Glossary

It adds thousands of numbers at the same time, over and over. That is exactly what AI training needs. This one is made only for a data centre. It holds far more memory than a games card, and it is meant to run for years without a break. It is kept apart here from the card in a home computer, because the two are sold to different people at prices that are not even close.

Think of it as: A ship engine, not a car engine. Same idea, but made to run day and night for years.

It was built for games. Drawing a moving 3D world means doing thousands of small steps at the same time. That is exactly what AI needs too. So a games part ended up in the middle of it all, and the world ran short. The one made for a data centre is called an accelerator, and it costs far more.

Think of it as: A hundred people each doing one small step at once, instead of one person doing all of it.

A machine keeps whatever it is working on in memory, because it can reach it there far more quickly. Memory forgets everything the moment the power goes off. RAM, DDR and desktop memory all mean this same thing. HBM is a much quicker kind, made in the same plants.

Think of it as: Desk space. The store room holds everything; your desk is what you can spread out in front of you right now.

An AI chip can only work as quickly as memory can feed it. So this kind is stacked, one on top of another, and put right beside it. Here is the catch. It is built in the same plants, on the same wafers, as the memory a home computer uses. More of one means less of the other. That is how AI quietly pushed up the price of memory in your home.

Think of it as: Keeping what you need right by your hand, not across the room.

This is where your work and your pictures sit when nothing is using them. Memory forgets when the power goes; this does not. NAND is the part itself. An SSD is a drive made from it, and NVMe is the fast way it joins the rest of the machine.

Think of it as: If memory is your desk, this is the store room behind you.

A general chip is good at everything and best at nothing. A chip built for one single kind of work can run far more quickly, and it costs less. That is why a big cloud company will design its own. The catch: if the work turns into something else, the chip is no use to anyone.

Think of it as: A key cut for one door. It opens that door beautifully and no other.

Nearly everything a computer does happens on one. Millions or billions of tiny switches are printed onto a thin square of very pure silicon, and those switches do all the adding, remembering and deciding. The word covers a whole family of things — the part that thinks, the part that remembers, the part that stores — and this atlas tracks several of them apart because they are made and sold very differently.

Think of it as: A city map printed so small it fits on your fingernail, with the roads carrying electricity instead of cars.

When something happens “in the cloud”, it happens in one of these. They hold thousands of machines in long rows, plus the power and cooling to keep them running. They are where nearly every AI chip in the world ends up, and they use enough electricity that where you put one is a real decision, not a detail.

Think of it as: A warehouse, but instead of boxes it is full of computers, and it never closes.

A chip can change how much power it pulls far more quickly than any power supply can follow. So hundreds of these sit right beside it, ready to fill in. Tantalum ones hold the most power for their size, and they keep working when hot. That is why a metal that is hard to find ends up right under every large chip.

Think of it as: A glass of water by the bed. It is there the second you want it.

More than a few feet of wire cannot carry data fast enough, so it is turned into light instead. A transceiver is the small plug at each end that makes the change. A network adapter is the card inside the machine that feeds it. When thousands of chips work on one AI job together, they talk all the time. So you need more of this as the group grows, not as the chips get better.

Think of it as: Like sending letters by train once they have to go a long way.

Chips are not made one at a time. They are built side by side on one round piece of silicon, then cut free at the end. Everything on that piece is fighting for the same space. So a factory that makes more of one product makes less of another.

Think of it as: One page of paper cut into many small cards. The page only holds so many.

The word for one single chip, once it has been cut out. Hundreds are built on one wafer and then cut free. Each one is a die. Some do not work at all, and yield is the count of the good ones.

Think of it as: One card, cut from the page of many.

Short for fabrication plant. The most advanced ones — the only places the best AI chips can be made — you could count on one hand. A new one costs tens of thousands of millions of dollars and takes two to three years. So when the world suddenly wants more, it has to wait. 'Leading edge' means the newest kind of work those plants can do.

Think of it as: One of these costs more than a great bridge, and needs about as long to build.

Building the chips is not the last step. An AI part is really several parts in one: the chip that does the work, and layers of memory beside it. They must be joined to a single base, with wire as thin as a hair. This step, not chip building, is where the whole chain slows to a stop. Adding more of it needs two to three years.

Think of it as: The parts are all made. This is the step that puts them together as one.

Nothing comes out right every time. Dust and tiny marks in the silicon mean some of every run is no good. Lift this number by a few points and the world gets far more chips than you would guess. That is why it matters much more than its size suggests. The word also has a money meaning, but not here.

Think of it as: Like a class test. What counts is not how many sat it, but how many passed.

A chip is built up in layers, one at a time. Light draws the fine detail of each layer on to the silicon, which is then washed clean. The best tools use a light so short that a line can be a few atoms wide. Only one company on earth can build those tools. Deposition, etch and metrology are the steps on either side: they add material, cut it away, and check the result.

