How should rented GPUs be counted without double counting?
By Nathan Benaich (Air Street Capital)
Inventory and compute-access methodology
Count the physical hardware once in its owner or operator inventory, and describe a customer’s access separately. If a cloud provider operates 10,000 GPUs and rents that same capacity to a lab, adding both figures does not create 20,000 unique GPUs. This is an illustration of overlapping access, not a claim about a particular company. A rental agreement describes a right to use capacity rather than an additional hardware installation.
How to read this finding
The index separates selected owner and operator GPU inventories from the frontier-lab compute-access chart. The latter includes commitments and power-capacity figures that cannot simply be added to GPU counts. Individual rows can also describe clusters or broader fleets at different stages of deployment. Do not assume every disclosed figure is mutually exclusive or sum the charts into an audited global total. Check the source, physical site or fleet scope, reporting date, and deployment status before aggregating entries.
Citation estimates use inputs through September 1, 2026; fleet counts have an October 1, 2026 cutoff. The research-topic sample covers January 1 to June 1, 2025. See each answer for its period and limitations.
Benaich, Nathan. “How should rented GPUs be counted without double counting?” State of AI Report Compute Index. Web page updated 2026-10-11; data periods as specified above.
