Skip to content
Hi, Bot

› Free · for curious people of any age · last checked November 18, 2026

How deep does AI go?

Start with one chatbot answer and keep asking what's underneath. You'll pass through math and machines, chips and factories, mines, rocks and laws, and come out the bottom among the stars.

82 stops on 10 levels, seen 6 ways: how it works, ideas and history, money, power and rules, people, and the planet. Every stop says where its facts come from, 151 sources in all.

It's made in the spirit of Claude Shannon, who invented information theory, built a maze-solving robot mouse, and never let the edge of a subject stop a good question.

› The map

Ten levels down, six ways across.

Each row is a level, from your question at the top to the stars at the bottom. Each column is a way of looking. Tap a stop to light up what it connects to, then keep going down.

0 The question

How it works
Ideas and history
People

1 The model

How it works
Ideas and history
Money
Power and rules
People

2 The math

How it works
Ideas and history

3 The computer

How it works
Ideas and history
Money
Planet

4 The building

How it works
People
Planet

5 The chip

How it works
Ideas and history
Money
Power and rules

6 The factory

How it works
Money
Power and rules
People
Planet

7 The ground

How it works
Money
Power and rules
People
Planet

8 The rocks

How it works
Ideas and history
Planet

9 The stars

How it works
Ideas and history

› 30 days

Follow a trail.

Each trail walks along real connections on the map. Take a stop every few days: read it, open one of its sources, try what it links to, and write down one question of your own.

› 30 years

Questions nobody has answered yet.

These are real open problems. People are spending careers on them, and some may outlast everyone working on them today. Each one starts from a stop on the map.

› Better questions

Ask like Claude Shannon.

In a 1952 talk at Bell Labs, Shannon described the moves he used on hard problems. Here they are as questions you can point at any stop, plus two of our own. On the map, open “Ask your own question” to aim them at the stop you're on.

  1. Shannon's move

    Simplify

    What is the smallest, simplest version of it I could explain on one page?

  2. Shannon's move

    Find a cousin

    What problem in a completely different field looks like it, and how was it solved?

  3. Shannon's move

    Say it another way

    How would I describe it without using any of its jargon?

  4. Shannon's move

    Zoom out

    Is it one example of a bigger pattern? What else fits that pattern?

  5. Shannon's move

    Break it apart

    What are the parts of it, and which part is the hardest?

  6. Shannon's move

    Work backward

    Start at the end: what had to be true just before it, and just before that?

  7. Our addition

    Ask who decides

    Who gets to decide how it is made, sold or used, and who doesn't get a say?

  8. Our addition

    Ask how we know

    What evidence would change my mind about it, and where would I find it?

Source: Claude Shannon, “Creative Thinking”, a talk at Bell Labs, March 20, 1952 (transcript) (opens in a new tab)

› 30 months

Every stop, level by level.

Everything on the map, written out with its questions and sources. Pick a level and work your way down.

Level 0

The question

You type. It answers.

Your question · How it works

You ask a chatbot a question.

You type a question and an answer appears a second or two later. It can feel like magic. Underneath that one answer sit layers of math, machines, factories, mines and laws, and every layer has a story.

This page follows that answer down. Each stop is a level deeper, or a different way of looking at the same level: how it works, the ideas behind it, the money, the power and rules, the people, and the planet. Every stop lists where its facts come from, so you can check them yourself.

Questions to ask

  • What actually happens between pressing Enter and seeing the first word?
  • Where, physically, is the computer that answered you?

Kids and chatbots · People

Lots of kids already use chatbots for school.

In a 2024 survey, about a quarter of US teens said they had used ChatGPT for schoolwork, twice the share from a year before.

The same survey found teens judge uses differently: just over half said using it to research a new topic is fine, but only 18% said the same about writing an essay. Knowing what sits underneath an answer, where it came from and why it can be wrong, is what turns a shortcut into a tool.

Questions to ask

  • When is using a chatbot for homework learning, and when is it skipping the learning?
  • How would you check an answer a chatbot gave you?

ELIZA, 1966 · Ideas and history

People have been chatting with programs since 1966.

In 1966 Joseph Weizenbaum at MIT published ELIZA, a program that chatted by matching patterns in your words and turning them back into questions. People knew it was a program and still felt understood.

ELIZA had no knowledge of the world at all, only rules like: if someone types "I am sad", ask "How long have you been sad?" Weizenbaum was surprised and worried by how readily people confided in it. The habit of reading understanding into fluent text is now called the ELIZA effect, and it matters more with today's far more fluent systems.

Questions to ask

  • Why do fluent words make us feel understood?
  • What would convince you a machine understands you, and what wouldn't?
Back to the map ↑

Level 1

The model

What a language model does with your words.

Tokens · How it works

Your words get chopped into tokens.

A model doesn't read letters or whole words. It reads tokens: common chunks of text, each with an ID number. A short common word is usually one token; a long rare word may split into several.

Most chatbots build their chunk list with byte-pair encoding. Start from single bytes, then keep gluing together the pair that shows up most often in a huge pile of text, until the list holds tens of thousands of chunks, or a couple of hundred thousand in some newer models. The trick began as a 1994 data-compression method and was adapted for language models around 2015.

Questions to ask

  • Why not just give every word its own number?
  • What happens to a word the tokenizer has never seen?

Embeddings · How it works

Each token becomes a long list of numbers.

Each token ID is swapped for a vector, a list of hundreds or thousands of numbers. Tokens used in similar ways end up with similar lists, so "cat" lands near "kitten".

Nobody types these numbers in. They start random and get nudged during training until the geometry carries meaning. In 2013 the word2vec project showed directions in this space can line up with ideas (king minus man plus woman lands near queen), though such neat results work less often than the famous example suggests. The same geometry can also soak up stereotypes from the text it learned from.

Questions to ask

  • If meaning is a direction, what would "halfway between two words" mean?
  • How could a list of numbers carry a bias?

Attention · How it works

Every word looks back at the words before it.

In a transformer, each token asks which earlier words matter to it and mixes in information from them. In "The trophy didn't fit in the suitcase because it was too big", attention is how "it" can lean toward "trophy".

The 2017 paper "Attention Is All You Need" built a whole model, the transformer, from layers of attention stacked dozens deep, each layer with many heads looking for different patterns. Standard attention compares every token with every other token, so doubling the length of the text roughly quadruples that work. That's one reason very long conversations are expensive.

Questions to ask

  • Why might looking at all the words at once beat reading one word at a time?
  • What gets harder when a conversation gets very long?

Next-token guessing · How it works

It guesses the next word, one at a time.

At the end of all those layers, the model gives a probability to every token it knows, say "Paris" 92% and "Lyon" 1%. One token is picked, added to the text, and the whole model runs again for the next one.

A function called softmax turns raw scores into probabilities that add up to 100%. A setting called temperature controls how adventurous the pick is: low temperature nearly always takes the top choice, high temperature takes more chances. A long answer is hundreds of these picks in a row, each one leaning on everything before it, including the model's own earlier words.

Questions to ask

  • If it only ever picks the next word, how can it seem to plan a whole answer?
  • Why might the same question get two different answers?

