AI Literacy
The map is drawn for someone else
Every guide to the AI supply chain is written for a person deciding what to buy. Here is the same map, redrawn for a woman deciding what to depend on.
You get one question to ask before you build anything on an AI tool, and an honest picture of which layer of it you can actually own.
I have been reading a lot of these maps lately, the ones that explain what is really underneath the AI tools we all use now, and there is something at the bottom of nearly every one of them that I could not stop looking at once I had noticed it.
A disclaimer. This is not investment advice.
Sit with that for a second, because it tells you exactly who the map was drawn for. Somebody who wants to know where the money is. Somebody deciding what to buy. That is a perfectly good question and it is simply not the one I have, and I do not think it is the one you have either, and once I saw that, I could not read another one of these things the same way.
I am not trying to work out which chip company wins. I want to know what happens to my work when a layer I have never heard of gets tight.
So this is the same territory, walked as someone who has to use the thing.
Six layers, and what each one actually does to you
The machines that make the chips. One company on earth makes the lithography machines that print the most advanced chips, and that company is ASML in the Netherlands. Here is the number that stopped me. In the whole of 2025, across the entire planet, ASML recognised revenue on 48 EUV systems. Forty-eight. That is roughly four machines a month, for the world, and every advanced AI chip in existence traces back through one of them. The newest generation is reported to cost somewhere near 380 to 400 million US dollars each, though ASML does not publish its prices, so treat that as the industry's figure rather than the company's. When people talk about the AI boom being unstoppable, I think about forty-eight machines.
The chip designers. Nvidia is the name everyone knows, and its data centre business took 193.7 billion US dollars in the year to January 2026, up 68 per cent. The part that matters for you is quieter: Nvidia does not make anything. It designs, and someone else manufactures. So the company at the centre of the story is itself standing on a layer it does not own, which should tell you something about how far down this goes.
The factories. That someone else is mostly TSMC, in Taiwan, which turns out more than 17 million wafers a year. Look at TSMC's own capacity page and count the geography: six of its big 12-inch fabs are in Taiwan, and there is one in Arizona, one in Nanjing, one in Japan. I went looking for the German plant everybody wrote about, and TSMC's own page leaves it off the capacity list entirely. Which is the thing about reading the primary source instead of the coverage. The announcements and the concrete are different timelines, and only one of them makes chips.
The memory. This is where it stopped being abstract for me, and I want to show you the actual words rather than paraphrase them, because a company saying this in a legal filing is saying it carefully. SK Hynix, which IDC put first in high bandwidth memory with 56.4 per cent of that market in early 2026, told its investors that the memory AI needs is "more complex and resource-intensive than traditional DRAM products and requires greater wafer input". And then, under a heading it wrote itself, "Traditional DRAM: A Structural Supply Constraint", it explained that the industry has "allocated their limited cleanroom space and capital expenditure to the production of HBM".
Read that again in plain words. There is only so much factory. AI memory is eating the space where ordinary computer memory used to be made.
You can follow that all the way to a shop. Memory contract prices have been reported rising steeply through 2026, and PC makers have been reported passing that on. Which means a woman who has never opened an AI tool in her life, who just wants to replace a dying laptop, is quietly paying for all of this. She is inside the supply chain and nobody sent her a map.
The buildings. The companies buying all this are spending money that has stopped meaning anything as a number. Meta lifted its 2026 capital spending guidance to 125 to 145 billion US dollars and said the reason for the increase was, among other things, memory pricing. So the constraint I just described, the cleanroom space, shows up ten billion dollars later on a public earnings call. Every one of these layers is the same thing wearing a different name.
The electricity. And here is where it gets genuinely strange. The International Energy Agency put data centres at about 415 terawatt hours in 2024, around 1.5 per cent of the world's electricity, heading for about 945 terawatt hours by 2030. Roughly a doubling and then some, in six years. The number matters less than what Satya Nadella, who runs Microsoft, said out loud on a podcast in November 2025. His problem now is finding somewhere to plug the chips in. He described GPUs sitting in inventory that he cannot switch on for lack of power.
