The physical stack behind artificial intelligence
AI is built from power, water, silicon, fiber, and time.
Every model query, training run, and AI agent depends on a vast physical system: transmission lines, substations, transformers, backup generation, liquid cooling loops, GPU racks, optical networks, chip supply chains, and the people who build and operate them.
Interactive educational model. Representative configurations. Not a blueprint for a specific facility.
Start here
Three ways in.
- Walk the campusA guided 3D tour from the power line to the GPU racks and out to the cooling towers. Click anything to learn what it is.Start the tour
- Where the power comes fromGrid, solar and batteries, gas turbines and small nuclear reactors, and how each one makes electrons.See the energy
- Inside a GPUZoom from a data hall into a single Blackwell chip, all the way down to the silicon atoms.Zoom in
Who it is for
Anyone who wants to understand what AI is physically made of: investors, policy people, students, builders. Engineers get the full specs in Pro mode. It is a learning model, not a design tool.
How accurate it is
Every number is tagged: vendor specs are exact, reported and typical figures are real ranges, live rates are physics estimates, and the 3D layout is illustrative. See how to read them
Right now, in a 300 MW campus
Electrons, water and signals, every second.
Live · 300 MW campus model
util 100%Electrons into the campus
2.82×10²²/s
≈1,506 A per phase at 138 kV
0 since you arrived
Through one rack busbar
1.56×10²²/s
2,500 A at 50 V DC · 125 kW
Water evaporated, campus
83.3 L/s
Pod coolant loop 26.7 L/s · out at 47 °C
0 L since you arrived
Signals on the AI fabric
113 Pb/s
NVLink inside 16 racks: 16.6 Pb/s
0.0 b moved since you arrived
Illustrative rates from the live pod model. In AC lines electrons mostly jiggle in place at millimeters per second; energy moves through the field near light speed. Charge rate is current divided by 1.602×10⁻¹⁹ C.
Four eras
From the computer room to the AI factory.
01 · 1990s to mid-2000s
Pre-cloud data center
The enterprise computer room.
1–5 kW/rack
Bottleneck: floor space and cooling distribution inside an existing building
02 · Mid-2000s to late 2010s
Cloud data center
The warehouse-scale computer.
5–15 kW/rack
Bottleneck: land, power and network reach in the right metro
03 · 2020s onward
Hyperscaler AI data center
A power plant's worth of accelerators.
30–150+ kW/rack
Bottleneck: grid interconnection, transformers and switchgear, and accelerator supply
04 · Current and emerging
Neocloud AI compute campus
The AI factory as a service.
40–150+ kW/rack
Bottleneck: gpu allocation, energized power and the speed of capital