Every search, video and AI answer runs in a building packed with computers, power gear and cooling. Walk through one, meet the biggest real sites, design and run your own, and see who owns the most.
Since you opened this page, the world's data centers have used0.0MWhAbout 15 MWh every second, from the IEA estimate of 485 TWh used in 2025. A typical US home uses about 10.5 MWh in a year.
12,960data centers in 181 countriesData Center Map, Nov 2026
1,300+hyperscale sites, 58% run by Amazon, Microsoft, GoogleSynergy Research
1+ GWIT power at the largest AI campus, Amazon's Project Rainier in IndianaEpoch AI estimate, Oct 2026
~20MH100-equivalent AI chips in use worldwideEpoch AI, mid-2026
Anatomy of a data center
U16 Patch panel
A data center is a building designed around three flows: electricity in, heat out, and data in both directions. Click any part of the building, or switch on a flow to trace where it goes.
PowerCold air / waterHeatData
Scroll sideways to see the whole building.
Inside the data hall
Server racks
The four systems inside
U15 Leaf switch
Every data hall runs on the same four systems. Compute does the work, storage remembers, the network connects everything, and power and cooling keep it all alive.
CPU or GPU?
A server CPU has dozens of powerful cores (the biggest server chips now have 128 to 192) and is good at varied, step-by-step work: websites, databases, operating systems.
A GPU has thousands of simpler cores that apply the same math to huge blocks of numbers at once. AI is mostly that kind of math, multiplying large grids of numbers, so AI data centers are built around GPUs and similar accelerators such as Google's TPUs and Amazon's Trainium.
Race the two chips on a job that splits into 1,024 independent pieces, then on one where every step needs the answer from the step before.
CPU, 16 fast coresreadyGPU, 1,024 simple coresready
Each square is one piece of work. Time is in clock ticks; a GPU core is slower per tick but there are far more of them.
One AI rack, by the numbers
NVIDIA's GB200 NVL72 is the rack many 2025–2026 AI clusters are built from. It behaves like one giant GPU.
GPUs
72
Blackwell, plus 36 Grace CPUs
Compute trays
18
plus 9 NVLink switch trays
GPU-to-GPU link
1.8 TB/s
per GPU, 130 TB/s across the rack
Power draw
120–140 kW
about 10× a typical cloud rack
Cooling
Liquid
cold plates on every chip
Next: Vera Rubin (late 2026)
170+ kW
per Meta’s first racks; 600 kW Kyber racks planned for 2027
Counting AI chips
Chips differ, so analysts convert them into H100-equivalents: how many NVIDIA H100 GPUs it would take to match a chip's 8-bit math speed.
By Epoch AI's count, the world had about 15 million H100e in January 2026, drawing more than 10 GW, and roughly 20 million in use by mid-2026. About 27.6 million had been sold by June 2026.
Only five companies (Amazon, Google, Meta, Microsoft and Oracle) held an estimated 71% of all AI compute at the end of 2025. See who owns what.
How far away is your data?
Typical round numbers for a modern server; bars use a log scale, so each step to the right is about ten times slower.
Hot, warm and cold
Hot data you need now lives in memory and flash (NVMe SSDs): microseconds away.
Warm data, such as last year's photos, sits on hard drives: cheaper per terabyte, milliseconds away.
Cold archives go to tape libraries, where a robot fetches a cartridge in about a minute. Tape is still the cheapest way to keep data for decades.
AI adds a fourth layer: HBM, memory stacked next to the GPU chip that moves 3 to 8 terabytes per second so the model's weights can be read for every word it writes.
Break drives, keep the photo
In a big data center, drives fail every day. Your photo survives because it is stored in pieces across three separate zones. Click drives to break them, or take out a whole zone.
Photo readable
A spine-leaf fabric
4 of 4 paths upFull capacity between every pair of racks.
Click a spine switch (top row) to fail it. Traffic spreads over whatever paths remain.
