Data centers & the world, in charts
The companion to the interactive map: the same dataset, turned into charts and graphs. Where data centers are, who owns them, what they draw from the grid, what that costs in carbon, water and money, what's still being built, what got cancelled — and how it all lands on the communities next door. Drill into a place on the map or read the full community-impact report.
The at-a-glance answer: who builds data centers, where, what they draw, and what it costs the world around them. KPI tiles up top, then the highest-signal charts before you drill into energy, water, communities, build-out and global reach. Every count is a row count of cited records; every energy/water/cost/carbon figure is an estimate from the shared model.
The boom is, at heart, an electricity story. Summing the shared model (Hyperscale 40 MW @ 0.85 load factor, Colocation 5 MW @ 0.55, Edge PoP 0.3 MW @ 0.6) over the sample yields an estimated annual draw; pairing each facility's country with its grid carbon intensity turns kilowatt-hours into tonnes of CO₂. These are order-of-magnitude estimates, not meter readings — this 250-site sample models to ~10% of the IEA 2024 global anchor, so no chart here is a world total.
The physical footprint beyond the meter: cooling water and farmland. Water is estimated from the model's annual kWh × LBNL water-intensity factors (1.8 L/kWh direct, 9.4 L/kWh including upstream power generation). Farmland is computed by testing each facility's coordinates against major breadbasket regions and applying the per-type acreage and food-foregone model. Cited national water and land statistics anchor the estimates.
When concentrated demand outpaces supply, the costs land on ordinary people: electricity bills, diesel-generator exhaust, round-the-clock noise, and stressed water and power systems. This section pairs cited national statistics with 22 place-specific cases communities have documented near data-center clusters. National stats are shown as sourced tiles/tables because most are string ranges that can't honestly be plotted as numeric series.
Who owns the infrastructure and where it sits. A handful of operators (Equinix, Meta, Google, AWS, Microsoft) hold most of the fleet, and the overwhelming majority of mapped facilities are in the USA. Counts use only directly recorded fields; power-draw and build-cost rankings derive from the shared model. Rankings reflect this sample's coverage, not global market share.
What's coming, what stopped, and how markets reacted. The 14 mega-projects under construction carry reported, cited MW and cost (~19 GW, ~$159B) so they can be summed honestly. The stalled projects mostly lack MW/cost, so we deliberately don't total their gigawatts or dollars — we chart status and reasons and surface only the disclosed figures. The 6 market reactions show Wall Street's mirror of the build-out.
Beyond the big facilities, the network edge spans 333 points across 160 countries — the thin global layer that puts content near users, while heavy compute lives in about a dozen countries. And every data center plugs into the real power system: 34,936 plants totalling 5.7 TW. The generation mix is shown as context — no facility is ever linked to a specific plant.
Build on this data
Every figure here comes from an open dataset you can query directly — facilities, operators, type, and estimated energy, carbon & cost, plus 34,936 power plants — over a simple REST API.
Method & sources
This dashboard visualises the same open dataset that powers the interactive map. It is a curated, cited sample of named facilities and network edge points — not a complete census of every data center on Earth.
What's a count vs. an estimate
Counts (facilities by operator, state, country, type; documented community cases; project statuses) are exact row counts of records, each of which cites its public source. Estimates (energy, power draw, carbon, water, electricity cost, build cost, farmland food-foregone) come from a shared model and are labelled as such on every chart.
The energy, cost & carbon model
Per-facility electricity is estimated, not metered. Each facility is assigned a representative total draw and load factor by type — Hyperscale 40 MW @ 0.85, Colocation 5 MW @ 0.55, Edge PoP 0.3 MW @ 0.6 — so daily kWh = MW × 1000 × 24 × load factor. The per-facility draws are calibrated against the IEA 2024 anchor of ~415 TWh/yr for data centers (~1.5% of world electricity) — so a full census would approach it; this curated 250-site sample sums to only ~10% of that global total, which is why every aggregate here is a floor, not a world figure. Carbon multiplies modeled kWh by approximate national grid intensity (gCO₂/kWh, Ember / Our World in Data ~2023); cost multiplies by approximate national industrial power price (USD/kWh, GlobalPetrolPrices / IEA 2023–24); build cost uses a per-MW range from JLL / Turner & Townsend cost indices. Water uses LBNL water-use-effectiveness factors (1.8 L/kWh direct, 9.4 L/kWh including upstream generation). Farmland flags facilities whose coordinates fall inside a major agricultural region and applies USDA ERS / Our World in Data yields. These are order-of-magnitude figures for awareness and comparison, not billing or compliance data.
Deliberate restraint
Where the data is too sparse to support a number, we don't invent one: we don't total the gigawatts or dollars of cancelled projects (most don't disclose them), we don't rank two same-type facilities against each other, and we don't read an edge-network operator split as market share (it reflects which vendor publishes its PoP list). The construction pipeline's MW and cost are reported per project, so those we sum.
Data sources
Facility locations from OpenStreetMap (ODbL) and cited provider/press sources; power plants from the WRI Global Power Plant Database (CC-BY-4.0). Compiled 2026. Figures are estimates where noted.