The richest company in the world, holding the most wanted object on earth, unable to use it because of the grid. That is where the bottleneck has moved.
Nobody actually knows what it costs
While I was reading, I kept trying to find out how much water an AI question uses, and I want to tell you honestly what I found, because the answer is a lesson in itself.
Google says a median Gemini text prompt uses about 0.26 millilitres, roughly five drops. Researchers at the University of California, Riverside, counting the water used to generate the electricity as well as the water used to cool the building, have been reported at something closer to ten to twenty-five millilitres.
That is a gap of up to a hundred times, between two sources that are both credible. They are measuring different things and neither is lying. There is simply no agreed way to draw the boundary yet. If someone quotes you one of those numbers without the other, they have taken a side in an argument they have not told you is happening.
I find that clarifying rather than depressing. We are all standing on infrastructure that nobody has finished measuring.
The seventh layer
So here is what I actually think, having walked the whole thing.
Six layers, and you own none of them. You are not buying a lithography machine. You cannot influence a cleanroom allocation in Icheon or the interconnection queue in Virginia. Every one of those layers is somebody else's decision, made for somebody else's reasons, arriving at your desk with no warning and no explanation. When Sam Altman wrote that "our GPUs are melting" and capped the free tier at three images a day, that was a chip constraint becoming a Tuesday, for millions of people who had never heard the word wafer.
The investor's map stops at six layers, and it is right to, because six is where the buying stops.
But there is a seventh, and it never appears on those maps for the simple reason that there is nothing there to invest in. It is your files. Your rules. The record of how you work, written down in words you chose, sitting on a machine in your house. Nobody draws it because nobody can sell it to you.
It is also the only layer of the entire chain that is yours.
I know how that sounds, so let me be concrete about what it buys and what it does not. It does not make you independent. I still call models that live in data centres I will never see, drinking power I cannot account for. Owning my files does not move a single wafer.
What it does is decide what a change six levels down actually costs me. When a price moves, or a model I liked is retired, or a rate limit appears overnight, I am not rebuilding myself. My voice, my values, the way I work, the things I have learned the hard way, none of that lived in the tool. It lived in files I own, in plain language, and I point them at whatever is standing up that week. I have run the same setup against a small model on my own laptop with the wifi off. It was slower and it was worse and it was completely, absolutely mine, and the point is that the choice existed at all.
That is the whole difference between an inconvenience and a catastrophe, and it is decided long before the thing goes wrong.
What to do with this
You do not need to become a semiconductor analyst. I am not one and I never will be. You need one question, and you can ask it before you adopt anything, before you build your business on a tool, before you let something become the place your thinking lives.
Which layer am I standing on, and what happens to me when it moves?
It moves. That is the only part of any of this I am certain about. Forty-eight machines a year, one island, three memory makers, a grid that takes half a decade to connect. Something in there is always tightening, and it never announces itself to you, it just arrives one morning as a price, or a limit, or a thing that used to work.
The women I want to reach have been handed maps their whole lives that were drawn for somebody else. This one is drawn for you, and the useful part of it is the layer at the top, the one you can put your hands on.
So put your hands on it today, and it costs you nothing but twenty minutes. Open a plain text file, somewhere you will find it again, and write down the things you would have to rebuild from scratch if your favourite tool vanished on Monday morning. How you like to be spoken to. The rules you have learned the hard way. The words you never use. What good looks like to you. It will feel almost silly while you are doing it, and then one day a price moves or a model you liked is retired, and you will point that file at whatever is standing up that week and carry on with your day.
That file is the seventh layer. It is the only part of the map you own, and it starts as an empty page 🤍
This is educational writing about how AI infrastructure works. It is not investment advice, and it is not a recommendation about any company. Where a figure is reported rather than filed, I have said so and linked the source, so you can go and check me.