The world runs on rock · Story 2 of 12

The Mind AI runs on metal

Every answer an AI model gives is computed on chips, in racks, in buildings that draw as much power as a city. Behind the software are copper, silicon, gallium and steel, a few factories in Taiwan, and a grid that has to grow to feed it.

Scroll to open the rack

Take it apart

One rack, 72 chips, one machine

This is a model of a rack built the way Nvidia's GB200 NVL72 is: 72 Blackwell GPUs and 36 CPUs in eighteen compute trays, joined through nine switch trays so they work as one giant accelerator, and cooled by liquid rather than air.³

Two miles of copper

Down the back runs a spine of more than 5,000 copper cables, about two miles of it, moving data between the chips at 130 terabytes a second.² Inside the rack, copper does this job, not optical fibre.

Layer by layer

Power shelves turn grid power into the low voltages chips use, through copper busbars
Compute trays the GPUs and CPUs: silicon, with the chip's dozen minerals
Cold plates copper blocks on every chip, with water running through them
Switch trays chips that connect every GPU to every other
Coolant lines carry the heat out to the building's cooling plant

Down to the chip

The chip from Story 1, times 72

Pull one tray and the accelerator from The Chip comes out: silicon, hafnium, tungsten, cobalt, copper and tantalum, 208 billion transistors at a time. A big AI data centre holds thousands of these racks.

SiCuHfWCoTaGaNiSn

A long aisle between rows of black supercomputer racks with honeycomb front panels — NASA's Pleiades supercomputer, Ames Research Center, 2008. Photo: NASA (public domain).

Rows upon rows

Then multiply by thousands

A data centre is aisle after aisle of racks. This is NASA's Pleiades supercomputer in 2008, long before the AI boom. Today's AI racks pack far more chips and power into the same floor space, which is why they need liquid cooling.

The power bill

415 TWhused by the world's data centres in 2024, about 1.5% of all electricity

945 TWhby 2030 in the IEA's base case: slightly more than Japan uses today

International Energy Agency, Energy and AI (2025).¹

The transformer wait

Copper and steel, in oil

Every new hall needs transformers: a core of thin silicon-steel sheets, copper windings wrapped round it, the whole thing sitting in oil. Wood Mackenzie says power transformer prices have risen 77% since 2019, and some high-voltage units now take years to arrive, up from about one in 2020 and 2021.⁵

Metal for the minds

The IEA's first estimate of data centres' mineral needs: by 2030 they could use more than a tenth of today's world gallium supply, for chips and power electronics, and China refines about 99% of it. Copper, aluminium, silicon and rare earths are on the same list.¹

The United States at night seen from space, cities glowing gold along a dark blue land — The United States at night, 2012. NASA Earth Observatory / Suomi NPP (public domain).

On the grid

Campuses the size of power stations

The new AI campuses plug into the same grid that lights these cities. Some are planned at one to five gigawatts each, as much as one to five large power stations produce.^6,7,9

Where the minds live

1 · Made in Taiwan

Nvidia's Blackwell chips are made by TSMC in Taiwan.⁴ Almost every AI lab's computing starts in the same few fabs.

2 · Project Rainier, Indiana

Amazon's campus at New Carlisle runs nearly 500,000 of its Trainium2 chips for Anthropic's Claude, and is planned to draw up to 2.2 gigawatts when finished.⁷

3 · Hyperion, Louisiana

Meta's campus in Richland Parish is planned at 5 gigawatts, with more than $50 billion of investment.⁹

4 · Stargate, Texas

OpenAI announced Stargate in January 2025 with SoftBank, Oracle and MGX: up to $500 billion of AI infrastructure in the US, with its first site at Abilene, Texas.¹⁰

5 · Anthropic

In October 2025 Anthropic agreed to use up to one million Google TPUs, well over a gigawatt of capacity in 2026.⁶ A month later it announced $50 billion for its own data centres in Texas and New York, built with Fluidstack.⁸

6 · Hangzhou

DeepSeek and Alibaba's Qwen team publish open-weight models that anyone can download and run.^11,12 They are trained on the same kinds of chips, under export rules that limit which ones they can buy.

7 · Gallium

China refined 99% of the world's primary gallium in 2025.¹⁵ Data centres could need more than a tenth of today's supply by 2030.¹

8 · Copper

The wiring, cold plates, busbars, cables and transformers are copper. Chile mined 23% of it in 2025, DR Congo 14%.¹⁹

9 · A Japan of power

The IEA's base case for data centres in 2030, about 945 TWh, is slightly more than all the electricity Japan uses today.¹

The two-way chokepoint

Chips one way, minerals the other

The US and its allies make the most advanced AI chips and the machines that print them. China refines most of the gallium, germanium, graphite and rare earths those industries need. Since 2022 each side has put rules on what the other can buy.

October 2022

The US requires licences to sell China the most advanced AI chips and the tools to make them.¹³

August 2023

China puts gallium and germanium exports under licence.¹⁵

October 2023

The US closes the gap that let Nvidia sell slower versions of its chips, the A800 and H800, to China.¹⁴ DeepSeek's V3 had been trained on H800s.

December 2023

China requires licences for exports of battery-grade graphite.¹⁶

December 2024 → November 2025

China bans gallium, germanium and antimony exports to the US;¹⁷ in November 2025 it suspends that ban until 27 November 2026, with licences still required.¹⁸

Most of these rules are licences, not outright bans, so each gate in the model holds back part of the flow, not all of it.

Counting chips

2,048

DeepSeek says it trained its V3 model on 2,048 Nvidia H800s. It put the final training run at about 2.8 million GPU hours, or about $5.6 million at rental prices, a figure that leaves out earlier research and experiments.¹¹ One cube is one chip.

Project Rainier

~500,000

Trainium2 chips running Anthropic's Claude in Indiana, according to Amazon.⁷ Each dot is 4 chips.

And more coming

1,000,000

Google TPUs in Anthropic's October 2025 agreement: "up to" a million.⁶ Whether two thousand or a million, every one of them is silicon from the same few refineries, wired, cooled and powered through copper.

Why "critical"