Think of it as: Hold a picture up to a light and it lands on the wall. Same idea, but far smaller.

Ore comes out of the ground as mostly rock. It is ground down and washed until far more of it is metal. That is a concentrate: a dust that gets sold and moved around the world. Smelting and refining later turn it into metal you can use. Those steps often happen in a different country, so a country can mine a material and still not control it.

Think of it as: Like washing river sand to keep the gold. What is left is what you carry home.

Smelting uses heat to drive off everything that is not the metal. Refining then cleans up what is left, often with electric power. It is the step that eats the power. And it is done in only a few parts of the world. Bauxite is the rock that aluminium starts as. Alumina is the white powder in between. A cathode is the finished copper at the end.

Think of it as: Cooking the rock until the metal runs out of it.

A chip is not just its tiny working parts. It also holds ten or more layers of very fine copper wire that join them up. The wire that runs up between layers is called a via. All this wire is a big part of why so much copper and tungsten is needed at all. How much the wire slows the power down is one of the things that hold back speed.

Think of it as: The streets of a city, laid on top of each other in a dozen layers.

In the factory it means the piece of material the circuit is grown on, nearly always silicon. In packaging it means the small, very fine board a finished chip sits on. That board carries hundreds of tiny joins out to something a computer can take. The word is used both ways here.

Think of it as: The floor a house stands on — or the board you carry it about on.

Nearly every join inside a computer is a small drop of this. That is why about half of all the tin in the world ends up here. Die attach is the same idea one size down: fixing a chip to its base so heat can get out. Where solder cannot take the heat, silver that has been sintered is used instead.

Think of it as: It works like glue, but it is made of metal and power can run through it.

A factory has a ceiling. Once it is running flat out, more orders do not produce more goods — they produce a queue and a higher price. Adding capacity means building more factory, which for the most advanced chips takes two to three years and enormous sums of money. This is why a sudden rush of demand shows up as a price rise rather than as more chips.

If people suddenly want far more of something than usual, that is a demand shock. What happens next depends on whether the sellers can make more. If they can, the price barely moves. If they cannot, the price is the only thing left that can change — which is most of what this atlas is about.

Memory is the clearest example. Prices rise, so everyone builds new factories. The factories all finish at once, there is suddenly too much supply, and prices crash — which stops anyone building, until shortage sends prices up again. It repeats for decades. A business in a cyclical market can be brilliant and still lose money at the wrong moment.

Drawing a chip and building one are two different lines of work. A fabless company does the first and pays a foundry to do the second. So it must fight for the same factory space as everyone else. It can be held back by how fast a new factory gets built, not by how much people want to buy.

Think of it as: You can draw the plan for a house. Someone else still has to build it.

A chain can only move as fast as its slowest part. Add more of anything else and nothing gets better until that one step opens up. So finding it matters more than counting the rest. Tightness is the 0 to 100 score used here: 50 means supply and orders are even, and above 50 means people are fighting over too few units.

Think of it as: A road with a bridge only one car wide. The rest of the road does not matter.

For a games card, days. For a large transformer, over two years. A long lead time is what turns a rush of orders into a price rise, not into more goods. If the goods cannot get there in time, the price is the only thing left that can move.

Think of it as: Order a book and it comes tomorrow. Order a house and you wait years.

Stock already in the shops holds off the first hit, before any price change. Four months of it will hold off a wave of orders far better than one and a half months. So the same wave reaches the price of one part quickly and another part slowly.

Think of it as: Food in the house means one bad day at the shop does not matter.

Here it is given as years of use at the rate the world uses it now. It is not a count of the years left before it runs out. Reserves grow when the price goes up, when new ground is found, and when better tools reach more of it. They fall back when those go the other way. A material that showed 20 years has often showed 20 years for the last 50.

Think of it as: The food in the shop is not all the food in the world.

Gallium and germanium have no mine of their own. They come out while aluminium and zinc are being cleaned up. So the amount the world gets is set by how much of those other metals people want. If the world wanted two times as much, it could not just make two times as much. The price can rise a long way without one more load turning up.

Think of it as: The skin comes with the meat. To get more skin, someone must want more meat.

A transformer brings the high power of the grid down to something a building can use. Switchgear keeps it safe and sends it where it is needed. Busbars and PDUs carry it round inside, and a UPS holds the load through a power cut. None of it is new or strange — it is copper and steel. But one big transformer can take over two years to turn up, and that is often the real reason a new building cannot open.

Think of it as: The way power gets from the street into the wall. Nobody thinks about it until it is late.

A kWh is one kilowatt used for one hour — about what a kettle uses in an hour. An MWh is a thousand of those. A TWh is a million MWh, which is the size at which a whole country is counted. A watt tells you how fast power is being used; a kWh tells you how much has been used.

