Zettabyte
AI Economics

How a GPU Marketplace Changes Procurement and Operating Responsibility

A GPU capacity marketplace represents a new business model for renting access to GPU clusters. At a high level, they aggregate multiple GPU providers so that a user can buy or compare their capacity from a single vendor. This still represents a purchase of rented GPU capacity, so any analysis of colocation vs cloud for AI workloads is still applicable upstream of this choice.

What Is a GPU Capacity Marketplace?

There are currently multiple approaches to providing this type of aggregation, and a few companies have published their approach to a GPU marketplace:

  • Fluence offers an auction-style marketplace, where a user can view a list of named clusters with an ask price based on GPU type and geographic region. A screenshot from October 15, 2026, of the Fluence auction shows H100 x176 clusters in South Korea for $2.30/GPU-hour; H200 x1024 clusters in US East for $2.76/GPU-hour; and B300 x1024 clusters in US Central for $3.92/GPU-hour. These prices are subject to change due to the auction format but can be used as a rough guide for recent GPU cluster rental costs.
  • Lightning AI is a development platform that offers a marketplace to rent GPUs, making it easier for companies to compare processors from hyperscaler and neocloud providers.

Some of the differences with a GPU capacity marketplace include:

  • Provider comparison: The ability to compare different providers, avoiding the need to find and compare options from multiple vendors.
  • Dynamic pricing: The price is always a snapshot, as with any rental. To get a sense of the dynamic price of GPUs, we recommend reading Ornn's H100 Pricing Index, which shows that H100 prices have been as high as $3.20/GPU-hour since May 2026, and were $2.67/GPU-hour as of August 11, 2026.
  • Cost versus tooling: Some marketplaces may allow the buyer to optimize the trade-off between cost and tooling, as neoclouds typically offer lower costs than hyperscalers but require more machine learning (ML) tooling to be built in-house rather than supplied by the cloud provider, as Semafor reports.

How Available Are the Clusters Listed on the Marketplace?

For rack-scale Blackwell clusters, a good indicator of how a marketplace sources capacity is how it explains cluster availability. GMI Cloud describes capacity that can be allocated on short notice or committed capacity with a provider.

This may differ from other listings on the same marketplace, so it is important to understand what "available now" means for each listing.

Does the Marketplace Take Responsibility for Operating the Cluster?

Depending on the structure of the contract, the GPU capacity marketplace may take responsibility for the operations of the cluster on behalf of the buyer.

In this case, SLA obligations (e.g., credits for downtime, escalation procedures and communications) may be handled by the marketplace, depending on the contractual terms. The underlying infrastructure may be owned and operated by a different company.

We have not seen an independent, audited comparison of how SLAs are passed through for different GPU capacity marketplace companies.

If going through a marketplace, a buyer should confirm:

  • Cluster ownership: The identity of the GPU cluster owner/operator and the region, as described in What Is a GPU Cluster?.
  • On-demand availability: Whether the marketplace uses a different model for "on demand" than the reserved and spot pricing models used by cloud providers.
  • SLA and support: How the SLA is structured, whether it is with the marketplace or the underlying provider, and how support tickets will be routed.
  • Pricing: Whether the listed price is a live price, how long that price will be held, and whether the final price may change based on availability, region or other allocation terms.
  • Compliance: How audit trails are passed through for compliance and who is responsible for signing them.