Epoch AI 估算最大 AI 数据中心算力每 7 个月翻一番

Learn more about this graph

Data comes from our Frontier Data Centers dataset, which includes computing capacity estimates for each data center over time. Computing capacity is the total peak floating point operations per second (FLOP/s) of AI chips in the data center. We express this as the equivalent number of NVIDIA H100 chips required.

We focus on computing capacity rather than power capacity because it is more relevant to AI capabilities. The frontier trend in computing capacity indicates the limits of what the AI industry is willing and able to build in a single facility. The data shows that this frontier is dominated by hyperscalers (Google, Microsoft, Amazon, Meta) along with SpaceXAI.

Code for the analysis is available here.

Data

The full dataset extends from 2019 to 2030, but it’s only likely to cover the true frontier between 2024 and 2028. We start at the launch of Colossus 1 in August 2024 — a milestone almost two years after ChatGPT, when the AI investment boom began to deliver completed data centers. We end in April 2028, giving the future extrapolation an equal timespan to the historical trend.

We estimate that the dataset only covers 26% of global AI computing capacity as of March 31st, 2026 (5.2 million operational H100-equivalents against 20.2 million sold through 2025, assuming a 3-month deployment lag). However, we’ve prioritized finding the largest AI data centers, so there’s unlikely to be many that are larger than the ones we’ve identified.

Computing capacity is measured using the maximum FLOP/s specification available for each chip at 8-bit numerical precision, then converted to H100-equivalents using the FP8 Tensor Core specification for the NVIDIA H100 (dividing the specification by 2 to account for sparsity, i.e., 1,979 teraFLOPS).

Most estimates are derived from each data center’s peak power capacity and the energy efficiency of the most-sold chips (which varies by company, e.g., Google mostly runs TPUs). Power capacity itself is estimated from satellite imagery, public disclosures, and permitting documents. A few sites, like Colossus 1 and 2, disclosed chip types and counts directly, giving more confident estimates. Full methods are in the dataset documentation and each data center’s “Calculations sheet” field.

Analysis

To analyze the frontier trend, we filtered to the single largest data center by computing capacity over time. Note that the largest data center by compute capacity is not necessarily the largest by power capacity, as some data centers have more energy-efficient chips than others. The filtering resulted in 10 past observations and 5 future projections at the frontier.

Since the frontier trend appeared to grow exponentially, we did a simple log-linear regression to the past observations, resulting in the 3.3x per year growth trend. This is equivalent to a doubling time of 7 months. The chart shows the 90% prediction interval from this regression as a shaded area.

The growth rate is uncertain both due to the small sample size and the noise in each data point. For the frontier points, we expect that 80% of the time, the true data point is within a factor of 1.3x of our estimated compute capacity, and within ±3 months of our estimated operational date. For other points we use a factor of 1.5x for compute capacity and ±6 months for the operational date, because there is often less public information and we have done less research on them.

To measure the impact of this noise, we ran 1,000 simulations that randomly perturb the data before filtering to the frontier and fitting the trendline. The resulting fits ranged from 2.3x/year at the 5th percentile to 5.1x/year at the 95th percentile, compared to our point estimate of 3.3x/year. This is an illustrative range rather than a true confidence interval.

Assumptions and limitations

We assume the data covers the frontier from August 2024 to the present, but there is likely to be at least one point missing. The dataset prioritizes the United States, where most data center capacity sits, so we may have missed a large facility internationally.

Although we don’t fit the trend to future data points, it should be noted that these are more uncertain than past data points. Existing data centers may expand, and new ones may be accelerated or (more likely) delayed. However, we reviewed all known data centers that could plausibly reach the frontier in the future—via expansion or a new build—and concluded that the frontier is unlikely to grow any faster than the current trend through 2028.

Note that the largest data center is not necessarily the largest AI training cluster. A single data center might not be used as one training cluster, even if all its servers are networked together. A cluster could also consist of several data center campuses networked together—this is uncommon today, but has been demonstrated at a sub-frontier scale.

Download this data

Frontier Data Center Computing Capacity

CSV, Updated Jun. 11, 2026

Explore this data

AI Data Center Map