Training · How it works

It learned by guessing, trillions of times.

Before it ever met you, the model was shown enormous amounts of text with the next token hidden. It guessed, got scored, and every one of its billions of internal numbers was nudged to make the right answer a little more likely.

Those internal numbers are called parameters or weights, and big models have hundreds of billions of them. Repeat the guess-and-nudge loop over trillions of tokens and patterns of grammar, facts, style and reasoning start to appear, because they help predict the next word. Nobody writes those patterns in, which is why even the builders have to test a model to find out what it can do.

Questions to ask

  • Why would predicting the next word teach a model facts or arithmetic?
  • What might a model learn that its builders never intended?

Training data · How it works

It read a huge slice of the internet.

Most of that text came from web pages copied by automatic crawlers, plus books, code, papers and more. What is in the pile, and what is missing from it, shapes what the model knows and how it talks.

One public archive, Common Crawl, has been copying billions of web pages since 2008. Builders filter it heavily for quality and remove duplicates, but the web over-represents some languages, places and points of view. A 2021 study of one big cleaned-up dataset found its filters removed text written in some dialects, and text about some groups of people, more often than others.

30-year question: Who should decide what goes into the data that trains the models everyone uses?

Questions to ask

  • Who is missing from the internet, and so from the model?
  • Should people be asked before their writing is used to train a model?

Fine-tuning · How it works

People taught it to act like a helpful assistant.

A model fresh from training only continues text. To make it answer questions, follow instructions and refuse harmful requests, builders train it more, on examples and on people's ratings of its answers.

A common recipe is reinforcement learning from human feedback: people compare two answers, a second model learns to predict which one they would prefer, and the chatbot is tuned to score well with that judge. Some builders also train against a written list of principles. Whose preferences and principles get used is a choice, and it shapes the chatbot's personality. Tuning for approval can also teach a model to tell people what they want to hear.

Made-up answers · How it works

Sometimes it makes things up, confidently.

A chatbot can invent a quote, a book or a court case and state it calmly. It isn't lying the way a person lies. It is producing text that fits the pattern of a good answer, whether or not the answer is true.

Researchers call these hallucinations. A 2025 paper argues that training and testing often reward a confident guess over saying "I don't know", the way a multiple-choice test rewards guessing. Letting a model look things up and cite sources helps but doesn't remove the problem. Checking a claim against a source you trust is still the reader's job.

30-year question: Can we build a model that reliably knows when it doesn't know?

Questions to ask

  • Why is "I don't know" hard for a model trained to always produce an answer?
  • Which kinds of questions are most likely to get a made-up answer?

Looking inside models · How it works

Nobody fully knows how it works inside.

We know the recipe exactly: every number and every multiplication. What those billions of numbers add up to, which ones store a fact and which ones do the arithmetic, is still mostly a mystery.

Interpretability researchers open models up the way biologists use a microscope. They have found internal features for ideas as specific as the Golden Gate Bridge, and traced small circuits that do jobs like copying a name from earlier in the text. It is a young science: the tools work on pieces of a model, not yet on a whole one at once.

30-year question: Can we read what a model is doing inside well enough to check that it is safe and honest?

Dartmouth, 1956 · Ideas and history

The name "artificial intelligence" comes from a 1955 proposal.

In 1955 four researchers, John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, asked for money to run a summer study of "artificial intelligence" at Dartmouth College. The name first appeared in that proposal, and the 1956 workshop helped make it stick.

They expected a small, well-chosen group could make real progress in a single summer. It took decades longer than they hoped, through booms and "AI winters" when funding dried up. Their proposal already lists questions that are still open: how machines could use language, form abstractions and improve themselves.

Questions to ask

  • Why do you think the founders expected AI to be easier than it was?
  • Which of their 1955 questions has AI answered, and which are still open?

Understanding · Ideas and history

Does it understand, or just predict?

People disagree, sharply. Some say a system that only predicts text can't mean anything by it. Others say predicting well enough requires something like a model of the world. Both sides include serious researchers.

In 1950 Alan Turing proposed swapping "can machines think?" for a test: can a machine hold a conversation you can't tell apart from a person's? In 1980 the philosopher John Searle replied with his Chinese Room argument: following rules for symbols, he said, is not the same as understanding them. A 2021 paper called large language models "stochastic parrots". Researchers are still testing what these models represent inside.

30-year question: What would count as evidence that a machine understands, and could we ever be sure?

Questions to ask

  • How do you know your friends understand you?
  • Could a system understand language without ever seeing, touching or doing anything?

The price of an answer · Money

Answers keep getting cheaper, and people keep asking more.

Companies sell the use of their models by the token. For the same quality of answer, prices have dropped fast: one 2025 analysis found falls of between 9 and 900 times per year, depending on the benchmark and how good an answer you need.

Cheaper things get used more. In 1865 the economist William Stanley Jevons noticed that more efficient steam engines made Britain burn more coal, not less, because coal power became worth using everywhere. Whether AI follows Jevons, with cheaper answers leading to so many more answers that total chip and electricity demand keeps climbing, is a big question for energy planners.

30-year question: Will cheaper AI answers mean the world uses less energy on AI, or more?

Questions to ask

  • Who pays for a "free" chatbot answer?
  • If each answer uses less energy, how can total energy use still go up?

AI laws · Power and rules

Governments are writing rules for AI, and they don't agree.

The European Union passed the first broad AI law in 2024. It sorts AI uses by risk: a few are banned, some must pass strict checks, and makers of the biggest general-purpose models get extra duties. Other countries are taking different paths.

China has had binding rules for recommendation algorithms, deepfakes and generative AI since 2022 and 2023, and since September 2025 AI-made content there must be labeled. The United States has no single national AI law: its rules come from executive orders, agencies and states, and a December 2025 presidential order set out to challenge state AI laws. In 2024 the Council of Europe opened the first legally binding international AI treaty for signature, and in 2026 the EU pushed its strictest rules, for high-risk uses, back to late 2027 and 2028. These rules change fast, so check the date on anything you read.

Data workers · People

Thousands of people label data and rate answers.

Behind "the AI learned" are people: workers who tag images, write example answers, rank chatbot replies, and read violent or disturbing text so it can be filtered out before users see it.

Much of this work goes through outsourcing firms and online platforms, often in countries where wages are lower. Reporters found that workers in Kenya, paid about 1.32 to 2 dollars an hour, labeled graphic text to make a well-known chatbot safer, and some said the work harmed their mental health. A 2021 guide from researchers and companies had already set out fair pay and support for this kind of work.

Questions to ask

  • Should the people who teach an AI be credited, the way authors are?
  • What would fair work look like for someone who reviews disturbing content all day?

The token tax · People

Some languages cost more tokens than others.

Tokenizers are mostly built from English-heavy text, so the same sentence can take many more tokens in another language. More tokens means a higher price, a slower answer and less room in the model's working memory.

A 2023 study compared translations of the same text and found that, with the same tokenizer, some languages needed up to 15 times as many tokens as others. Speakers of those languages pay more for the same service and fit less into it. Bigger, more balanced token lists can narrow the gap, but they don't close it on their own.

