out of independent data centers. Founded in 2017 by Dmitry Kurinskiy, Ponomarev and Tom Trowbridge, the company has raised $14 million from investors including 1KX, Multicoin, Distributed Global, Blockchange, Tiger Global and Protocol Labs. The report was produced with research partners NEAR AI, io.net, Boundless, Phala, Gensyn, Fortytwo, Theta, Compute Labs and GAIB.