The AI buildout has a physical layer.
A journal on the physical constraints behind the AI buildout: power, grid connection, data centres, cooling and the supply chains that gate them. Written to be understood by anyone.
Read the pieces below, or follow the power down from the grid to the chip.
-
Layer 01 · The grid connection
- Hong Kong has the plugIts grid builds to demand, and the chip needs a licence
- The equipment queueTransformer and switchgear makers are building to keep the wait, not to end it
- America rations the plugWho may be cut, who must hold on, who pays
- Johor and SingaporeThe plug, not the chip, decides who gets to build
-
Layer 02 · Power on site
-
Layer 03 · Power inside the building
-
Layer 04 · Cooling
-
Layer 05 · The network between chips
-
Layer 06 · Memory
-
Layer 07 · Chips and the factories that make them
-
Layer 08 · Who runs the capacity
-
Layer 09 · Data centres in orbit
Or follow the power down, stop by stop.
01 · The grid connection
It starts with a wire
Power for an AI data centre arrives the way it arrives anywhere else: over high-voltage lines from the public grid.
01 · The grid connection
The substation is the gate
High-voltage power arrives from the public grid through a substation built for the site.
What binds. The wait for a connection, and the transformers and switchgear behind it, now runs to years.
01 · The grid connection
Then underground, to the site
The last stretch usually runs as buried cable into the site itself. Follow it in.
02 · Power on site
The site, up close
Close up, the site is more than one building. The grid connection is one way in, and more sites now add a second.
02 · Power on site
Making power next door
Turbines, fuel cells, batteries and, one day, small reactors make or store power beside the building.
What binds. Skipping the grid queue means joining the queue for the equipment instead.
03 · Power inside the building
Inside the building
Power is stepped down, backed up and carried to each rack, then down to below one volt at the chip.
Switchgear and batteries fill one room. Copper busways carry power overhead to every rack, and coolant runs beneath the floor.
03 · The network between chips
Light between the racks
Thousands of chips swap partial results constantly, over copper inside the rack and light everywhere else.
What binds. A slow link leaves the most expensive silicon in the building waiting.
04 · One rack
One rack, seventy-two chips
In Nvidia's flagship rack, 72 Blackwell processors sit in compute trays above and below a block of switch trays. The switches join all 72 so they act as one machine.
04 · Power inside the building
Power down the back
Each rack is fed from a copper bar running down its back, with power shelves converting the supply for the trays.
What binds. At a megawatt per rack, the copper needed at today's low voltage stops fitting.
04 · Cooling
Heat out the side
Almost every watt that goes into a chip comes back out as heat, and liquid now carries most of it away.
What binds. Air-cooled racks ran 10 to 40 kilowatts; the AI racks replacing them draw 140.
05 · One tray
Four chips on a tray
Slide out one compute tray. Two boards each carry a Grace processor and two Blackwells, with network ports along the front.
05 · Power inside the building
Below one volt, right beside the chip
Power stages ringed around each processor take the supply down to below one volt. Cold plates sit on top, with coolant running through them.
06 · The chip
Two dies acting as one
Lift out one B200. At its centre sit two slabs of logic made by TSMC in Taiwan, printed with light from machines only ASML in the Netherlands builds.
06 · Memory
Memory beside it, wiring beneath it
Eight stacks of high-bandwidth memory flank the dies. Everything rests on TSMC's packaging, then on a substrate built in Japan or Taiwan around Ajinomoto's film.
What binds. For most of the time a model is answering, the chip is waiting on memory, not maths.
07 · Chips and the factories that make them
Printed with light
Each layer of the die is printed by a machine that makes extreme ultraviolet light by firing a laser at droplets of tin. Mirrors carry that light to a mask, then onto the wafer.
07 · Chips and the factories that make them
Many dies at once
Every wafer carries many dies, printed layer by layer. ASML describes the technology as unique to itself, and its mirrors come from ZEISS.
What binds. A new chip factory takes two to three years to build and one to two more to fill.
08 · The clock on each part
Where the time goes
Some parts could be bought elsewhere within months. Others need a new factory or years of testing first. The slowest link sets the pace for everything above it.
09 · Who runs the capacity
All of it, for one answer on your screen
The companies that turn power, buildings and chips into computing that someone else can use.
What binds. Buyers pay for megawatts switched on, not megawatts signed.
10 · Data centres in orbit
Or take the whole hall off the planet
Satellites in near-constant sunlight run AI chips on solar power and shed the heat by radiating it into cold space. No grid connection, no queue, and the sun almost never sets.
10 · Data centres in orbit
The plug is free, and the heat is the bill
In a vacuum there is no air or water to carry heat away, so it can only glow off radiators. Every kilogram of radiator and panel has to be launched.
What binds. Every kilogram of radiator and solar panel must be launched, so the weight of each kilowatt decides the cost.