Questions to ask

  • Who decides which languages a tokenizer is good at?
  • How could you make a tokenizer fair to every language?
Back to the map ↑

Level 2

The math

The ideas every model is built from.

Information · How it works

Information is surprise you can count.

In 1948 Claude Shannon showed how to measure information in bits. A fair coin flip carries 1 bit. A message you could have guessed carries almost none. The less predictable something is, the more information it holds.

Shannon's entropy is the average surprise of a source, and it sets hard limits: how small a file can be compressed, and how fast data can cross a noisy wire without errors. Much of modern communication, from phone calls to zip files to Wi-Fi, builds on ideas from that paper. A language model is, in a precise sense, a machine for being less surprised by the next token.

Questions to ask

  • Which carries more information: a weather report from a desert, or from a rainforest?
  • Why can't you zip a zip file and make it much smaller?

Shannon's guessing game · Ideas and history

In 1951, people played "guess the next letter" for science.

Shannon asked people to guess a text one letter at a time and counted their misses. From that he estimated that English carries roughly 1 bit per letter, far below the roughly 4.7 bits it would carry if letters were random.

His 1948 paper had already generated text one statistical step at a time, by hand: he picked each next word using a real book and got strings that sounded eerily like English. The mathematician Andrey Markov had counted letter patterns in a Russian poem back in 1913. A chatbot plays the same guessing game with a far better memory for what came before. Its training goal is, quite literally, to lower the surprise Shannon measured.

Loss · How it works

Training means shrinking the surprise.

The score a model gets during training is called its loss. For a chatbot it is usually cross-entropy: how surprised the model was by the real next token. Lower is better.

If the model gave the right token a 50% chance, the surprise is 1 bit; at 25% it is 2 bits; at 100% it is zero. Average that over trillions of tokens and you get one number the builders watch fall during training. It is Shannon's measure of information, put to work as a teacher.

Questions to ask

  • Why punish a confident wrong guess more than an unsure one?
  • Could a model get a very low loss and still be bad at what people want?

Matrix math · How it works

It's mostly multiply and add, trillions of times.

Almost all the work inside a model is multiplying grids of numbers called matrices: multiply pairs of numbers, add up the results, repeat. One answer can take trillions of these small steps.

Linear algebra, the math of vectors and matrices, is the language every layer is written in. Each layer turns a token's list of numbers into a new list by multiplying it by a grid of learned weights. Because each output can be worked out separately, the job splits neatly across thousands of small processors, which is exactly what graphics chips were built to do.

Questions to ask

  • Why does it matter that each multiplication can happen at the same time as the others?
  • What is a matrix, in your own words?

Gradient descent · How it works

It learns by walking downhill.

Picture the loss as a landscape of hills and valleys. Training feels for the downhill direction and takes a small step. Calculus tells it which way is down, for billions of weights at once.

The trick that makes this practical is backpropagation, popularized in 1986: the chain rule from calculus, run backward through the layers to share out the blame for a wrong answer. The step size, called the learning rate, is a balancing act. Too small and training crawls; too big and it overshoots the valley.

Questions to ask

  • How could you find the bottom of a valley in thick fog?
  • What could go wrong if the landscape has lots of small dips?

Rough numbers · How it works

AI chips do math with rough numbers, on purpose.

Scientific computers usually store each number in 64 or 32 bits. AI chips increasingly use 16, 8 or even 4 bits per number. Rougher numbers mean less memory to move and more math per second.

It works because neural networks shrug off small errors: a weight of 0.4375 instead of 0.4372 rarely changes the answer. Formats like bfloat16 and FP8 keep a wide range of sizes while dropping fine detail. Halving the bits roughly doubles how many numbers fit through the memory pipe, which is often the real bottleneck.

Questions to ask

  • When is a rough answer good enough, and when isn't it?
  • Why could moving a number cost more energy than doing math with it?

Scaling laws · How it works

Bigger kept working, in a predictable way.

In 2020 researchers found that a model's loss falls along a smooth, predictable curve as you add parameters, data and computing. That curve turned "build a bigger model" into a plan that companies could bet billions on.

A 2022 study nicknamed Chinchilla showed many models were too big for the amount of data they saw: for a fixed budget, model size and data should grow together. Nobody has a full theory of why these curves hold. Whether they keep holding, and whether lower loss keeps turning into more useful abilities, is one of the most expensive open questions in the world.

30-year question: Why do bigger models improve so predictably, and when will it stop?

Questions to ask

  • If a curve has held for years, how confident should you be that it continues?
  • What could run out first: data, chips, electricity or money?

Logic in switches · Ideas and history

A 21-year-old showed that switches can do logic.

In 1937 Claude Shannon, a 21-year-old student at MIT, noticed that circuits of on-off switches follow the same rules as George Boole's 1854 algebra of true and false. That meant any logical rule could be built out of switches.

His master's thesis turned circuit design from trial and error into math: AND, OR and NOT, combined, can add numbers, compare them and make decisions. Today's chips use transistors instead of clicking relays, but they are still Shannon's switches, billions of them. The psychologist Howard Gardner called it possibly the most important master's thesis of the century.

Questions to ask

  • How could on-off switches add 2 and 3?
  • What else in your home is a switch making a decision?

The perceptron · Ideas and history

A machine learned from examples in 1958.

Frank Rosenblatt's perceptron was a simple artificial neuron: add up weighted inputs and fire if the total passes a threshold. It learned its weights from examples. The US Navy, which paid for the work, said it expected such machines to one day walk, talk and see, and newspapers ran the story.

It built on a 1943 idea by Warren McCulloch and Walter Pitts, who modeled neurons as logic switches. In 1969 Marvin Minsky and Seymour Papert proved that a single layer of perceptrons can't learn some simple patterns, and interest faded for years. Stacking many layers and training them with backpropagation is how today's networks got past that limit.

Back to the map ↑

Level 3

The computer

What computing is, and what it costs.

Computation · How it works

A computer is a machine that follows rules for symbols.

In 1936 Alan Turing imagined the simplest possible computer: a machine reading and writing symbols on a tape, one step at a time, following a table of rules. He showed that one such machine could imitate any other.

That universal machine is why the same chip can run a game, a spreadsheet or a chatbot: they are all just different rule tables. When AI people say "compute" they mean how many arithmetic steps you can afford, roughly chips times speed times time. Turing also proved there are questions no computer can ever answer, however much compute it has.

Questions to ask

  • If any computer can imitate any other, why build special AI chips?
  • What kind of question could no machine ever answer?

FLOP counts · How it works

AI is measured in floating-point operations.

A FLOP is one arithmetic step on a number with a decimal point. The largest training runs now use more than 10²⁵ of them, a 1 followed by 25 zeros. At one per second, that would take far longer than the age of the universe.

The computing used to train the biggest models has grown roughly four to five times every year since 2010, according to Epoch AI. The number matters outside the lab too: the EU AI Act presumes that a general-purpose model trained with more than 10²⁵ FLOP has "high-impact capabilities", which puts it in the systemic-risk group with extra duties. A count of multiplications has become a legal line.

