Amazon, Google, Meta, Microsoft, and Oracle collectively hold an estimated 71% of the world’s cumulative AI compute as of Q4 2025, measured in H100-equivalents of computing power. This is up from 63% in Q1 2024.
| Quarter | End date | Owner | H100-equivalents | Share of global AI compute |
|---|---|---|---|---|
| 2024 Q1 | 2024-03-31 | Top 5 hyperscalers combined | 1.5e+6 | 62.5 |
| 2024 Q1 | 2024-03-31 | 5.0e+5 | 20.8 | |
| 2024 Q1 | 2024-03-31 | Microsoft | 4.5e+5 | 18.9 |
| 2024 Q1 | 2024-03-31 | Meta | 2.5e+5 | 10.5 |
| 2024 Q1 | 2024-03-31 | Amazon | 2.0e+5 | 8.4 |
| 2024 Q1 | 2024-03-31 | Oracle | 9.3e+4 | 3.9 |
| 2024 Q2 | 2024-06-30 | Top 5 hyperscalers combined | 2.2e+6 | 63.7 |
| 2024 Q2 | 2024-06-30 | 7.1e+5 | 20.6 | |
| 2024 Q2 | 2024-06-30 | Microsoft | 6.4e+5 | 18.8 |
| 2024 Q2 | 2024-06-30 | Meta | 3.8e+5 | 11.1 |
| 2024 Q2 | 2024-06-30 | Amazon | 3.1e+5 | 9.0 |
| 2024 Q2 | 2024-06-30 | Oracle | 1.5e+5 | 4.2 |
| 2024 Q3 | 2024-09-30 | Top 5 hyperscalers combined | 3.1e+6 | 64.5 |
| 2024 Q3 | 2024-09-30 | 9.7e+5 | 20.5 | |
| 2024 Q3 | 2024-09-30 | Microsoft | 8.7e+5 | 18.5 |
| 2024 Q3 | 2024-09-30 | Meta | 5.3e+5 | 11.3 |
| 2024 Q3 | 2024-09-30 | Amazon | 4.6e+5 | 9.8 |
| 2024 Q3 | 2024-09-30 | Oracle | 2.1e+5 | 4.4 |
| 2024 Q4 | 2024-12-31 | Top 5 hyperscalers combined | 4.5e+6 | 67.4 |
| 2024 Q4 | 2024-12-31 | 1.4e+6 | 20.9 | |
| 2024 Q4 | 2024-12-31 | Microsoft | 1.2e+6 | 18.6 |
| 2024 Q4 | 2024-12-31 | Meta | 7.7e+5 | 11.7 |
| 2024 Q4 | 2024-12-31 | Amazon | 7.6e+5 | 11.5 |
| 2024 Q4 | 2024-12-31 | Oracle | 3.1e+5 | 4.7 |
| 2025 Q1 | 2025-03-31 | Top 5 hyperscalers combined | 6.4e+6 | 69.7 |
| 2025 Q1 | 2025-03-31 | 2.0e+6 | 21.9 | |
| 2025 Q1 | 2025-03-31 | Microsoft | 1.7e+6 | 18.5 |
| 2025 Q1 | 2025-03-31 | Amazon | 1.1e+6 | 12.2 |
| 2025 Q1 | 2025-03-31 | Meta | 1.1e+6 | 11.9 |
| 2025 Q1 | 2025-03-31 | Oracle | 4.8e+5 | 5.2 |
| 2025 Q2 | 2025-06-30 | Top 5 hyperscalers combined | 8.6e+6 | 71.1 |
| 2025 Q2 | 2025-06-30 | 2.8e+6 | 23.4 | |
| 2025 Q2 | 2025-06-30 | Microsoft | 2.2e+6 | 18.0 |
| 2025 Q2 | 2025-06-30 | Amazon | 1.5e+6 | 12.5 |
| 2025 Q2 | 2025-06-30 | Meta | 1.4e+6 | 11.7 |
| 2025 Q2 | 2025-06-30 | Oracle | 6.7e+5 | 5.5 |
| 2025 Q3 | 2025-09-30 | Top 5 hyperscalers combined | 1.1e+7 | 71.3 |
| 2025 Q3 | 2025-09-30 | 3.8e+6 | 24.3 | |
| 2025 Q3 | 2025-09-30 | Microsoft | 2.8e+6 | 17.4 |
| 2025 Q3 | 2025-09-30 | Amazon | 2.0e+6 | 12.4 |
| 2025 Q3 | 2025-09-30 | Meta | 1.8e+6 | 11.6 |
| 2025 Q3 | 2025-09-30 | Oracle | 8.9e+5 | 5.7 |
| 2025 Q4 | 2025-12-31 | Top 5 hyperscalers combined | 1.4e+7 | 70.7 |
| 2025 Q4 | 2025-12-31 | 5.0e+6 | 24.8 | |
| 2025 Q4 | 2025-12-31 | Microsoft | 3.4e+6 | 16.8 |
| 2025 Q4 | 2025-12-31 | Amazon | 2.5e+6 | 12.1 |
| 2025 Q4 | 2025-12-31 | Meta | 2.3e+6 | 11.3 |
| 2025 Q4 | 2025-12-31 | Oracle | 1.1e+6 | 5.6 |
| 2026 Q1 | 2026-03-31 | Top 5 hyperscalers combined | 1.3e+7 | 75.8 |
| 2026 Q1 | 2026-03-31 | 5.5e+6 | 31.0 | |
| 2026 Q1 | 2026-03-31 | Microsoft | 3.3e+6 | 18.9 |
| 2026 Q1 | 2026-03-31 | Meta | 2.0e+6 | 11.1 |
| 2026 Q1 | 2026-03-31 | Amazon | 1.5e+6 | 8.6 |
| 2026 Q1 | 2026-03-31 | Oracle | 1.1e+6 | 6.2 |
Many AI labs (including OpenAI and Anthropic) largely depend on these hyperscalers for access to R&D and inference compute.
Epoch's work is free to use, distribute, and reproduce provided the source and authors are credited under the Creative Commons BY license.
Learn more about this graph
We calculate the aggregate amount of compute (in H100-equivalents) held by Amazon, Google, Meta, Microsoft, and Oracle, as a share of the global total each quarter. These five largest hyperscalers own a combined 71% of compute as of Q4 2025, rising from 63% in Q1 2024.
Data
Chip ownership data comes from Epoch AI’s AI Chip Owners dataset, which tracks chip acquisitions by major entities from Q1 2022 through Q4 2025. The dataset covers chips manufactured by Nvidia, Google, AMD, Huawei, and Amazon. We focus on the period 2024 through 2025 because Google’s TPUs are only tracked from 2023 onward and using less than one year of past data results in unstable aggregate figures.
Assumptions and limitations
Our Chip Ownership data does not capture all global chip ownership, and has weaker coverage prior to 2023. Q1 2026 data is excluded as it was not yet finalized at time of analysis.
The H100-equivalent unit uses a chip’s highest 8-bit operation/second specifications to convert between chips. The actual utility of a particular chip depend on workload assumptions, so H100e does not perfectly reflect real-world performance differences across chip types.


