Act 1 of 6
Hover or tap a country for its figures
Tools · buildings · money · water · heat
Every country in the Data Center Map directory, set against who owns the AI tools people actually use, the size of its economy, its freshwater stress, and how much it has warmed. The buildings sit in 185 countries. The tools are owned from 12.
Act 1 of 6
How to use this map. It plays on its own, holding each act for seven seconds. Press Pause to stop it, click any of the six acts to jump straight there, or use the left and right arrow keys to step. Hover, tap, or tab to a country for its full figures, including its own warming stripes. The key under the map changes with each act.
On a phone. Press + on the map to zoom in, or pinch it. The map grows taller as it zooms, so north and south stay in view while you drag east and west. Fit puts the whole world back.
The four tiers
The tier is one ratio: a country's share of the world's top AI tools, divided by its share of the world's data centres. At 1 it owns tools in proportion to the buildings it hosts. Above 1 it owns more than it hosts. Below 1 it hosts more than it owns.
Tier 1 is 1 or above, 9 countries. Tier 2 is below 1, 3 countries. Tier 3 owns no tool and hosts ten or more data centres, 72 countries. Tier 4 owns no tool and hosts fewer than ten, 101 countries.
The tools are the 81 distinct products in a16z's Top 100. 67 of them are placed in 12 countries and 14 I could not place, so the shares are of the 67.
The denominator is small, so the tier boundary is thin. Seven of the twelve countries hold exactly one product, and ten hold two or fewer. Only the United States at 36 and China at 18 survive a single reclassification. Take away Sweden's one product and it falls from tier 1 to tier 3. Spain, Italy, and France are one product away from the same drop. Read tiers 1 and 2 as a small group rather than a ranking.
How to read these bars. Each bar splits the world total across the four tiers, darkest for tier 1. Hover or tap a segment for the exact figure behind the percentage. Compare the facilities bar against the tools bar: that gap is what the tiers measure. The four tiers are spelled out underneath.
Buildings and tools
Facility counts and tool ownership, for the same countries.
| Country or region | Facilities | Share of facilities | Top-100 tools | Share of tools |
|---|
How to read this table. Compare the two share columns across a single row. Facility share is of all 13,207 listings, and tool share is of the 67 products whose owner I could place. On a phone each row becomes a card.
Sources: Data Center Map country directory; a16z Top 100 Gen AI Consumer Apps, 7th edition. Europe (29) is the United Kingdom, Germany, France, Italy, Spain, the Netherlands, Sweden, Poland, Ireland, Finland, Denmark, Norway, Switzerland, Austria, Belgium, Portugal, Czech Republic, Romania, Greece, Hungary, Iceland, Luxembourg, Slovakia, Slovenia, Croatia, Bulgaria, Estonia, Latvia, and Lithuania. It excludes Russia, Turkey, Ukraine, and the Balkan states outside that list.
602 of them, behind only the United States, and it owns none of the top 100 AI tools. Germany has 585 facilities and owns none. Japan has 269 and owns none. India has 348 and owns none.
Who owns the tools
a16z ranks the top 50 AI products on the web by monthly visits and the top 50 mobile apps by monthly actives, both for August 2026. Stripping the overlap leaves 81 distinct products.
These are consumer products. The ranking holds no enterprise AI, no defence, no scientific computing, and no chip design, and it is a single month's traffic. A country can run a serious national AI programme and own nothing on this list. The country is the headquarters of the company that owns the product, which is a proxy for control rather than a measure of it: where a model runs, whose law reaches it, and who can switch it off are separate questions.
How to read these cards. The large figure is how many of the 81 products that country owns, and the small type names them, up to four before it counts the rest. The last line says how many of that country's attributions rest on a single source. The dashed card holds the products whose headquarters I could not confirm at all.
SeaArt has a Chinese founding team, a Hong Kong or Singapore entity, and a Japanese arm. Manus runs from Singapore under a Cayman parent, with the original Beijing company still registered. For a lot of AI tools the corporate address is a choice rather than a fact, and the unverified band is the honest version of that.
They are not missing at random. Two have confirmed Chinese founding teams, and nine of the fourteen come from the mobile ranking, where ownership is hardest to trace. Leaving them out of the shares most likely understates China and flatters everyone else.
