Operating AI data centers
Data centers serving AI workloads today. A further 81 operating sites are under review and not shown.
Status: preliminary and in progress. Figures will change as they are verified.
Exhibit 1
Ten largest operating AI data centers by IT load, MW. Select a data center to open its profile.
Exhibit 2
Median months from ground-break to first power, by company, live AI campuses; companies with two or more campuses. Select a company to see the campuses behind its median.
Exhibit 3
Months from ground-break to first power, by IT load of the first phase brought online, live AI campuses. Hover over a point for the campus; select it to open the profile.
Exhibit 4
Live AI IT load by company, MW, with cumulative share of the total
Exhibit 5
Takeaways so far
All operating AI data centers
Search or sort the full list. Select a row to open its profile.
Upcoming AI data centers
Data centers that are rumoured, announced, permitted, in early works or under construction, and are AI or likely AI.
Status: preliminary and in progress. Power is not shown for upcoming sites.
Exhibit 6
Ten largest AI data centers under construction, ranked by reported campus capacity. Select a data center to open its profile.
Exhibit 7
Upcoming AI data centers by country and project status, number of sites, top 12 countries
Exhibit 8
Upcoming AI data centers by US state, number of sites. Select a state to list its sites.
Exhibit 9
Upcoming AI data centers by type of builder, number of sites, split by AI classification
Exhibit 10
Takeaways so far
Data risks
What our review of the data and our visits to two sites under construction say about how far to trust this list.
All upcoming AI data centers
Search or filter the full list. Select a row to open its profile.
Unit economics of a token
What one gigawatt of AI compute costs to build and run, how many tokens it produces, and how the money splits between the developer who owns the building, the neocloud that rents out the chips, and the model provider that sells the tokens.
Status: illustrative model. Every number below recalculates when you change a filter. Rental rates, API prices, power prices and land comparables were looked up once in October 2026; they are not live.
Rest is training
Rest is free tiers and internal use
Depreciation period for servers
Neocloud billed utilization
All assumptions
Exhibit 11
From a gigawatt to a price per token, one gigawatt of grid power in the selected county, filled with the selected chip and serving the selected model. Annual figures.
Exhibit 12
Where the price of a million tokens goes, $ per 1M tokens, blended input and output at the selected ratio
Exhibit 13
Revenue and operating profit per gigawatt, by layer, $ billions per year
Model provider profit is token revenue less compute cost (rent, or the full cost of the gigawatt when integrated). It excludes research staff, sales and other operating costs.
Exhibit 14
List price and our cost to serve, by model, $ per 1M tokens, blended, log scale. The gap is the price and cost delta that the provider keeps.
Show as table
Exhibit 15
Serving cost per 1M tokens, by chip, selected model, $ per 1M blended tokens
Exhibit 16
Neocloud return on invested capital, by chip, after tax, at the selected rental rate; line marks the cost of capital
Exhibit 17
Cost of a GPU-hour against what it rents for, by chip, $ per GPU-hour sold
Exhibit 18
What the depreciation period does to the neocloud, selected chip, chip useful life from 3 to 7 years
Return on invested capital
Rent needed to earn the cost of capital, $ per GPU-hour
Exhibit 19
Output tokens per second per chip against speed per user, selected model and precision, log scale. Faster answers mean smaller batches and fewer tokens per chip.
Modelled: decode time per step is the larger of memory time (weights touched plus each user's KV cache, at 70% of rated bandwidth) and compute time (at 40% of rated FLOPs). Replica size is chosen to maximise output per chip. Real serving stacks add speculative decoding and prefill/decode splits that this simple model leaves out.
Exhibit 20
Model provider gross margin after training and free usage, at the selected model's price; outlined cell is the current setting
Monetized share of inference
Exhibit 21
The same gigawatt in each county, selected chip and model
Show county inputs
Exhibit 22
Assumptions and sources
Inputs marked est. are our estimates where no public figure exists. Model sizes for closed models are not disclosed; the defaults are assumptions you can change under All assumptions.
What's next
Guidance needed.
- Visit or speak with a few sites, utilities and county offices to verify what the data show.
- Compare announced online dates with the dates sites actually came online.
- Expand the list of data centers, starting with the 81 operating sites held for review.
- Estimate power ourselves for a sample of sites.
- Test published third-party estimates against filings and against each other.
- Project AI data center supply for the next six quarters.
Appendix
Data sources
How the site's data were built
Data Collection: Parameters and Their Fill Rates
Each data center is described by parameters of three kinds: those retrieved from public announcements and filings, those observed or deduced from satellite imagery, and those estimated with a model. This section shows how complete each parameter is today.
Status: preliminary and in progress.
Exhibit 23