• Hyperscalers list H100, H200 and B200 GPUs at 1.98 times the GPU-cloud median price.

• The H100 shows the widest gap at 2.8x; the B200 gap is 1.4x.

• A three-year GPU commitment cuts the median price 48% against the same provider's on-demand rate.

Hyperscalers List GPUs at 1.98x

Hyperscalers listed GPU compute at 1.98 times the specialized GPU-cloud median in a study Fluence released on Monday, spanning the H100, H200 and B200 across published hourly rates. The report, The AI Compute Race, was produced with research partners NEAR AI, io.net, Boundless, Phala, Gensyn, Fortytwo, Theta, Compute Labs and GAIB. Its central argument is simple: where a company buys a GPU-hour sets what it pays. The spread runs from 1.4 times on the B200 to 2.8 times on the H100, and the same H200 lists from $4.08 to $10.60 an hour depending on the provider. Commitment length matters as much as vendor choice: a one-year reservation cuts the median price 21% against the same provider's on-demand rate, a three-year term 48%, and interruptible spot capacity lists at 53% of on-demand. Buyers hoping for relief will not find it in the trend line. Seventeen of the 22 GPU models tracked by price aggregator GetDeploying cost more than a year earlier, and median H100 rents sit 13% to 30% higher depending on the window measured. The report attributes the drift to three forces. Open-weight models now carry more usage: Chinese models' share of OpenRouter tokens among companies in the US, Europe and the Global South rose from 6-13% to 57-67% in under eight months, so any firm holding GPUs can serve a near-frontier model on hardware it controls. The biggest buyers locked up the newest supply: the four largest hyperscalers guide $720 billion to $740 billion of 2026 capital spending, and the four largest disclosed deals between AI labs, Meta and neoclouds total about $233 billion, much of it for capacity not yet delivered. The GPU-hour is also becoming a financial asset: CME, ICE and Nodal Exchange have announced GPU futures on separate price indexes, and lenders are advancing credit against hardware, with Lambda's investment-grade loan priced at 6.78% and Sharon AI's at 9.95%. “The GPUs are the hard part,” Fluence co-founder Evgeny Ponomarev said of the open-weight shift. The report closes with four cost levers: compare the identical chip across provider types, commit when the workload is steady, match the GPU generation to the model, and look beyond hyperscalers.

How the 1.98x Gap Was Measured

The 1.98x figure is a median of published hourly rates, not a claim that every contract costs twice as much at a hyperscaler. The sample behind it varied by model: the H100 comparison covered three hyperscalers against seven GPU clouds, the H200 three against six, and the B200 two against six, because Azure does not publish a rate for its eight-GPU B200 product. The report converts the full price of such bundles into a per-GPU-hour figure. List prices also travel with conditions. Google Cloud's pricing page did not display region labels, and Azure's H200 rate is quoted for the US East 2 region, while hyperscaler instance rates may bundle CPU, memory and network where some specialists price the GPU alone. The report's sensitivity test found that excluding a single hyperscaler moved the measured H100 gap by 15 to 52 percentage points, so the headline multiple depends on the sample as much as the market. The H200 range of $4.08 to $10.60 an hour shows the dispersion: identical silicon listed more than 2.5 times apart. Independent tracking points in the same direction with different magnitudes. GetDeploying's November 9 weekly data counted 17 of 22 comparable GPU models higher and five lower over the past year, with a median change of 15%; the report's own reading of the same set puts the median increase at 19%. Both set the three-year discount at 48%, but the one-year discount reads 21% against 25%, and spot capacity at 53% against 56% of on-demand. One caveat sits with the publisher: Fluence operates a GPU cloud that aggregates capacity from independent data centers, so it sells compute in the market it is pricing. The report covers rental rates only and presents no figures linking the price gaps to FLT, the token on which the Fluence DAO runs on Ethereum.

Pre-Contracted Capacity and 2026 Spending

Read together, the findings and the caveats leave buyers a narrower but clearer playbook. The 1.98x median is directional rather than exact, yet the direction is unambiguous, and the 48% three-year discount is the single largest cost lever for steady workloads. With 17 of 22 models setting higher prices over the past year, the rental market has behaved like a support and resistance map in slow motion, building floors instead of breaking them. The $720 billion to $740 billion of hyperscaler capital spending and roughly $233 billion of disclosed neocloud deals point to continued scarcity of the newest GPUs, which makes provider comparison and early commitment the two moves the data supports.