H100 rental price
median price per GPU-hour$2.55 / GPU-hr
Live chart · updates with every published snapshot
Most AI compute owned by one company
The largest reviewed corporate stockpile of AI accelerators in NVIDIA H100 equivalents, shown against every tracked owner with uncertainty bands, estimated unit counts, and accelerator power draw.
H100-equivalents owned5.45M H100eheld by Alphabet
Leader share of tracked total34.1%Alphabet
Lead over no. 21.5×Microsoft
Tracked total15.97MH100e across 6 owners
- Alphabet5.45M34.1%
- Microsoft3.57M22.4%
- Amazon2.51M15.7%
- Meta2.4M15.0%
- Oracle1.2M7.5%
- CoreWeave831.33K5.2%
Bars are median estimates in NVIDIA H100 equivalents. The thin band across each bar spans the p5 to p95 uncertainty interval, and shares are of the total across tracked owners.
| Owner | H100e, median | p5–p95 interval | Share | Est. accelerators | Accelerator power | Silicon | Coverage |
|---|---|---|---|---|---|---|---|
| Alphabet | 5.45M | 4.09M – 7M | 34.1% | 7.17M | 3.1 GW | Google · Nvidia | Partial |
| Microsoft | 3.57M | 2.69M – 4.77M | 22.4% | 2.35M | 2.1 GW | AMD · Nvidia | Partial |
| Amazon | 2.51M | 2M – 3.2M | 15.7% | 2.79M | 1.7 GW | Amazon · Nvidia | Partial |
| Meta | 2.4M | 1.79M – 3.23M | 15.0% | 1.54M | 1.4 GW | AMD · Nvidia | Partial |
| Oracle | 1.2M | 864.19K – 1.69M | 7.5% | 703.8K | 681 MW | AMD · Nvidia | Partial |
| CoreWeave | 831.33K | 756.76K – 912.71K | 5.2% | 511.02K | 484 MW | Nvidia | Complete |
5.45M H100e on every publication so far. A line appears here as soon as the number moves.
About this metric
AI compute ownership is the hardest form of power in the AI economy: models change monthly, but accelerator fleets take years and tens of billions of dollars to assemble. This chart ranks every reviewed corporate owner in H100 equivalents, a single scale that makes mixed fleets of different accelerator generations directly comparable, and shows how much of the tracked total the largest holder controls.
Figures are Epoch AI median estimates of owned compute, normalized by peak dense 8-bit operations per second against an NVIDIA H100 baseline, with uncertainty bands preserved on the underlying board. Ownership means the accelerators a company controls, not the capacity it rents out or the utilization it achieves.
Whoever owns the most compute sets the terms for everyone renting it. Concentrated ownership shapes model training roadmaps, cloud pricing, and even national industrial policy. Tracking the top holding over time shows whether the compute base is consolidating further or whether challengers are closing the gap.
Related charts