Questions to ask

  • Is the amount of computing a good way to decide which AI needs extra rules?
  • What could a model trained with less compute still do well?

GPUs · How it works

Chips built for video games ended up running AI.

A graphics processing unit has thousands of simple cores that each do small sums at the same time, originally to color millions of pixels for games. That happens to be exactly the shape of neural-network math.

In 2012 a network called AlexNet, trained on two gaming GPUs, won an image-recognition contest by a wide margin and set off the deep-learning boom. Since then chipmakers have added special units just for matrix math and low-precision numbers. A CPU is a few brilliant cores; a GPU is an army of simple ones.

Questions to ask

  • What other jobs are like painting a million pixels at once?
  • When is one brilliant worker better than an army of simple ones?

The memory wall · How it works

Moving numbers is slower than doing math with them.

Chips can multiply far faster than memory can feed them numbers. When a chatbot writes an answer, the bottleneck is usually not the math. It is hauling the model's weights out of memory for every single token.

Engineers named this the memory wall in 1995, and it has grown taller. One 2024 study found that chips' peak math speed rose about 3 times every two years over two decades, while the bandwidth of the main memory chips rose about 1.6 times. AI chips fight back with high-bandwidth memory: stacks of memory chips placed right beside the processor, joined by thousands of tiny wires.

Questions to ask

  • Why might a chip spend most of its time waiting?
  • How is this like a fast cook with a slow pantry?

Chip-to-chip networks · How it works

Thousands of chips must work as one.

No single chip can hold a big model or train it fast enough, so thousands to hundreds of thousands work together, constantly swapping results over very fast links and fiber-optic cables.

Training splits the work: different chips hold different layers, slices of layers or batches of data, then sync up. The network between them can decide how fast the whole machine runs. At this scale something is always breaking: Meta reported 466 interruptions in 54 days while training one model on up to 16,000 GPUs, most of the unplanned ones caused by hardware faults.

Questions to ask

  • Why is it hard to make 10,000 workers act like one?
  • If a few chips fail every day, how would you keep the job going?

Inference · How it works

Answering you is its own huge job.

Training happens once; answering, called inference, happens billions of times a day. Your question travels to a data center, joins a batch of other people's questions, and the model writes the reply token by token.

Batching is the trick: the model's weights are read from memory once per step and shared across many users, so each answer costs less. Long answers, "reasoning" models that work through steps before replying, and pictures and video all take more steps. A 2022 paper by Google researchers found that inference made up about three-fifths of the company's machine-learning energy use in the years they studied.

Landauer's limit · How it works

Physics sets a minimum energy for forgetting.

In 1961 Rolf Landauer at IBM showed that erasing one bit of information must release a tiny, fixed amount of heat. Information isn't just an idea. It is physical.

At room temperature the minimum is about 3 × 10⁻²¹ joules per bit, written kT ln 2, and experiments confirmed it in 2012. One step of arithmetic in a real chip flips many transistors and uses far more energy than that limit, so physics leaves plenty of room to improve. Reversible computing, which avoids erasing bits at all, is one long-shot idea for getting closer.

30-year question: Could computers ever run near the physical minimum energy, and what would it take?

Theseus the mouse · Ideas and history

Shannon built a robot mouse that learned a maze.

Around 1950 Claude Shannon built Theseus, a mouse moved by magnets that found its way through a maze of movable walls and then remembered the route. Its brain was a bank of telephone relays under the floor.

After exploring once by trial and error, Theseus could run straight to the goal. Move the walls and it hunted again, updating its memory. It was one of the first machines to learn from experience, built from the same kind of switches Shannon had turned into logic in 1937.

Questions to ask

  • What is the difference between finding the way and learning the way?
  • What would you build if you followed your curiosity the way Shannon did?

The 20-watt brain · Ideas and history

Your brain runs on about 20 watts.

A human brain uses roughly 20 watts, about as much as a dim light bulb, and learns language from a few years of listening and talking. Meta's team reported that the power use of one 16,000-chip AI training cluster could swing by tens of megawatts in an instant.

Artificial neurons were inspired by real ones, but they are far simpler: a real neuron is a living cell with thousands of connections that change with chemistry and time. Brains and chips work on very different principles, so comparisons are tricky. Still, the gap in energy and data is a clue that there is much left to learn about learning.

30-year question: How does the brain learn so much from so little data and energy?

Questions to ask

  • What can a five-year-old do easily that AI finds hard?
  • What could chip designers borrow from brains?

The cost of training · Money

Training a top model costs a fortune, and the bill keeps growing.

The chips and electricity for one frontier training run cost tens of millions of dollars, and by one careful estimate that cost grew about 2.4 times a year from 2016 to 2024. Counting staff and experiments, developing a top model has passed 100 million dollars.

Chips are the biggest single cost, and research staff can add a third to half of the total. Because only a handful of companies and governments can pay that much, cost shapes who gets to build the most capable models, and pushes everyone else toward smaller models or toward sharing open ones.

Questions to ask

  • If only a few companies can afford the biggest models, who decides what those models are for?
  • What could make training cheaper?

Energy per answer · Planet

One answer uses a little energy. Billions of answers add up.

In 2025 Google reported that the median text prompt to its Gemini assistant used about 0.24 watt-hours of electricity, less energy than watching TV for nine seconds, plus about five drops of water for cooling.

Independent estimates for other chatbots land in a similar range for short text answers, but long "reasoning" answers, pictures and especially video can use many times more. Not every company publishes these numbers, and they don't all measure them the same way. Multiply by billions of prompts a day, add training and everything else, and it adds up.

Questions to ask

  • Is a tiny cost per use still small when billions of people use something every day?
  • Why might two companies measure energy use differently?
Back to the map ↑

Level 4

The building

Data centers: racks, power, water and neighbors.

Data centers · How it works

The answer comes from a warehouse full of computers.

Your question travels to a data center: usually a windowless building filled with rows of racks, each rack packed with servers, each server holding a handful of AI chips. A typical AI-focused data center uses as much electricity as 100,000 homes.

Inside, the work is mostly three jobs: getting electricity in, getting heat out and moving data around. The newest AI racks can draw over 100 kilowatts each, many times an older rack, which is pushing builders from fans to liquid piped right to the chips. The International Energy Agency says the largest data centers now being built will use 20 times as much power as today's typical AI site.

Questions to ask

  • Why would a company build a data center somewhere cold, or somewhere dry?
  • What does "the cloud" actually weigh?

Undersea cables · How it works

Your question may cross an ocean through a cable.

Data between continents mostly travels as light inside glass fibers, in cables laid along the sea floor. Several hundred of these cables carry most of the data that crosses oceans.

Each fiber is about as thin as a hair, and its core is glass mixed with a little germanium so it guides light. Special ships lay the cables and repair them when anchors or undersea landslides cut them. Because so much depends on so few cables, who builds, owns and protects them is now a question for navies and diplomats as well as engineers.

Questions to ask

  • What happens to the internet when a cable is cut?
  • Who should protect cables that run through international waters?