Read each line as the country the product was built in, then the country of the company that owns it now.
Wealth against capacity
Income per head along the bottom, data centres per million people up the side, both on log scales. Circle size is population, colour is the tier.
How to read this chart. One dot per country. Hover or tap a dot for that country's full figures. Circle size is population and colour is the tier. Both axes are logarithmic, so each labelled gridline is ten times the one before it.
Sources: Data Center Map country directory, read 12 November 2026. World Bank: GDP (current US$) and total population, 2024.
The United Kingdom and Germany sit high on the right and own none of the top 100 tools. Ireland, Iceland, and Finland sit higher still and own none. China sits low and owns 18 of them. India has 348 facilities and owns none.
Who pays the water cost
Water stress is SDG 6.4.2: freshwater withdrawn each year as a percentage of what renews. Above 100% a country is draining groundwater or desalinating. These eighteen hold ten or more facilities, sit at 40% or higher, and own none of the top 100 AI tools.
| Country | Facilities | Water stress | Warming vs 1961 to 1990 | Population | Tier |
|---|
How to read this table. Sorted by water stress, highest first. Every country here has ten or more data centres, owns none of the top 100 AI tools, and withdraws at least 40% of its renewable freshwater each year.
Sources: FAO AQUASTAT SDG 6.4.2, served as World
Bank indicator ER.H2O.FWST.ZS, latest year per
country (2022 for almost all). Copernicus ERA5 via Our World in Data, rebased to 1961 to 1990.
A facility's water use depends on its cooling, size, and load. The directory records none of that. A country figure tells you the water was already contested before the servers arrived, and little else. Most operators never publish their draw.
Who pays the heat cost
One stripe per year, 1940 to 2025, against each country's own 1961 to 1990 average. The form is Ed Hawkins's, drawn in the Slow AI palette. The data is Copernicus ERA5.
How to read these stripes. One stripe per year, 1940 on the left and 2025 on the right. Blue is cooler than that country's 1961 to 1990 average and orange is warmer. The figure on the right is the mean of the last ten years. Hover any country on the map above to see its own stripes.
An anomaly measures how much a place has changed against its own past. It tracks latitude, so high-latitude countries sit at the top and tropical islands at the bottom, and it says nothing about who is harmed. It also carries no attribution: nothing here links a country's warming to the data centres inside it.
Sources: Copernicus Climate Change Service, ERA5 monthly averaged data on single levels, 1940 to present, via Our World in Data; rebased by me from the published 1991 to 2020 baseline to the 1961 to 1990 WMO reference period. Stripe design: Ed Hawkins, University of Reading, 2018.
Every country in the directory
Click a heading to sort. Facilities per trillion is the count over nominal GDP: how built out a country is for the size of its economy.
| Country | Facilities | Per million | Per $trn GDP | GDP $bn | Tools | Build : host | Water stress | Warming | Tier |
|---|
How to use this table. Click or press Enter on any column heading to sort by it, and again to reverse the order. Type in the box above to find a country, and scroll inside the table to reach all 185 of them. On a phone it carries the name, the facilities, and whatever you have sorted by, and a tap on any row gives you that country's full figures. Tools is how many of the 67 placed top-100 AI products that country owns, and Build : host is its share of those divided by its share of the world's data centres. Build : host reads n/a for any country owning no tools, because the ratio is zero there whether the country hosts one data centre or six hundred, so it carries nothing the Tools column has not already said. Sorting still ranks those countries last. Elsewhere n/a means the figure does not exist, and 0 means it is genuinely zero.
Provenance
Each field, its source, the exact dataset, and the date I pulled it.
NY.GDP.MKTP.CD) and population (SP.POP.TOTL), pulled 12 November
2026. Taiwan is not a World Bank member, so its figures come from UN national accounts.ER.H2O.FWST.ZS, pulled
12 November 2026. Latest year per country, 2022 for 176 of the 178 with a value.Where this came from
A newsletter about knowing when to use AI and when to leave it alone, read in 179 countries. Free on Fridays, paid on Wednesdays. This atlas was built for it, and the arguing happens over there.
£100 for the year, or £10 a month, and you get every Wednesday piece in full, plus the whole Slow AI archive. Paid readers can email me and get an answer.