Every chatbot answer is made by a very long chain of physical things: sand, light, memory, glue, drills, pipes, turbines. This guide follows one AI chip from sand to your screen, names the big players at each step, and explains how each one gets paid.
Nib saysHello. I am Nib, and I have dug through every layer of this mountain. The short version: goblins who sell shovels in a gold rush get paid once per goblin. Goblins who sell candles get paid every night the goblins dig. There are five kinds of goblin in this mountain, and knowing which one you are looking at explains most of what follows.
Your question travels to a data centre and lands on a rack of AI chips. The flagship NVIDIA rack today, the GB200 or GB300 NVL72, holds 72 GPUs wired together so they behave like one giant chip. One rack draws roughly 120–140 kilowatts, about as much power as 100 American homes. It is cooled by liquid, not fans, and it is reported to cost around $3 million.
A large AI campus holds thousands of these racks and needs a gigawatt or more, the output of a nuclear reactor. By 2025–26, Microsoft, Amazon, Alphabet and Meta together were spending several hundred billion dollars a year building this. Every company in this guide is, one way or another, catching a share of that spending.
Ultra-pure silicon is grown into a crystal ingot and sliced into 300 mm discs called wafers.
Shin-Etsu, SUMCO, GlobalWafers, Siltronic
In a fab, the wafer goes through hundreds of steps over about three months: coating, printing patterns with light, etching, depositing thin films. The finest layers are printed with extreme-ultraviolet (EUV) light from machines that cost well over $150 million each and come from exactly one company. An AI GPU die is huge, so only a few dozen fit on one wafer.
Foundry: TSMC (far ahead), Samsung, Intel · Tools: ASML (the only EUV maker), Applied Materials, Lam Research, Tokyo Electron, KLA (inspection)
AI chips starve without fast memory. HBM (high-bandwidth memory) stacks 8 to 16 thin DRAM chips on top of each other, connected by tiny vertical wires drilled through the silicon.
SK hynix (the leader), Samsung, Micron · Stacking bonders: Hanmi, ASMPT, Hanwha
Needles on a probe card touch each chip on the wafer to find the bad ones before anything expensive is attached. The industry calls this finding the "known good die".
Probe cards: FormFactor, Technoprobe, MJC · Testers: Advantest, Teradyne
The GPU dies and HBM stacks are placed side by side on a slab of silicon, then the whole thing is mounted on a package substrate. It has its own section below.
TSMC (most of it), ASE/SPIL, Amkor · Tools: Shibaura, ASMPT, BESI, Towa, Disco · Substrates: Ibiden, Unimicron, with film from Ajinomoto
Burn-in runs each chip hot for hours so weak ones fail in the factory, not in the data centre. Then system-level test runs it like a real computer for hours more. Each chip sits in a custom socket during both.
Sockets: Enplas, Yamaichi, WinWay, ISC · Test systems: Advantest, Teradyne, Aehr · Test houses: KYEC
AI boards are thick slabs with 20 to 40 layers of copper and special low-loss laminate. They are riddled with tens of thousands of drilled holes, and the hard materials wear drill bits out fast.
Boards: Victory Giant, Gold Circuit, TTM, WUS · Laminate: Elite Material · Drill bits: Dingtai, Union Tool, Topoint
Contract manufacturers assemble trays of GPUs, CPUs and switch chips into a finished rack with cold plates, pipes and power shelves.
Foxconn (Hon Hai), Quanta Computer, Wistron/Wiwynn, Supermicro, Dell, HPE
Thousands of racks must talk to each other at enormous speed. Inside a rack, copper; between racks, light carried by optical transceivers.
Switch chips: NVIDIA, Broadcom · Switches: Arista, Cisco · Optics: Coherent, Lumentum, InnoLight, Eoptolink, Fabrinet · Cables and links: Amphenol, Credo, Astera Labs
Transformers, switchgear, backup generators, batteries, chillers and cooling towers. Grid connections now take years, so some campuses build their own gas turbines or fuel cells.
Power gear: Schneider, Eaton, Vertiv, ABB, Delta, Hitachi Energy · Cooling: Vertiv, Trane, Munters · Generation: GE Vernova, Siemens Energy, Mitsubishi Heavy, Bloom, Caterpillar
Cloud companies rent the chips to AI labs and businesses. The lab's model reads your question, predicts the answer one token at a time, and sends it back. That last step is paid per token or by subscription.
Clouds: Microsoft Azure, AWS, Google Cloud, Oracle, CoreWeave, Nebius · Labs: OpenAI, Anthropic, Google DeepMind, Meta, xAI
CoWoS stands for Chip-on-Wafer-on-Substrate, and the name describes the two assembly steps. First, the chips go on a wafer: GPU dies and HBM stacks are placed onto a thin slab of silicon, the interposer, which is cut from a wafer and carries millions of microscopic wires between them. Second, that assembly goes on a substrate: a sturdier package base that connects to the circuit board. The point is distance. Memory millimetres from the GPU can feed it data many times faster than memory elsewhere on the board.
Nib saysThis is why a bonder maker can disappoint even in a boom. It sells the machine for step 1. TSMC buys those machines when it adds new CoWoS capacity, then uses them for years. More GPUs made next year does not mean more bonders sold next year. It means more machines only if capacity keeps growing faster and faster.
Each card names the largest listed companies with tickers, the chokepoint where one supplier holds most of the power, and which goblin it is, meaning how it gets paid (explained in the next section). The five goblins are shovel, blacksmith, blueprint, candle and toll-bridge.
Every company in the stack is paid in one of five ways. The dotted line in each sketch is AI activity, which grows fast and then levels off. The coloured line is what happens to that goblin's revenue. The AI boom is the same in every sketch, and the outcomes are very different. This is why CoWoS volume can keep growing while a bonder maker's revenue does not.
The shovel goblin is paid when capacity is added. Revenue follows the speed of building, so it rises and then falls even while activity stays high.
Bonders, lithography tools, transformers, construction
The blacksmith goblin is paid for every item forged. Revenue follows production, with swings when chip generations change. It is the biggest goblin by money.
NVIDIA, TSMC, HBM makers, contract manufacturers
The blueprint goblin is paid for each new chip design. Custom parts are thrown away when the design changes, not when they wear out. Revenue dips if volumes flatten.
Burn-in sockets, probe cards, photomasks
The candle goblin is paid as the product burns down. Consumption follows the level of activity, and harder materials push it higher. The risk moves to competition and price.
Drill bits, polishing pads and slurries, test contacts
The toll-bridge goblin is paid every time someone crosses. Rent, subscriptions, tokens, royalties, electricity bills, service contracts.
Clouds, AI labs, power sellers, landlords, maintenance
Nib saysA candle that everyone can make becomes cheap. A bridge with no other way across becomes a toll road. And even the best bridge is a bad investment if you paid too much for it.
The shovel, blueprint, candle and toll-bridge names came from a conversation with ChatGPT. The blacksmith was added here.