06Compute About 5 minutes
Who shows up in the compute data?
Nobody publishes a complete census of the world's AI computing power. Epoch AI assembles partial data from public sources. Make seven guesses about what that data shows, then see each answer with its source, its definition and its limits.
- Questions 1 to 3
- Questions 4 and 5
- Questions 6 and 7
- Your score
? Question 1 of 7
! The answer
Data: Epoch AI chip owners (updated May 12, 2026), GPU clusters (updated March 13, 2026, no longer maintained), AI data centers and AI models (both retrieved September 26, 2026). How these are measured
4 Your score
Who owns the means of producing intelligence, and what that ownership earns, is a central question in Intelligence Capitalism by Bharat Rao.
Sources and method
- Epoch AI, "Data on AI chip owners," updated May 12, 2026, epoch.ai/data/ai-chip-owners, file cumulative_by_designer.csv. CC BY 4.0.
- Epoch AI, "Data on GPU clusters," updated March 13, 2026, epoch.ai/data/gpu-clusters, file gpu_clusters.csv. CC BY 4.0. Epoch AI marks this dataset as deprecated and estimates that it covers 10 to 20 percent of global GPU cluster performance as of March 2025.
- Epoch AI, "Data on AI models," large-scale AI models file, epoch.ai/data/ai-models, file large_scale_ai_models.csv. CC BY 4.0.
- Epoch AI, "AI data centers," retrieved September 26, 2026, epoch.ai/data/ai-data-centers, files data_centers.csv and data_center_timelines.csv. CC BY 4.0. Field definitions: records documentation.
- Epoch AI, "Data on AI models," notable AI models file, retrieved September 26, 2026, epoch.ai/data/ai-models, file notable_ai_models.csv. CC BY 4.0. Field definitions: records documentation.
Method. Question 1 sums Epoch AI median estimates of AI chips acquired since the first quarter of 2022, in H100 equivalents, across the six chip designers Epoch AI covers (Nvidia, AMD, Google, Amazon, Huawei and Cambricon), for the latest quarter in which every owner has complete data (December 31, 2025). The denominator is that tracked total (about 20.3 million H100 equivalents), not all AI compute in the world: chips bought before 2022 and chips from other designers are outside it, and retirements are not subtracted. The speculative estimate of smuggled chips in China (about 0.66 million H100 equivalents) is excluded, matching the default view in Epoch AI's explorer. Owners that Epoch AI does not attribute individually are grouped as all others (about 16.6 percent). Epoch AI estimates ownership, not use: many AI developers rent the chips they train on, and rented capacity counts for the owner. An earlier draft of this page described the result as a share of the world's AI chip computing power; that wording overstated what the data covers and was corrected on September 27, 2026. Question 2 counts existing clusters, keeping only the latest phase of clusters built in phases so capacity is not double counted. Question 3 counts models trained with more than 10^23 operations by the country of the developing organization; models from teams in several countries are grouped as multinational. Questions 4 to 7 each reveal one value exactly as Epoch AI publishes it in a single row of its data, with no calculation by us beyond converting units (watts to megawatts, dollars to millions). Question 4 is the IT power (MW) of the data center with the highest value in Epoch AI's timelines file, using each site's latest record dated on or before September 26, 2026; this equals the Current power field in data_centers.csv, although Epoch AI's documentation describes that field as total facility power, so we use the clearly defined timeline field and also quote Epoch AI's facility power estimate for the same site. Question 5 is the median Number of Units in the row for Google TPUs through the fourth quarter of 2025, the same quarter as question 1, with Epoch AI's 5th and 95th percentile estimates. Questions 6 and 7 are the Training compute cost (2023 USD) of Llama 3.1-405B and the Training power draw of Grok 3 in the notable AI models file, with Epoch AI's confidence rating for each model. The comparison charts show other published rows of the same field, sorted by value. Epoch AI data are collected from public sources and are not exhaustive.