Electricity · Planet

AI is changing how much electricity the world needs.

Data centers used about 1.5% of the world's electricity in 2024, according to the International Energy Agency, which expects that to more than double by 2030, with AI the biggest reason.

The growth is uneven. In the United States, data centers used about 4.4% of all electricity in 2023 and could reach between 6.7% and 12% by 2028, packed into a few regions. Utilities are building new lines and power plants, tech companies are signing deals for nuclear, gas, solar and wind power, and some local grids are struggling to keep up. Forecasts disagree a lot, so check who made one and why.

30-year question: How much electricity will AI need by 2030, and where will it come from?

Questions to ask

  • Who should pay for new power lines: the data center, or everyone's electric bill?
  • Why do forecasts of AI's electricity use disagree so much?

Cooling and water · Planet

Every watt a chip uses turns into heat.

All the electricity a data center uses ends up as heat that has to go somewhere. Many data centers cool themselves by evaporating water, the way sweat cools you.

How much water depends on design and climate. Evaporative cooling saves electricity but uses water; closed liquid loops and dry coolers use little water but more power. There is also hidden water at the power plants that make the electricity. Researchers argue that AI's water use should be measured, reported and planned for, especially in dry places.

Questions to ask

  • Is it better for a data center to use more water or more electricity?
  • Where does a data center's heat go, and could it warm homes instead?

Data center neighbors · People

Every data center is somebody's neighbor.

Northern Virginia has one of the largest clusters of data centers in the world. The buildings bring tax money and some jobs, along with noise, new power lines and arguments about who pays for the grid.

A 2024 study for Virginia's legislature found the industry brings in local tax money, and also that a state sales-tax break saved data center companies 928 million dollars in a single year. It warned that their growing demand could require new power plants and lines that households may help pay for. Similar debates are happening near new sites around the world, and communities increasingly ask for a say in where data centers go and how they use power and water.

Questions to ask

  • Would you want a data center in your town? What would you ask for in return?
  • Who should get a vote on where one is built?
Back to the map ↑

Level 5

The chip

Billions of switches on a sliver of silicon.

Transistors · How it works

A chip is billions of tiny switches.

A transistor is a switch with no moving parts: a small voltage on its gate lets current flow or stops it. The biggest AI chips hold more than 100 billion of them, with features smaller than most viruses.

The first transistor was built at Bell Labs in December 1947. The kind in nearly every chip today, the MOS transistor, was invented there in 1959: a layer of metal (now other materials), a thin glassy layer of silicon dioxide and the silicon itself. Wire them together in Shannon's logic patterns and you get adders, memories and processors.

Questions to ask

  • Why build a computer from switches instead of dials?
  • How small can a switch get before it stops working reliably?

Moore's law · Ideas and history

For 50 years, chips doubled their switches about every two years.

In 1965 Gordon Moore noticed that the number of parts on a chip was doubling every year; in 1975 he revised that to every two years. Engineers made the guess come true for half a century.

Moore's law isn't a law of nature. It is a forecast an industry used as a schedule. Around 2005 a related trend broke: chips could no longer simply run faster without overheating. Designers added more cores and special-purpose units instead, and now they also stack chips and pack them side by side. Each new generation costs more to reach than the last.

30-year question: What comes after shrinking silicon transistors?

Questions to ask

  • What is the difference between a law of nature and a forecast that people work to make true?
  • If shrinking stops, how else could chips get better?

Bell Labs, 1948 · Ideas and history

The transistor and information theory came from one lab, weeks apart.

Bell Telephone Laboratories announced the transistor at the end of June 1948. The next month its journal began publishing Shannon's theory of information. The hardware and the math of the digital age went public weeks apart, from one company.

Bell Labs, the research arm of the US telephone monopoly, could afford long bets: physicists, mathematicians and engineers shared hallways and lunch tables. It went on to make the first practical silicon solar cell, Unix and the C programming language, and work behind several Nobel Prizes. Historians still argue about which mix of money, freedom and people made it work, and whether it could happen again.

Chip design · How it works

Chips are designed in software, mostly by companies that don't make them.

No one draws billions of transistors by hand. Designers describe a chip in code, and design software from a few specialist companies turns that into a layout, checks it and simulates it before anything is built.

Most AI chip companies are "fabless": they design chips and pay a foundry to make them. Many license ready-made building blocks, such as processor cores from Arm, or use the free and open RISC-V instruction set. Because a handful of companies dominate chip-design software, cutting off access to it has become a tool of trade policy too.

Questions to ask

  • Why would a chip company choose not to own a factory?
  • What is the difference between an open standard and an open market?

The best AI chips · How it works

The best AI chip isn't just the fastest one.

AI chips are compared on math speed with rough numbers, how much fast memory they carry, how well thousands of them link up, how much work they do per watt and what software runs on them. Price, and whether you can actually get them, count too.

By most analysts' estimates, Nvidia's data-center GPUs are the most widely used as of 2026, partly thanks to CUDA, programming software it has built up since 2007 that many AI tools assume. Google's TPUs, used inside its own data centers since 2015, AMD's Instinct GPUs and chips designed in-house by cloud companies compete hard. Public benchmarks such as MLPerf let anyone compare systems on the same tasks.

Questions to ask

  • Why might software matter as much as the chip itself?
  • Would you pick the fastest chip, the cheapest, or the one you can get this year? Why?

Packaging · How it works

Today's best chips are several chips joined together.

A top AI "chip" is really a package: one or two big processor dies plus stacks of memory, set side by side on a slab of silicon wiring. How you join the pieces now matters as much as how you make them.

Giant dies are risky, because one speck of dust can ruin the whole thing, so smaller "chiplets" give more good parts per wafer. Advanced packaging, such as TSMC's CoWoS, links the pieces with wires far finer than any circuit board's. In 2023 and 2024 this packaging step, not the chips themselves, was one of the tightest bottlenecks on AI chip supply.

Questions to ask

  • Why might two medium chips beat one giant one?
  • Where else does putting parts closer together make things faster?

Supply and demand · Money

When everyone wants the same chip, its price soars.

After ChatGPT launched in late 2022, demand for AI chips jumped far faster than factories could grow. Buyers waited months, and prices and profits for chipmakers climbed.

Nvidia's data-center revenue rose from about 15 billion dollars in its 2023 fiscal year to about 115 billion two years later, and kept climbing. Supply grows slowly because every step, from fabs to packaging lines to memory, takes years and billions of dollars to expand. High prices pull in competitors and new factories, which is how markets usually respond. Whether today's demand will last is a bet investors argue about every day.

Questions to ask

  • Why can't factories just make more chips when demand jumps?
  • What happens to prices when the new factories finally open?

Export controls · Power and rules

Some countries ban selling top AI chips to rivals.

Since October 2022 the United States has restricted exports of the most advanced AI chips, and the machines that make them, to China. The Netherlands and Japan added limits on chipmaking tools. China has answered with limits on minerals.