Think of it as: A watt is how fast the water runs. A kWh is how much water came out.

Computers turn nearly all the power they draw into heat. It costs yet more power to carry that heat back out. A figure of 1.2 means that for every 1 kW the computers use, the building pulls 1.2 kW from the grid. The other 0.2 goes on keeping things cold, and on power lost along the way. 1.0 means none of that, which is how a home computer in a room is counted here.

Think of it as: You pay once to run the machine, then again to carry its heat away.

It is counted in grams of warming gas for each kWh. A grid run on water or nuclear power sits under 100. A grid run on coal is over 600. The same machine can give off ten times as much, for no reason but where it is. So where you put a data centre matters more than how you build it.

Think of it as: The same machine is clean in one country and not clean in the next. Only the wire is different.

A company that sells almost nothing but AI chips is heavily exposed to AI: if that market booms it does very well, and if it stalls it suffers. A company that sells a hundred different things is barely exposed to any one of them. Neither is better — more exposure means bigger swings in both directions.

Think of it as: Standing in one spot in the rain versus under a tree. One gets everything the weather does; the other gets some of it.

A quick way to see how dear a company is, next to the money it makes each year. A high number means people expect that money to grow a lot. A low number means they doubt it, or they think something may go wrong. A company that makes no money has no number at all, so the number is left out.

Think of it as: A shop that makes £10 a year and sells for £200. That is twenty years of money.

Much like the P/E number, but two things are different. Money the company must pay back is counted as part of what you would take on. And the earnings used are what the work itself brings in, before tax and before costs that only sit on paper. That is more fair for a company that owns a lot of heavy plant, like a mine or a chip works. As before, a company with nothing to show has no number.

Think of it as: When you buy a shop, you also take on the money it still has to pay back.

What it would cost, on paper, to buy every share in a company. It is the simple way to say how big a company is. It matters because one small company can hold a step in the chain that the whole world needs. The ones buying from it can be a hundred times the size.

Think of it as: One egg has a price. Times all the eggs in the box, and that is the price of the box.

Rather than buy one company, you buy a fund that holds a basket of them and moves as a single line. If one of them fails, it hurts you less. The cost is that you also get less of the one you were really after. A proxy here means anything used to stand in for something you cannot buy on its own.

Think of it as: You back the whole team, not one member of it.

When a company is drawn next to the price of a part, both lines are set to start at 100. So the company line is a level, not a price. Turning it into another kind of money would mean nothing, and a money sign on it would be a lie. A basket is several companies rolled into one line.

Think of it as: Like a mark out of 100. The mark is not money.

For AI chips it is counted in FLOP/s — how many small sums it does each second. A petaFLOP/s is a thousand million million of them. BF16 names the way those numbers are written down. There are several ways to count, and two numbers counted in different ways cannot be set side by side. So one way is used here all the way through.

Think of it as: How many cars a road can carry each hour, not how fast one car goes.

This is the way a card has always turned a 3D world into a picture you can see. A score from that test is the best way to line up cards built ten years away from each other. 'Street price' means what people really pay in a shop, not the price the maker asks for. The 'representative product' is the card in the same place in the market at each point in time.

Think of it as: One race that everyone has to run, year after year.

A card costs $300 and a tonne of copper costs $9,000. Draw both on one picture and one line looks dead flat. Set both to 100 at the start and you see how far each has moved instead. So 150 means half as much again. The numbers up the side are no longer money. This word is used here for two other things as well: a stock market index, and the score on the materials page.

Think of it as: Line everyone up at the same start, then watch who moves the most.

This is the big idea here. A card can cost twice what the last one cost and still be better value, because it does three times the work. Share the price out over the work done, and you can ask whether computing is costing less, not just whether it is costing more. The two questions often give answers that point different ways. A benchmark point is one unit of work, set so a part from 2015 counts as 100.

Think of it as: Two cars, one at two times the price. If it can carry three times as much, it is the better buy.

On an even scale, 100 to 200 takes the same space as 1,000 to 1,100. But the first is twice as much and the second is hardly a move at all. A log scale gives equal space to equal times as much. So a line that grows 10% every year comes out straight. Use it when one picture has to hold both very small and very large numbers.

Think of it as: Like counting 1, 10, 100, 1000 with the same space between each.

$500 in 2015 bought more than $500 buys now. Set the two side by side and the past looks better value than it really was. To be fair, the old numbers are lifted up into money of today. CPI is the measure of how much everything has gone up, and it is what does the lifting.

Think of it as: A train ride cost less in 1990. So did everything else.

A score of 82 works like one country holding 82% of all the world can make. A material from just one place is easy to cut off. One order from a government, one fire or one dry year will do it. How much of it sits in the ground does not come into it.

Think of it as: One road into town. It works fine until the day it does not.