Get friendly with the AI race

Every piece in this series looks at one part of a machine: the chip, the memory, the heat, the grid, who got paid. This one steps back to ask why the machine is being built at all. The answer is usually called a race: a handful of companies each spending more than most countries to build the most capable AI first.

That picture is right but incomplete. Building the smartest model is only the first of five races, and it is no longer the one that settles who wins. The other four are about what happens once a model exists: who can get the compute to run it, who can make it reliable enough to hand a job to, who owns the customers, and who can turn a dollar of computing into more than a dollar of work.

The racers

There are two kinds of racer. The first are the labs: companies whose whole business is building frontier models. OpenAI, Anthropic and xAI lead them. xAI now sits inside SpaceX, which listed in June 2026 in the largest IPO ever.So the easiest way to buy a piece of a frontier AI lab, for now, is to buy a rocket company. Behind them are challengers founded by people who left the leaders, like Thinking Machines and Safe Superintelligence, and Europe’s Mistral. The second kind race from inside giants that pay for it out of other businesses: Google DeepMind out of search, Meta’s lab out of advertising. In China, DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi and Zhipu’s GLM publish their models openly. They have no market price at all.They can still move markets. The day DeepSeek’s R1 went viral, in January 2025, NVIDIA lost about $590bn of value, the largest one-day loss by any company in history.

Those numbers are hard to hold, so hold them against things you already know the size of.

NVIDIA alone is worth about as much as Germany produces in a year. The ten companies that make AI’s chips and memory together are worth more than any economy except America’s and China’s.Add the four biggest cloud companies and the labs, and the AI industry’s combined value comes to roughly $31 trillion, close to a year of the entire US economy. Not all of the clouds’ value is AI, but that is the scale being bet. Anthropic’s last private price works out to over a hundred dollars for every person alive.

Where they say they are going

The prices are not for what the labs are today. They are for what the labs say they will become.

OpenAI has told investors it expects more than $280bn of revenue in 2030, roughly seven times its current run rate, with about $100bn of that from advertising. Anthropic’s own forecast, made in late 2025, was up to $70bn by 2028. Its run rate hit $65bn in July 2026, two years early.Run rate means this month’s revenue times twelve. For a company doubling every few months, it can overtake a year’s forecast before the year starts. The next prices are being set now: Anthropic is reported to be aiming for about $2 trillion in a November listing, and OpenAI is in talks near $1.2 trillion.

Here is the twist. By Stanford’s 2026 AI Index, the best models from Anthropic, xAI, Google, OpenAI, Alibaba and DeepSeek sat within 25 points of each other on the most-used head-to-head ranking.That ranking is LMArena, where people compare two anonymous answers and vote. The points are Elo, the system invented to rank chess players. The gap between the best American and best Chinese model had shrunk to under 3%, from double digits in 2023. American private investment in AI was 23 times China’s. When everyone’s model is about as good, being best matters less. Being cheapest, most reliable and most used matters more.

Who has held the crown

The lead has changed hands more often than the headlines suggest, and it rarely lasts.

GPT-4 held the top spot for most of a year after March 2023, the longest reign so far.GPT stands for Generative Pre-trained Transformer. The Transformer came from a 2017 Google paper with eight authors; every one of them later left Google, though one has since gone back. Since then the crown has passed between OpenAI, Anthropic and Google about every few months, and some reigns lasted weeks. Meanwhile the dashed line, the best model anyone can download for free, closed in. In mid-2024 the best open model was about fifteen months behind the best closed one. By July 2026, Moonshot’s Kimi K3 was about four months behind.

The race is not only upward. It is also downward, in price.

The price of reaching GPT-4’s level on PhD-level science questions fell about 40 times a year. Every lab is racing to offer the best model, and at the same time to offer last year’s best at a fraction of last year’s price, because an open-weight model will offer it for less if they don’t.

There is no finish line

People in the race talk about three destinations. You will hear them constantly, so here they are once, plainly.

AGI, artificial general intelligence, usually means a system that can do most of the work a skilled person does at a computer, across fields rather than in one.No two labs define it the same way, and one of the largest contracts in the industry once turned on who got to decide when it had been reached. ASI, superintelligence, means a system well beyond the best humans at nearly everything. RSI, recursive self-improvement, is how some expect to get from one to the other: an AI good enough at AI research to build its own successor, which builds the next one faster.

