Marketplace funnel to private cloud
Voltage Park
The key strategic point is that both companies use cheap, flexible marketplace supply to win customer logos, then convert the best workloads into reserved, dedicated GPU deals where the real economics live. In practice, that means a startup can get machines fast through a self serve path, while larger buyers can move into longer contracts for whole clusters, guaranteed capacity, and lower unit costs. Fluidstack has already made that shift at scale, and Voltage Park is building the same ladder from small jobs to committed infrastructure.
-
Fluidstack makes the split unusually clear. Its marketplace business sources underused GPUs from data centers and carries roughly 13 percent gross margins, while its Private Cloud uses dedicated colocated clusters on 2 to 3 year contracts with 25 to 50 percent paid up front and roughly 85 percent gross margins. That is the clearest template for why the dual model matters.
-
Voltage Park mirrors that structure with on demand rentals for developers and test workloads, then 6 to 12 month reserve contracts for enterprises that want guaranteed capacity and discounts. Its TensorDock acquisition adds a long tail marketplace style funnel, so a one GPU developer can start small and later move onto larger dedicated H100 clusters.
-
Customer behavior explains why this design keeps showing up. GPU cloud buyers often treat providers as interchangeable, switch in a day or two, and choose mostly on price, reliability, and whether the latest NVIDIA inventory is actually available. A marketplace gets them in the door quickly, but reserved capacity is what reduces churn and makes revenue more predictable.
This model is heading toward a barbell. The low end becomes an increasingly commoditized spot market for developers, and the high end consolidates around operators that can finance large GPU fleets and lock in multi year capacity deals. The winners will be the companies that turn fast provisioning into long duration infrastructure relationships before hyperscalers close the availability gap.