The rules aim to slow a rival's military and AI progress. Critics argue they also push that rival to build its own chip industry faster. The rules have been tightened, loosened and rewritten several times; in January 2026, for example, the US began reviewing license requests for some previous-generation AI chips case by case instead of refusing them by default. Chips have also been smuggled around them. Behind it all sits an older agreement, the 1996 Wassenaar Arrangement, through which 42 countries coordinate controls on goods with both civilian and military uses.

Chip trade · Power and rules

A chip crosses many borders before it's finished.

One AI chip might be designed in the United States, built with machines from the Netherlands, Japan and the US, made and packaged in Taiwan, and put into a server in Mexico, before it ships to a data center somewhere else.

No country makes the whole chain alone. A 1996 World Trade Organization deal, the Information Technology Agreement, removed tariffs on chips and computers among its members to keep that chain cheap. Recent years have brought new tariffs, subsidies and restrictions as governments try to bring more of the chain home; in January 2026, for example, the US put a 25% tariff on certain advanced chips. Each choice trades cost against security.

Back to the map ↑

Level 6

The factory

How chips are printed, and who can print them.

Silicon wafers · How it works

Chips start as slices of one giant crystal.

Ultra-pure silicon is melted, and a small seed crystal is slowly pulled up out of the melt, growing a single perfect crystal taller than a child. It is sliced into wafers 300 millimeters across and polished flatter than almost anything else people make.

The method is named for Jan Czochralski, who discovered it in 1916, reportedly after dipping his pen into molten tin instead of his inkpot. Chip silicon must be about 99.999999999% pure, which engineers call eleven nines. Dozens to thousands of chips are made on each wafer at once, then cut apart.

Questions to ask

  • Why does a chip need one perfect crystal instead of many small ones?
  • What does eleven nines pure actually mean in atoms?

Printing with light · How it works

Chips are printed with light, layer by layer.

The wafer is coated with a light-sensitive chemical, light shines through a stencil called a mask, and the pattern is etched or filled in. Repeat that dozens of times, layer on layer, and you get billions of transistors and the wiring between them.

A leading-edge chip goes through hundreds of steps over roughly three to four months in the factory, all inside cleanrooms whose air is far cleaner than an operating room's, because one speck of dust can ruin a chip. The patterns are now much smaller than the wavelength of the light that prints them, which takes clever optics, chemistry and repeated exposures.

EUV machines · How it works

A machine the size of a bus prints the smallest chips.

To print the tiniest features, machines from one Dutch company, ASML, fire a laser at about 50,000 drops of molten tin every second. Each drop flashes into a plasma hotter than the surface of the Sun and glows with extreme ultraviolet light.

That light, 13.5 nanometers long, is swallowed by air and by glass, so it travels in a vacuum and bounces off mirrors instead of passing through lenses. Zeiss, which makes the mirrors, has said that if one were the size of Germany, its biggest bump would be about a tenth of a millimeter high. The newest machines reportedly cost several hundred million dollars each, and getting EUV to work took decades of research.

Questions to ask

  • What happens to the world's chips if only one company can make this machine?
  • Why would this light need a vacuum?

The cost of a fab · Money

Only a few companies can afford the best chip factories.

Building a leading-edge chip factory, or fab, costs many billions of dollars: in 2025 TSMC said it planned to spend 165 billion dollars in Arizona on six fabs, two packaging plants and a research center. Only a handful of companies still make the most advanced chips: TSMC, Samsung and, by some counts, Intel.

In 1987 Morris Chang founded TSMC in Taiwan on an idea that was unusual then: make chips for other companies instead of designing your own. That foundry model let designers skip the factory bill and let TSMC spread its costs over many customers. Scale and experience compound, so the leader tends to pull further ahead. In 2021 an industry study estimated that 92% of the world's most advanced chipmaking capacity was in Taiwan.

Taiwan · Power and rules

Most of the world's best chips come from one island.

Taiwan, an island slightly smaller than Maryland and Delaware combined, makes the large majority of the world's most advanced chips. Some call that concentration a "silicon shield"; others call it a single point of failure.

The People's Republic of China claims Taiwan as its territory, while Taiwan has its own elected government, and tension across the Taiwan Strait shapes chip policy worldwide. Taiwan also sits where two tectonic plates collide. After its strongest earthquake in 25 years, in April 2024, TSMC said more than 80% of its chipmaking tools were running again the next day. Building fabs elsewhere spreads the risk, but it costs more and takes years.

30-year question: How should the world lower the risk of depending on one place for its most advanced chips?

Questions to ask

  • What makes something a single point of failure, and how do engineers avoid them?
  • What would happen to your phone, school or hospital if chip supply stopped for a year?

Chips laws · Power and rules

Governments are paying billions to build fabs at home.

In 2022 the United States passed the CHIPS and Science Act, with about 52.7 billion dollars for chip manufacturing, research and training. The European Union, Japan, South Korea, China and India run big chip programs of their own.

The money aims at security and jobs: if a war, pandemic or disaster cut off chips from abroad, a country could still make its own. Critics worry about subsidy races between allies and about paying companies that would have built anyway. In 2025 the US government turned about 8.9 billion dollars of chip money it had promised Intel into a 9.9% stake in the company, a sign the rules are still being written.

Questions to ask

  • Should taxpayers help pay for factories that private companies own?
  • What does a country need besides money to build a chip industry?

The fab's shopping list · Power and rules

A fab's shopping list is a map of world politics.

Fabs need hundreds of ultra-pure chemicals and gases. When one country cuts off one ingredient, chipmakers around the world feel it.

In 2019 Japan tightened exports to South Korea of three chemicals its chipmakers relied on, including photoresists, the coatings that capture the light patterns, and South Korea took the dispute to the World Trade Organization. In 2022 Russia's invasion of Ukraine shut down two Ukrainian companies that purified about half the world's chip-grade neon, a gas used in the lasers of many lithography machines. Each time, companies scrambled, found other sources and learned to keep bigger stockpiles.

Questions to ask

  • Which ingredient in your favorite food comes from farthest away?
  • Is stockpiling a smart backup, or a waste?

Fab water · Planet

Chip factories need rivers of ultra-pure water.

Wafers are rinsed again and again in water purer than anything in nature, and a large fab can use millions of gallons a day. During Taiwan's severe drought in early 2021, chipmakers including TSMC arranged to truck water in as a precaution and said production was not affected.

Ultrapure water has had almost everything removed, even dissolved minerals, so it leaves no trace on parts only a few dozen atoms wide. Fabs now recycle much of their water and build plants to clean wastewater for reuse. Where a fab is built, water rights and droughts matter as much as power and roads.

Questions to ask

  • What makes water pure, and why isn't drinking water pure enough?
  • In a drought, who should get water first: farms, homes or factories?

The people who build it · People

The whole stack runs on skilled people.

Behind every answer are researchers, chip designers, fab technicians in head-to-toe cleanroom suits, electricians, cooling engineers, miners, geologists, lawyers and diplomats. Many of these jobs take years of training.

When the United States set out to build new fabs, a 2023 industry study projected that about 67,000 of the new chip-industry jobs expected by 2030 could go unfilled for lack of trained technicians, engineers and computer scientists. Skills, not just money and machines, decide how fast an industry can grow.