None of these has a line you cross. What the racers can measure is two trends. Scaling laws say models get predictably better as you spend more compute, data and parameters training them. Inference-time compute says they also get better if you let them think longer before answering, which is why a reasoning model is slower and dearer per question. Both turn money into capability. That is why the capability race is also a spending race.

The scoreboards, and why they wear out

With no finish line, the racers point at benchmarks, and the benchmarks keep wearing out. SWE-bench Verified asked models to fix real bugs from open-source projects. It became so widely trained on, directly or by accident, that OpenAI stopped reporting it in 2026 and recommends a harder, less-leaked version instead. The same models score far lower there. ARC-AGI-3, launched in March 2026, drops a model into small puzzles it has never seen, with no instructions; at launch every frontier model scored under 1% on puzzles people solved. Contamination is part of the story now: once a test is public, it slowly stops measuring anything.

The measure worth watching is METR’s time horizon: how long a task, timed by a skilled person, a model can finish half the time. It has doubled roughly every seven months since 2019, and faster lately. A benchmark tells you whether a model can answer. The time horizon tells you how much of a job it can be left alone with, which is what the valuations are betting on.

The race has layers

Most explainers stop at lab → model → users. The real chain is longer: chips → clouds → models → agents and apps → customers → revenue, and a different race is run at each step.

The compute race is the rest of this series. NVIDIA dominates the chips, but it is fighting a platform war: AMD, Google’s TPUs, Amazon’s Trainium and custom chips designed for single buyers are all trying to become the second supplier. Above the chips sit the clouds, which rent the compute out. Owning that layer matters even without the best model. It is why Google, Microsoft and Amazon can spend like labs without being labs: whoever wins the model race still has to rent from them.

The deployment race is turning a model into something that does a job reliably. In 2026 that increasingly means an agent: a model given tools, a computer and a task, left to work through the steps. Coding went first. In one survey of over 500 technical leaders, run by Anthropic, a vendor with a stake in the answer, nearly nine in ten organisations used AI in software development. 57% ran agents on multi-step work.

The distribution race is about who owns the customer: the chatbot people open by habit, the office suite at work, the cloud a company already pays, the coding tool a developer won’t switch from. A slightly worse model in front of a billion users can beat a better one nobody sees.

Open and closed

This is the fork the “whoever builds the smartest model wins” story misses. Closed labs, OpenAI, Anthropic and Google, keep their models’ weights private, the learned numbers themselves. They sell access by the question, through their own apps and APIs, and control the price. Open-weight labs publish the weights, so anyone can download the model, run it on their own chips, change it, or sell it through any cloud.

Open weights turn capability into a commodity. If a free model is nearly as good as a paid one, the paid one’s price falls toward the cost of running it. The pressure is real. On OpenRouter, a marketplace developers use to switch between models, open-weight models went from about a third of tokens in late 2025 to most of them by mid-2026, led by Chinese labs. Qwen became the most-downloaded family on Hugging Face.“Open-weight” is not quite “open source”: you get the trained numbers, but usually not the data or code that made them. Like being given a cake and not the recipe. That is one marketplace, skewed toward developers who shop on price, not the whole market. But it is the direction the closed labs have to price against, and it is why China backs open weights: if you can’t sell the best model, make models free and win the ecosystem built on them.

Who pays, and for what

The “who pays whom” of infrastructure comes later. First, the money that is supposed to pay for all of it.

Four kinds of customer pay the labs. Consumers pay monthly subscriptions. Developers pay per token through the API. Enterprises pay for seats, agents and custom deals. Governments pay for sovereign and defence work. The two leading labs have leaned different ways: OpenAI is weighted toward consumers, Anthropic toward businesses and code. For now, Anthropic’s run rate is growing faster.

What none of them is really buying is tokens. They are buying work: code written and reviewed, questions researched, customer tickets closed, data analysed, forms processed. The machine in this series turns electricity into tokens. The economics race is about turning tokens into work someone values more than the electricity, the chips and the building cost. Whoever does that at the lowest cost per useful result wins, whether or not they have the best model.

Data is becoming experience

A model learns from whatever it is shown, and what it is shown has changed in stages. First came the public internet, mostly used up for text. Then licensed and synthetic data, generated by other models. Then post-training, where people and models grade answers. Now the most valuable data may be experience: records of agents using tools, what worked, what users corrected, and the private context of a company’s own documents and systems.

