OWNER GROUP / COMPUTE
US-based cloud groups
Estimated cumulative AI accelerator capacity owned by seven named US-based cloud and model groups. Ownership is not the same as physical location, usage or access.
OWNER GROUP / TIME SERIES / LOGUS-based cloud groups
Observed estimate · Q4 202531.2 ZOPSUS-based cloud groups
5th–95th percentile25–38.7 ZOPSView evidence
Measurement 5th–95th percentile
The sum of the owner group's members as recorded for each time period.
MEMBERS / Q4 2025Select an owner for the next level
CHIP TYPE / LATEST QUARTERSelect an owner
DRILLDOWN READY
Select Amazon, Microsoft or another owner to see its estimated chip mix.
METHOD LIMIT
The group describes ownership by company identity, not physical location or state control.
The chip-type view comes from Epoch AI's separate chip-type table and may differ slightly from the ownership total due to source aggregation.
PRIMARY SOURCEData on AI Chip OwnersDocumentation ↗
VERIFIABLE EVIDENCE / 5 CLAIMSOpen evidence +
Claims, calculations, sources and known limitations for this view. Source data and dashboard interpretation are kept separate.
The curve estimates cumulative AI accelerator ownership in the Epoch AI dataset; it does not measure actual use, access, physical location or effective model capacity.
- Coverage follows Epoch AI's ownership model and is not a complete inventory of all computing worldwide.
compute/claim/measurement-scopeH100e is a comparison model that normalises different AI chips against the Nvidia H100's theoretical maximum dense 8-bit performance; it is not a count of H100 cards.
CALCULATION
H100e = peak dense 8-bit OPS/s ÷ 1.979e15; ZOPS = H100e × 1.979e15 ÷ 1e21- Memory, networking, software, precision, utilisation and workload can make actual performance higher or lower than this comparison.
compute/claim/h100e-modelThe series' 5th and 95th percentiles sum the source groups' respective bounds as a transparent display approximation.
CALCULATION
series_low = Σ owner_low; series_high = Σ owner_highEpoch AI: Data on AI Chip Owners →1 measurement
- The summed bounds are not a joint probabilistic confidence model.
compute/claim/uncertainty-modelThe US-cloud drilldown exposes every group-total and ownership point with a stable measurement ID in the form compute/us-cloud/total/<dato> or compute/us-cloud/<eier>/<dato>; each point retains lineage to the same time-versioned source observations summed by the chart.
Epoch AI: Data on AI Chip Owners →114 measurements
- Drilldown IDs identify derived display points; they do not replace atomic source facts.
- Owner identity is not the same as physical location or actual access.
compute/claim/us-cloud-drilldown-method15,746,640 H100e is the aggregate median for seven time-versioned US-based owners at 2025-12-31; their share of the world series is 75%.
CALCULATION
group_total = sum(member_owner_medians); share_pct = group_total ÷ world_median × 100Epoch AI: Data on AI Chip Owners →2 measurements
- The group describes ownership by company identity, not physical location or state control.
- The chip-type view comes from Epoch AI's separate chip-type table and may differ slightly from the ownership total due to source aggregation.
compute/claim/us-cloud-drilldown-latest