Most traffic never leaves the building
Inside a data center most packets go from server to server ("east–west"), not out to the internet. A spine-leaf network gives every pair of racks several equal-length paths, so one broken switch only trims capacity.
AI clusters run two extra networks. Scale-up links GPUs inside a rack: NVLink moves 1.8 TB/s per Blackwell GPU. Scale-out links racks across the hall over InfiniBand or Ethernet at 400 to 800 Gb/s per port.
Connections to the outside world arrive by carrier fiber at the meet-me room, where networks hand traffic to each other.
Where 100 kWh goes
PUE (power usage effectiveness) is total facility energy divided by the energy that reaches the IT equipment. 1.0 would mean zero overhead.
The split of the overhead between cooling, power conversion and lighting is an illustrative assumption. Average 1.54 from the Uptime Institute 2025 survey.
Which cooling can handle the rack?
Approximate practical limits per rack; real limits depend on the room and the equipment.
Types of data centers
U14 Storage shelf
Data centers range from a closet with a few servers to AI campuses that draw as much power as a city. Pick a type to see where it sits on the scale.
From closet to gigawatt
IT power per site, log scale. 1 MW powers roughly 1,000 US homes.
Tiers: how much downtime is allowed?
The Uptime Institute rates a facility's design from Tier I (one path for power and cooling, no spares) to Tier IV (everything doubled, so any single failure is survivable).
Most colocation and cloud facilities are built to Tier III: every component can be serviced without switching off the IT load.
Availability figures are the commonly cited design targets for each tier.
Major data centers, up close
U12 1U server
Twenty-one real sites, from Google's first company-built data center and a server farm on the seabed to AI campuses that went from bare land to nearly a gigawatt in under two years. Pick a site for its specs, or tick up to three to compare.
AI factoryCloud and internetRecord-breaker or experiment
Power running today, MW
Epoch AI estimates of IT power for the US AI campuses; company figures for Cortex, Yotta and Inner Mongolia. Pick a bar for details.
Side by side
The AI build-out race
IT power at nine of the biggest AI campuses, estimated from satellite images, permits and company statements. Each step is a building coming online; dashed lines are projections.
Follow work through the machine. Send a request across the internet, watch GPUs turn your prompt into an answer, or train a model on a cluster.
What are you doing?
Where is the data center?
Scroll sideways to see the whole path.
Event log
Where the time went
Model size (parameters)
GPU
Checking for live answers
Tokenizesplit into pieces
Schedulejoin a batch
Prefillread the prompt
Decodeone token at a time
Streamback to you
Press send to watch your prompt go through a GPU server.
Tokens
Math units busy0%
Memory bandwidth busy0%
Your KV cache0 MB
Batch slots
Answer
Speed and energy come from a simplified model of the hardware you picked. For comparison, Google measured 0.24 Wh for a median Gemini text prompt (2025) and OpenAI puts an average ChatGPT query at about 0.34 Wh.
Start from
GPU type
Compute uses the standard estimate of 6 × parameters × tokens. Frontier-scale is an illustrative setting; labs don't publish these numbers.
Ready
Training loss
Loss measures how wrong the model's next-word guesses are. It drops fast, then slowly.
Real clusters fail often: Meta reported 419 unexpected interruptions in 54 days while training Llama 3 405B on 16,384 GPUs, most of them hardware faults. Regular checkpoints limit how much work is lost.
Build your own data center
U7 Expansion bays
Build a data center step by step: choose a place, pick the computer parts, fill a room with racks, plug it in, keep it cool, and copy it into a campus. Then run it for a year and get a report card. Every choice changes the bill, the carbon, the water and whether it stays online.
Choose a starting design below, or start from an empty hall.
Go through the six steps on the left. The picture on the right changes as you do.
Press “Run one year” at the bottom to see what happens and get your grades.
Words to know
Rack
A tall metal cabinet, about the size of a fridge, filled with servers.
Hall
A big room full of racks, with cooling and power around it.
GPU
A chip that does the math for AI very fast, thousands of sums at once.