Questions to ask

  • Which job in this whole rabbit hole would you most like to try?
  • How long does it take to train a chip technician, and who pays for it?
Back to the map ↑

Level 7

The ground

The metals and minerals, and who digs them up.

Silicon from quartz · How it works

Chips are made from rock: purified quartz.

Silicon is the second most common element in Earth's crust, after oxygen, and the two together make quartz, the main mineral in a lot of sand. Chip silicon starts as quartz rock, heated with carbon in giant electric furnaces.

That makes metallurgical silicon, about 99% pure, which mostly goes into aluminum alloys and chemicals. A small share is turned into a gas and back into a solid called polysilicon to reach the eleven-nines purity chips need. China made almost 80% of the world's silicon materials in 2025, according to the US Geological Survey, much of it for solar panels.

Questions to ask

  • Silicon is everywhere, so why is chip-grade silicon hard to get?
  • How do you make something purer than nature ever does?

Gallium and germanium · How it works

Two obscure metals help run power supplies and the internet.

Gallium goes into fast power chips and radio chips; germanium goes into the glass of fiber-optic cables and into some sensors and chips. Neither is mined for itself. Both are recovered while refining other metals.

China produced 99% of the world's primary low-purity gallium in 2025 and is the leading producer of germanium, according to the US Geological Survey. Since 2023 China has required licenses to export both, and in December 2024 it banned their sale to the United States. Other countries are restarting recovery at aluminum and zinc refineries, which takes years.

Questions to ask

  • Why is it risky when a metal is only a side product of something else?
  • What would a new country need to start producing gallium?

Rare earths · How it works

Rare earths aren't rare, but they're hard to separate.

The 17 rare-earth elements make the strongest permanent magnets there are. In a data center they spin hard drives and fans; they're also in motors, wind turbines and phones. Despite the name, some are about as common in rock as copper.

The hard part is separation. Rare earths are chemically almost identical, so pulling them apart takes hundreds of stages of chemical extraction. China mines about two-thirds of the world's rare earths, according to the US Geological Survey, and does about 90% of the separating, according to the International Energy Agency. That lead took decades to build, and other countries are now racing to build mines, refineries and magnet factories.

Questions to ask

  • If an element is common in rock, what makes it "critical"?
  • Could we design motors and hard drives that need fewer rare earths?

Copper · How it works

Every chip and power line runs on copper.

Inside a chip, the wires linking billions of transistors are copper, a change IBM announced in 1997. Outside, copper carries power through data centers, transformers and the grid. Chile mines more of it than any other country.

AI's appetite for power means more copper for cables and grid upgrades, on top of copper for electric cars and renewables. New copper mines have typically taken well over a decade to go from discovery to production, and the ore in older mines is getting poorer. Copper can be recycled again and again without losing quality.

The byproduct trap · Money

You can't just dig up more gallium.

Gallium comes from refining aluminum ore, germanium from refining zinc, and helium from processing natural gas. Their supply depends on demand for something else, so a price spike doesn't quickly bring more.

Economists call that inelastic supply: a zinc refinery won't double its zinc output just because germanium got pricey. Recovering a byproduct needs extra equipment that only pays off if prices stay high, and one country can dominate simply because it refines the main metal. That's why governments keep stockpiles, and why small markets are easy to squeeze.

Questions to ask

  • If a price doubles, why doesn't supply always follow?
  • What is a byproduct in your kitchen or your school?

Mineral politics · Power and rules

Minerals have become bargaining chips.

Since 2023 China, which dominates the processing of many critical minerals, has put export controls on gallium, germanium, graphite, antimony and several rare earths, often soon after new US chip rules.

Other countries are answering with critical-mineral lists, stockpiles, money for new mines and refineries, and deals with mining countries; the US list grew to 60 minerals in 2025. In November 2025 China suspended its newest rare-earth controls for a year and paused its ban on gallium, germanium and antimony sales to the United States until late November 2026, while still requiring licenses. A mine or refinery takes years to build, while a policy can change overnight.

Mining and human rights · People

Some electronics metals are dug by hand, sometimes by children.

Most of the world's cobalt, used in rechargeable batteries, comes from the Democratic Republic of the Congo, and some of it is dug by hand in dangerous small mines. Human-rights groups have documented children working in them.

Tin, tantalum, tungsten and gold are all used in electronics, and tin even fuels EUV machines. They have been mined in conflict zones where the money paid for armed groups. Laws in the United States (2010) and the European Union (in force since 2021) make companies trace and report where these "conflict minerals" come from, and industry programs audit the smelters. Tracing a metal through many traders is hard, and progress is real but uneven.

E-waste · Planet

Most old electronics are never recycled.

The world threw away about 62 million tonnes of electronics in 2022, and only about 22% of it was documented as properly collected and recycled. Old servers, phones and laptops are full of copper, gold and rare earths.

Data centers swap out chips every few years as faster ones arrive, adding to the pile. Some metals, like copper and gold, are recovered well; many others, including rare earths, are rarely recovered, because separating tiny amounts from mixed scrap is hard and costly. Urban mining, treating old electronics as ore, is a growing field.

30-year question: Can we recover the metals in old electronics cheaply enough to need fewer new mines?

Questions to ask

  • Where does your family's old phone end up?
  • Should companies have to take back what they sell?
Back to the map ↑

Level 8

The rocks

Why the minerals are where they are.

Ore deposits · How it works

Metals pile up where heat, water and time concentrate them.

Most metals are spread so thinly through rock that mining them would be hopeless. An ore deposit is a rare place where geology concentrated one element hundreds or thousands of times over, usually through magma, hot water or millions of years of weathering.

Plate tectonics drives much of it. Where one plate dives under another, melted rock rises, and hot, salty fluids carry metals upward and drop them in cracks as they cool. The same collisions build mountains and shake places like Taiwan and Chile with earthquakes. A map of mines is really a map of Earth's history.

Questions to ask

  • Why are there giant copper mines in the Andes but not in the middle of Kansas?
  • What would a geologist look for to find a new deposit?

Copper's volcano roots · How it works

Most copper comes from the roots of ancient volcanoes.

About 60% of the world's mined copper comes from porphyry deposits: huge bodies of rock that formed a few kilometers under volcanoes, above ocean plates sinking into the mantle.

As the magma cools it releases metal-rich fluids that crack the rock above and fill it with tiny veins of copper minerals. That's why great copper belts follow subduction zones, old and young: the Andes, the southwestern United States and the Pacific Ring of Fire. A typical deposit is under 1% copper, so mines move mountains of rock to get it.

Questions to ask

  • How can a volcano be a metal factory?
  • If ore is under 1% copper, what happens to the other 99%?

Rare-earth rocks · How it works

Rare earths gather in strange volcanic rocks and weathered clays.

Many of the richest rare-earth ores formed in carbonatites, unusual rocks that crystallized from magma rich in carbonate. Mountain Pass in California is one; Bayan Obo in Inner Mongolia is the largest rare-earth deposit known.

