The AI Compute Race report costs GPU compute outdoors the most important patrons: a three-year dedication cuts the median worth by 48%, whereas 17 of twenty-two tracked GPU fashions rose in worth over the previous yr
DALLAS, Texas – November 13, 2026 – Fluence has printed The AI Compute Race, a report that costs GPU cloud compute for each firm that’s not a frontier lab or a hyperscaler. It was produced with analysis companions NEAR AI, io.net, Boundless, Phala, Gensyn, Fortytwo, Theta, Compute Labs and GAIB. Its central discovering: the place an organization buys a GPU-hour units the worth. Throughout H100, H200 and B200 fashions, hyperscalers checklist GPUs at 1.98 occasions the GPU-cloud median, from 1.4 occasions on the B200 to 2.8 occasions on the H100. The identical H200 lists from $4.08 to $10.60 an hour.
Dedication size issues virtually as a lot as supplier alternative. In opposition to every supplier’s personal on-demand charge, a one-year reservation cuts the median worth by 21% and a three-year time period by 48%. Spot capability lists at 53% of on-demand, for workloads that may be interrupted.
Patrons who count on GPU costs to fall could must recheck their budgets. Seventeen of the 22 GPU fashions tracked by GetDeploying value greater than a yr earlier, and the median H100 rents for 13% to 30% extra, relying on the window measured.
The report additionally traces why costs are quickly rising:
- Demand moved to open fashions. Chinese language fashions’ share of OpenRouter tokens amongst firms within the US, Europe and the International South rose from 6–13% to 57–67% in beneath eight months, in accordance with CNBC. Open weights let any firm with GPUs serve a near-frontier mannequin on {hardware} it controls.
- The largest patrons locked up the most recent provide. The 4 largest hyperscalers information $720–740 billion of 2026 capital spending, and the 4 largest disclosed offers by AI labs and Meta with neoclouds complete about $233 billion, three of them “as much as” quantities and far of it for capability not but delivered.
- The GPU-hour is changing into a monetary asset. CME, ICE and Nodal Alternate have introduced GPU futures, every selecting a special worth index, whereas lenders are already lending in opposition to GPUs: Lambda’s investment-grade-rated mortgage priced at 6.78% and Sharon AI’s, with no score disclosed, at 9.95%.
“Open-weight fashions are shut behind one of the best closed ones, and firms are switching to them. Any firm with GPUs can now serve a near-frontier mannequin on {hardware} it controls. The GPUs are the laborious half,” mentioned Evgeny Ponomarev, Co-founder of Fluence.
The report closes with 4 methods to pay much less for a similar compute: examine the identical chip throughout supplier sorts, commit when the workload is regular, match the GPU technology to the mannequin, and look past the hyperscalers, since a lot of the most recent neocloud capability is contracted to labs years forward.
The total report is obtainable at fluence.ai/compute-race-2026.
About Fluence
Fluence Cloud is a GPU cloud supplier with capability from unbiased information facilities (fluence.ai). Based in 2017 by Dmitry Kurinskiy, Evgeny Ponomarev and Tom Trowbridge, the corporate has raised $14 million from traders together with 1KX, Multicoin, Distributed International, Blockchange, Tiger International and Protocol Labs.