RAM
A computer’s short-term memory: the desk where it keeps the work it is doing right now.
Storage
Where data is kept for the long term, on hard drives or flash. The warehouse.
kW and MW
Units of power. 1 MW is 1,000 kW, about what 800 US homes use on average.
H100e
AI computing power counted in NVIDIA H100 chips. 1,000 H100e is the math of 1,000 H100s.
PUE
Total power divided by computer power. 1.2 means 20% extra goes to cooling and losses.
Where to build. Cool weather makes cooling cheap, cheap and clean electricity lowers the bill and the CO₂, and the local grid decides how much power you can get.
Average temperature by month
Site figures are rounded estimates for this sandbox. Fast grid power is a rough guess at what the local utility could connect within about two years.
What goes inside the racks. The GPU is the brain that does AI math. RAM is the desk where a server keeps the work it is doing right now. Storage is the warehouse that keeps all the data.
GPU for the AI racksBars show 8-bit AI math per GPU compared with an H100, then memory per GPU.
RAM in each cloud server
Each cloud server has 128 CPU cores and is split into small virtual computers (VMs) that customers rent. A small VM needs 1 core and 8 GB of RAM.
Storage type
Two kinds of memory: servers use DDR5 RAM, while each GPU carries HBM, memory stacked right beside the chip. HBM is many times faster but far more expensive, so a GPU has much less of it. Google (TPU) and Amazon (Trainium) also design AI chips, but use them only in their own clouds. Specs are rounded; prices are rough late-2026 estimates, after an AI-driven memory shortage pushed RAM and flash prices up.
Fill the room with racks. Pick a rack below, then click the floor. Every rack turns its power into heat, so each one needs cooling nearby, and network racks connect them all. Watch the dot on each rack: green means cooled.
Wall air handlers, 200 kW each
Plug it in. The servers need power every second of every day. Connect to the grid, add backup generators and batteries for outages, and build your own gas or solar plant if the grid can’t supply enough.
peak
Backup design
Keep it cool. Almost every watt the servers use becomes heat, and it has to leave the building. Choose how: with chillers, evaporating water, outside air, giant radiators or cold seawater.
How heat leaves the building
ASHRAE recommends 18–27 °C at server inlets. Warmer air means more hours of free cooling and less chiller time.
Make it a campus. Big data centers copy one hall design many times. Choose how many halls and buildings, and how busy the servers will be.
AI training keeps servers about 85–90% busy; cloud servers average much less.
Swipe sideways to see the whole hall.
Dot on each rack: green cooled, amber warm, red overheating
18 °C38 °C rack inlet air
Design check
Run it for a year
How the sandbox works
Each air handler adds 200 kW of cold air to the room. Without containment a rack can take only about 12 kW from room air; with contained aisles, about 25 kW. In-row coolers cover 70 kW within two positions, and CDUs move 1.2 MW of liquid cooling within six positions, both across the two rows that share a cold aisle. Liquid-cooled GPU racks send 80–95% of their heat to liquid. GPU racks use rounded vendor specs: 8-bit math is compared with an H100’s 1,979 dense TFLOPS (Vera Rubin figures are early), and the biggest model assumes 70% of GPU memory holds 8-bit weights. Cloud servers draw 500 W plus 0.1 W per GB of RAM, cost $14,000 plus about $20 per GB of RAM, and host one small VM per core or per 8 GB of RAM, whichever runs out first. Storage racks hold 36 TB hard drives, 122 TB QLC or 30.7 TB TLC flash drives; a checkpoint of a 405B-parameter model with its optimizer state is about 5.7 TB. Monthly cooling power follows each site’s average temperatures, estimated wet-bulb temperature, the cooling plant and the air temperature you choose. Power losses run 3.5–5% depending on the backup design. Costs use about $3M per MW for the building, $3.5–6M for electrical gear and $1.6–3.2M for cooling, plus the racks. AI compute uses H100-equivalents (32 per H100 rack, about 180 per GB200 rack, about 630 per Vera Rubin rack). In the yearly run, each generator fails to start 3% of the time, less with longer battery runtime because operators can retry. Everything is a simplified model for learning, not for engineering.
