Topics
Thinking machines · every notable AI model since 1950
Thinking machines, counted against the people
- 107notable AI models released in 2025
- ×4.3a year: how fast the computing to train them grew, 2010–2025
- $71invested in AI by companies in 2025, for every person alive
- 68.5 millionpeople: the electricity AI data centres could use in a year
In 2025 the world released 107 notable AI models: computer programs that learned a skill from data, and that Epoch AI counts as important. Since 1950 there have been 994.
the largest of their day (Epoch's "frontier" models) · every other notable model
How to read this: one dot is one model, placed at the day it was published. The higher the dot, the more computing it took to train. Each line up is 1,000 times more: 10³ is a thousand steps of arithmetic, 10²⁷ a thousand million million million million. Only the 519 models whose training Epoch AI could estimate are drawn.
Verified 2026-09-30 · 73 canaries
In plain words
What is it?
An AI model is a computer program that learned a skill, like writing, seeing or playing a game, by practising on huge amounts of data instead of being told every step. Teaching it is called training.
Training takes computing: billions upon billions of small sums, done by thousands of chips in a data centre. This page counts the notable models, how much computing they took, who made them, and what they cost in money and electricity. The biggest of each year are frontier models.
As everywhere on this site, each figure is set against the people alive: the world's 107 notable models of 2025 beside its eight billion people.
Why it matters to you
- Your electricity. AI data centres could use 8 of every 1,000 units the world makes.
- Your money. Companies invested $71 for each person alive in 2025, nearly half of it in one country.
- Who decides. 93 of every 100 notable models came from two countries that hold 21 of every 100 people.
- Pace. Training computing doubled every 5.7 months; at their pace, the world's people take 65 years.
Training computing against people · 2010–2025
How fast it grows
From 2010 to 2025 the world's people grew by 17 of every 100. The computing used to train AI models doubled about 32 times.
Training computing: doubled 32 times
The world's people: about a quarter of one doubling
At the pace of those years, training computing grew about ×4.3 a year: it doubled every 5.7 months. For the largest models alone it was ×4.6. The world's people grew ×1.0107 a year (1.07 of every 100). At that pace they would take 65 years to double.
How to read this: one filled square is one doubling, twice as much as before. The first row is the computing used to train notable AI models; the second is the world's people over the same 15 years. The growth is the trend through the 457 models since 2010 with a known training compute, not any one model.
Models, papers and people · by country, 2025
Who builds it
Of every 100 notable models in 2025, 60 had a maker in the United States and 33 in China. Those two countries hold 21 of every 100 people. China writes the most AI research papers: 18 of every 100.
Notable models, of every 100
United States59.8
China32.7
South Korea7.5
United Kingdom0.9
France0.9
Canada0.9
Hong Kong0.9
Singapore0.9
India0
Indonesia0
Germany0
Italy0
Japan0
AI research papers, of every 100
- United States11.9
- China18.0
- South Korea1.4
- United Kingdom3.3
- France2.0
- Canada1.9
- Hong Kong1.1
- Singapore0.9
- India7.5
- Indonesia3.3
- Germany2.9
- Italy2.1
- Japan2.0
People, of every 100
- United States4.2
- China17.2
- South Korea0.6
- United Kingdom0.8
- France0.8
- Canada0.5
- Hong Kong0.1
- Singapore0.1
- India17.8
- Indonesia3.5
- Germany1.0
- Italy0.7
- Japan1.5
How to read this: each bar is a country's share of the world's notable AI models, AI research papers or people in 2025. A full bar would be all 100. A model or paper made in two countries counts once in each, so a column can add up to more than 100 (4 of the 107 models had makers in two countries).
AI research papers, every year
Scientists published 297,775 papers about AI in 2025, more than twice as many as in 2015 (128,052).
How to read this: one column for each year, 2000 to 2025; the taller the column, the more papers. The count is OpenAlex's: works whose main subject its computers class as artificial intelligence.
Every country in these lists
| Country | Models, 2025 | Models since 1950 | AI papers, 2025 | People, 2025 |
|---|---|---|---|---|
| United States | 64 | 647 | 35,401 | 347.3 million |
| China | 35 | 160 | 53,589 | 1416.1 million |
| South Korea | 8 | 23 | 4,293 | 51.7 million |
| United Kingdom | 1 | 89 | 9,895 | 69.6 million |
| France | 1 | 50 | 5,932 | 66.7 million |
| Canada | 1 | 57 | 5,529 | 40.1 million |
| Hong Kong | 1 | 15 | 3,322 | 7.4 million |
| Singapore | 1 | 16 | 2,715 | 5.9 million |
| India | 0 | 2 | 22,203 | 1463.9 million |
| Indonesia | 0 | 0 | 9,847 | 285.7 million |
| Germany | 0 | 31 | 8,675 | 84.1 million |
| Italy | 0 | 3 | 6,225 | 59.1 million |
| Japan | 0 | 11 | 6,004 | 123.1 million |
The lists hold every country with a notable model in 2025 and the ten that wrote the most AI papers.