A second kind, ion-adsorption clay, formed in southern China and Southeast Asia as tropical rain slowly rotted granite and left rare earths clinging loosely to clay. These clays are easy to leach and rich in the heavy rare earths that magnets need, but mining them, including widespread illegal mining in Myanmar, has poisoned rivers and land. Where these ores formed millions of years ago decides which countries hold the cards today.

Bauxite · How it works

Aluminum ore is ancient soil, and it hides gallium.

Bauxite, the ore of aluminum, formed over millions of years where hot, rainy climates washed nearly everything soluble out of rock and left aluminum behind. Traces of gallium came along for the ride.

That's why big bauxite deposits sit in tropical or once-tropical lands such as Guinea, Australia, Brazil, India and Jamaica. Bauxite holds about 50 parts per million of gallium on average, which is recovered from the liquid used to refine aluminum oxide. A country that refines lots of aluminum can make gallium cheaply; a country that doesn't can't easily start.

Questions to ask

  • How can ancient weather decide where a modern factory gets its metal?
  • What else might be hiding in a refinery's leftovers?

Helium · How it works

The helium a fab uses was made by radioactive rocks.

Fabs use helium to cool wafers and to keep air away from delicate steps. Almost all of Earth's helium was made underground, as uranium and thorium slowly decayed over hundreds of millions of years, and it collected in some natural gas fields.

Helium is so light that once released it escapes to space, so on human timescales it is effectively non-renewable. It is only worth extracting where gas fields happen to hold enough: mainly in the United States and Qatar, with Russia and Algeria also producing. That has made supply prone to shortages, and fabs are learning to capture and recycle it.

Questions to ask

  • Why would the lightest gas be hard to keep?
  • How can a rock make a gas?

Spruce Pine quartz · Planet

A small town in North Carolina mines ultra-pure quartz.

Around Spruce Pine, North Carolina, rocks called pegmatites hold quartz pure enough to make the crucibles that melt silicon for chip crystals. Most of the world's high-purity quartz comes from this one small area.

The pegmatites crystallized deep underground more than 300 million years ago, during the mountain-building that shaped the Appalachians, cooling slowly enough to grow large, clean crystals. Hurricane Helene flooded the area in late September 2024 and shut both mines for about two weeks; stockpiles softened the blow. One small place, one storm, a worldwide wobble.

Questions to ask

  • What other industries might depend on one small place?
  • Should companies keep bigger stockpiles of things that come from one place?

Deep-sea nodules · Planet

Metal-rich lumps cover parts of the deep sea floor.

Across huge stretches of the Pacific, 4,000 meters down or deeper, the sea floor is scattered with potato-sized nodules rich in manganese, nickel, cobalt and copper. They grow only millimeters every million years.

Companies want to collect them for batteries and electronics. Scientists warn that the nodules are home to animals found nowhere else, many not yet named, and that a mined area could take far longer than a human lifetime to recover. The International Seabed Authority, set up under the UN Law of the Sea, has negotiated mining rules for years without finishing them, and some countries want a pause. Meanwhile the US moved in 2025 to allow seabed mining under its own law, and in January 2026 it finalized permit rules.

30-year question: Can we mine the deep sea without destroying ecosystems we barely understand?

Questions to ask

  • Who owns the deep sea floor?
  • How do you weigh a metal people need now against life nobody has discovered yet?

Deep time · Ideas and history

Ores took millions of years to form; we use them up in decades.

Earth is about 4.54 billion years old. The ores in a chip took millions to billions of years to form. We dig up, use and throw away many of them within a single human lifetime.

Geologists measure deep time with radioactive clocks inside minerals, the same kind of decay that makes Earth's helium. The idea that Earth is unimaginably old took a long time to accept: in 1788 James Hutton wrote that he could find "no vestige of a beginning". Thinking in deep time changes what we call renewable, and what we think we owe the future.

Questions to ask

  • Which resources are renewable if you count in millions of years?
  • What would you need to make your own metal, starting from rock?
Back to the map ↑

Level 9

The stars

Where the atoms came from.

The first atoms · How it works

The hydrogen in a fab's water is 13.8 billion years old.

In the first minutes after the Big Bang, the universe made hydrogen, helium and a pinch of lithium, and almost nothing else. Nearly every hydrogen atom in the water that rinses a wafer dates from then.

Predicting how much of each light element the early universe made, then measuring it in ancient gas clouds, is one of the strongest tests of the Big Bang theory. Everything heavier, including the silicon, oxygen and metals in a chip, had to be made later, inside stars.

Questions to ask

  • How can anyone know what happened in the first few minutes of the universe?
  • Which atoms in your body are the oldest?

Star furnaces · How it works

The silicon in a chip was made inside a star.

Big stars fuse light atoms into heavier ones in layers, like an onion: helium into carbon, carbon into neon, oxygen into silicon. When such a star explodes as a supernova, it scatters the silicon and oxygen that later became quartz.

Our solar system formed about 4.6 billion years ago from a cloud enriched by generations of these stars. Fusion stops paying off at iron, so once a giant star's core turns to iron it collapses. As Carl Sagan put it, we are made of star stuff, and so is every chip.

Questions to ask

  • Why does a star's furnace stop at iron?
  • If stars stop at iron, where did the gold in a chip come from?

Star collisions · How it works

Some of Earth's gold and rare earths were likely forged when dead stars collided.

Elements heavier than iron need a flood of neutrons to build up. Some are made slowly inside aging giant stars; many of the heaviest, like gold and some rare earths, are thought to come from violent events such as two neutron stars smashing together.

In August 2017, detectors felt the gravitational waves of two neutron stars merging about 130 million light-years away, and telescopes then saw the glow of freshly made heavy elements. The neodymium in a hard-drive magnet may have come from either kind of star. Exactly where the heavy elements come from is still being worked out.

30-year question: Which cosmic events made the universe's heavy elements, and in what shares?

Questions to ask

  • How can a gravitational wave tell us what a star collision made?
  • Why are heavy elements rarer than light ones?

Reading starlight · Ideas and history

We learned what stars are made of by splitting their light.

In 1925 Cecilia Payne showed that stars are made mostly of hydrogen and helium, by reading the dark lines in their light spread into a rainbow. The famous astronomer Henry Norris Russell called the result impossible, so she played it down in her thesis. A few years later he agreed she was right.

Spectroscopy turns light into information: every element absorbs its own pattern of colors, like a barcode. The idea of decoding a signal to learn about its source runs from astronomy to Shannon's information theory. Nearly everything we know about the universe arrived as light, measured carefully.

Questions to ask

  • How could you tell what something is made of without touching it?
  • Why do you think experts doubted Payne at first?
Back to the map ↑

› How this page stays honest

A rabbit hole with exits.

  • Most rabbit holes online are built to keep you scrolling. This one is built to send you out: nothing here picks what you see next or plays by itself, and every stop ends in sources that leave this site.
  • The sources are mostly original papers, government data and official documents, plus careful explainers and news reports.
  • Numbers that change carry a year. Laws, prices and export rules move fast, so check a source for today's version.
  • Where experts disagree, the stop says so. Open questions are left open.
  • Every stop and source was last checked on November 18, 2026. Spot a mistake? Tell us.