Experiment lab
U5 KVM console
Two quick experiments: test airflow in a cold aisle, or cut the utility feed and watch the backup systems take over. To design a whole facility, use the Build sandbox.
18 °C38 °C
Each dot is a parcel of air, colored by temperature.
All inlets in range
Server inlet temperatures, top to bottom. The line marks 27 °C, the top of ASHRAE's recommended range.
Without containment, hot exhaust curls over the top of the rack and gets pulled back into the top servers. Missing blanking panels let it leak straight through the middle.
Click a running generator to trip it.
Alarm log
Backup design
IT loadUPClock00:00.0Battery A
100%, charging
Battery B
100%
GeneratorsOff, standing by
Simulation speed
Batteries hold the load for about 5 minutes here; generators take roughly 10 seconds to reach speed. N needs all three generators, N+1 can lose one, 2N runs two complete power paths and every server has two power supplies.
Who has the most
U4 Compute server
Counting buildings, the United States is far ahead of every other country. Counting AI chips, five cloud companies own most of the world's computing power, and nearly all of the biggest AI campuses are in the US.
Circle area shows the number of data centers listed for each country. The US has more than the next 18 countries combined.
Ranking
Show as a table
Source: Data Center Map, 12,960 listed facilities, November 2026. Counts measure buildings, not capacity: one AI campus can outweigh dozens of small sites.
AI compute owned, in H100-equivalents
Estimated holdings of AI chips, converted to how many NVIDIA H100s would match them.
Sources: Epoch AI estimates reported by Network World (April 2026); share from Epoch AI, up from 63% in early 2024.
Google leads, mostly on its own chips
Google holds about a quarter of the world's AI compute, and around 4 million of its roughly 5 million H100-equivalents are its own TPUs rather than NVIDIA GPUs.
Microsoft, Amazon and Meta follow. Amazon's total leans on its own Trainium chips, which power Project Rainier in Indiana, the largest known deployment of non-NVIDIA AI compute.
AI labs such as OpenAI and Anthropic mostly rent capacity from these companies or build with partners. The single largest AI site, Colossus 2, belongs to xAI, now part of SpaceX: about 1.1 million H100e at one campus.
The new gigawatt belt
Circle area shows IT power. A 1 GW site draws about as much electricity as a million US homes.
IT power estimates from Epoch AI's Frontier Data Centers hub, built from satellite images, permits and company statements (latest estimates, August to November 2026; 2028 figures as summarized by Tech Jack Solutions). Planned figures can change. Profiles of each site are in Major sites.
2026 capital spending, latest company outlooks
Most of it goes to data centers, servers, chips and networking. Ranges show where a company gave a range.
Big four, 2025
~$410B
Big four, 2026 outlook
~$725–750B
Growth
about +80%
Outlooks as of late October 2026: Amazon about $220B; Alphabet $195–205B; Microsoft about $175B for calendar 2026 after a lease reclassification; Meta $135–145B. Sources: Bloomberg via FourWeekMBA, AI Weekly, Tom's Hardware / FT.
Building a country's worth every year
Epoch AI estimates a frontier AI data center costs about $44 billion per gigawatt of server power: roughly $30 billion for IT hardware and $14 billion for the building, power and cooling.
At that rate, the big four's 2026 budgets alone could fund on the order of 15 gigawatts of new AI capacity, more than the entire Northern Virginia market, the world's largest, which had about 4.5 GW in mid-2026.
Power is now the bottleneck: vacancy in Northern Virginia fell to 0.2%, and suppliers of power equipment report backlogs stretching past 2030.
The numbers
U2 UPS
How much power data centers use, how efficient they are, and where it is heading. Hover or tap any chart for exact values, or open its table.
Electricity used by the world's data centers
Demand roughly doubles by 2030 in every IEA scenario. In the base case it reaches 945 TWh, slightly more than all of Japan uses today.