Investment · of every 1,000 dollars
The money
Companies put $581.7 billion into AI in 2025: about $71 for every person alive.
Most of it was in one country. In the words of the AI Index: “Its investments ($285.9 billion) were 23.1 times greater than those of the next-highest country, China ($12.4 billion).” For each person living there, that is about $823 in the United States and $9 in China.
How to read this: the 1,000 squares are everything the world made in 2025, in dollars of that year (the World Bank's measure of output). The filled squares are what companies invested in AI, as the Stanford AI Index reports it. "For every person alive" divides it by everyone, babies included.
Data centres · of every 1,000 units of electricity
The power
AI data centres could draw 29.6 gigawatts in 2025. Run flat out for a year, that is the electricity of about 68.5 million people, about as many as live in the United Kingdom.
How to read this: the 1,000 squares are all the electricity the world used in a year. The filled ones are the most AI data centres could use. It is an upper limit: data centres do not run at full power all the time. "The electricity of N people" divides it by what the world uses for each person, 3,787 kilowatt-hours a year (electricity per person).
Limits
What this page cannot say
- How many people use AI. The only figures are companies' own claims about their users. They are not counts anyone can check, so they are not shown.
- Electricity actually used. The one open figure is how much power AI data centres could draw. How much they did use is estimated only in reports this site may not reproduce.
- Patents. The world count of AI patents is published without terms that allow it here.
- This year. Epoch AI already lists 75 models for 2026, but the year is not over, so the page stops at 2025.
- "Notable" is a judgment. Epoch AI counts a model when it set a record, was widely cited, was historically important or widely used. Training computing is Epoch's estimate, known for 519 of 994 models.
- Papers are sorted by a machine. OpenAlex's computers decide each paper's main subject, so some AI papers are missed and some others counted.
Back to the count
How this rests on the census
AlexNet, the model that began today's rush of learning machines, was published in 2012. Everyone aged 12 or younger in mid-2025 was born in 2012 or later: 1.73 billion people, 21 of every 100. They have never known a world without it.
aged 12 or younger everyone else
How to read this: one column for each year of age, from newborn babies on the left to people aged 100 and over on the right. The taller the column, the more people of that age were alive in mid-2025 — 8.23 billion in all, counted by the UN.
In those years the computing used to train notable models doubled about 27 times.
Questions
Questions people ask
How many AI models were released in 2025?
Epoch AI counts 107 notable AI models published in 2025, and 994 since 1950.
Which countries make the most AI models?
The United States: 64 of the 107 notable models of 2025 had a maker there. China had 35. A model made in two countries counts in both.
How fast is AI training growing?
The computing used to train notable models grew about ×4.3 a year from 2010 to 2025, doubling every 5.7 months.
How much electricity does AI use?
AI data centres could draw 29.6 gigawatts in 2025. Run flat out for a year that is 259 terawatt-hours, the electricity of about 68.5 million people, as many as live in the United Kingdom.
How much money goes into AI?
Companies invested $581.7 billion in 2025 (Stanford AI Index): $71 for every person alive, about 4.9 of every 1,000 dollars the world made.
How many research papers about AI are written?
297,775 in 2025, by OpenAlex's count, against 128,052 in 2015.
Where do the numbers come from?
Models from Epoch AI (CC BY 4.0), papers from OpenAlex (CC0), money and power quoted from the Stanford AI Index 2026, output from the World Bank, electricity from Ember, people from the UN. Every source is on the references page.
Apparatus
Sources and method
Method: notable models are Epoch AI's list, fetched on 2026-09-19; a model counts once in every country where one of its makers is based. The growth of training computing is a straight-line trend through the logarithm of each model's computing against its publication date, 2010–2025. The world's people are the UN's mid-year counts. AI papers are OpenAlex works whose main topic is in its subfield "Artificial Intelligence", fetched on 2026-09-19, counted by the countries of their authors. Money and power are the Stanford AI Index 2026's figures, quoted: its licence allows no redrawn charts, so the squares are our own division: investment over the World Bank's world output in current dollars, and 29.6 gigawatts × 8,760 hours over Ember's world electricity use.
Sources: UN WPP 2024 · Epoch AI · OpenAlex · World Bank · Stanford AI Index 2026 · flag-icons. Every source, its licence and how to cite it →
Verified 2026-09-30 · 73 canaries