Hyperscalers list GPUs at nearly 2× the GPU-cloud price, Fluence research finds - Crypto Briefing
Fluence's AI Compute Race report reveals hyperscalers list GPUs at 1.98 times the GPU-cloud median price, with commitment lengths significantly affecting costs.
Fluence has published The AI Compute Race report, which prices GPU cloud compute for companies that aren't frontier labs or hyperscalers. The central finding reveals that where a company buys a GPU-hour significantly impacts the price. Across H100, H200 and B200 models, hyperscalers list GPUs at 1.98 times the GPU-cloud median price.
Price Variations
The price differences vary by model, ranging from 1.4 times on the B200 to 2.8 times on the H100. For example, the same H200 GPU lists from $4.08 to $10.60 per hour depending on the provider. Commitment length also affects pricing significantly.
A one-year reservation cuts the median price by 21% compared to on-demand rates, while a three-year term reduces it by 48%. Spot capacity lists at just 53% of on-demand prices, though for workloads that can be interrupted.
Rising Costs
Seventeen of the 22 GPU models tracked now cost more than a year earlier, with the median H100 renting for 13% to 30% more depending on the measurement window. The report attributes rising prices to several factors.
Demand has shifted toward open models, with Chinese models' share of OpenRouter tokens among US, European and Global South companies rising from 6-13% to 57-67% in under eight months. The four largest hyperscalers are guiding $720-740 billion in 2026 capital spending, while AI labs and Meta have disclosed deals totaling about $233 billion with neocloud providers.
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.
Financialization
The GPU-hour is increasingly becoming a financial asset, with CME, ICE and Nodal Exchange announcing GPU futures that settle on different price indices. Lenders are already providing financing against GPUs, with Lambda's investment-grade-rated loan at 6.78% and Sharon AI's at 9.95%.
The report suggests four ways to reduce compute costs: compare the same chip across provider types, commit when workloads are steady, match GPU generation to model requirements, and look beyond hyperscalers since much of the newest neocloud capacity is contracted to labs years ahead.
Related