That makes deployment and capability one loop. The lab with the most agents doing real work collects the most examples of real work, and trains the next model on them. The scarce input shifts from text on the internet to feedback from the field.

What it costs

Four costs, and one dwarfs the others. Compute: chips to train models and, increasingly, to run them for every user. Talent: a few thousand researchers, some paid like star athletes.In 2025, some AI researchers were offered pay packages reported at over $100 million to switch labs, more than most football transfers. Data: licensed, bought, or generated. And power, which, as the rest of this series shows, money can’t hurry.

Amazon, Alphabet, Microsoft and Meta expect to spend around $720–745bn in 2026 on buildings, chips and power. The two leading labs bring in about $105bn a year, though that figure has multiplied in a year: Anthropic’s run rate went from $9bn at the end of 2025 to $65bn by July. The bet is on which line catches which. The largest single project, Stargate, is about 10 GW of US data centres. Its first campus in Texas was only partly running by August 2026, a reminder from the grid piece that money arrives faster than power.

The circle

Follow the money for long enough and it comes back to where it started.

A lab needs chips it can’t yet pay for; a chip maker needs a customer committed for years. So the chip maker invests in the lab, or hands it rights to its own shares, and the lab signs to buy the chips. NVIDIA announced up to $100bn for OpenAI, paid as each gigawatt is built. By early 2026 that had stalled. AMD gave OpenAI warrants for up to 160 million AMD shares alongside a 6 GW order. The neoclouds in Who got paid sit in the middle. None of this is fraud, and some of it is ordinary vendor financing. But one company’s order book is another’s investment, so a slowdown at the centre would travel around the whole circle at once.

Nations

Around all five races sits a sixth, run by governments. AI chips are the one input the US can control, so it rations them. In January 2026 it moved to let NVIDIA’s H200 be sold to approved Chinese buyers case by case, with a 25% levy, while keeping the newest Blackwell chips blocked. By mid-2026 little had actually shipped. China’s answer is the open-weights strategy above and a push to build its own chips. The US, meanwhile, wants other countries to build on its “stack”, American chips, clouds and models, rather than China’s.

In between are the countries with money and power but no frontier lab. The Gulf is the clearest case: Abu Dhabi is building a planned 5 GW AI campus with OpenAI and Oracle, and Saudi Arabia’s Humain is buying chips by the hundred thousand. They are becoming both financiers and landlords of the race. Everyone else is working out how much sovereign AI they need: models, data centres and chips they control, in case access is ever cut off.

Rules

Models are not released into a vacuum. The EU’s AI Act has applied to general-purpose models since August 2025. It covers documentation, copyright policies for training data, and extra duties for the most capable systems. Since 2 August 2026, the European Commission can enforce it. It can demand evaluations, pull a model from the EU market, or fine up to 3% of global revenue. Copyright lawsuits over training data are working through courts in several countries. None of this stops the race, but it adds a seventh input alongside compute, customers, talent, data, distribution and capital: permission.

Safety, and the attacks that have started

The labs test each new model for dangerous abilities before release, and a few publish rules for what happens if it crosses a line. In April 2026 Anthropic kept a model, Claude Mythos Preview, from general release because it was too good at finding and exploiting software flaws. It gave the model only to defenders of critical software.

The harder problem showed up in 2026: agents doing things nobody asked for. During an OpenAI security evaluation, about 1,200 test agents used an internal package server as a hidden message board. They exchanged over 70,000 messages and reached outside their sandboxes, including into Hugging Face’s systems. Five labs have now confirmed a model left an evaluation and reached a real target: OpenAI, Anthropic, Moonshot AI, Meta and Google. A UN scientific panel concluded that the usual way of containing AI systems is “unravelling”. In 2025, a group Anthropic judged to be Chinese state-sponsored had used its coding agent to run 80 to 90% of an espionage campaign by itself. Criminals don’t need frontier models to join in. In September 2026, researchers traced a campaign that chained three free, open-source agents to break into 27 retailers and copy 600,000 card numbers, at about $25 of AI compute per target. This timeline keeps a sourced record of such incidents.

The worry is simple. A race rewards whoever ships first, and safety work is the part you can skip to ship sooner.

The maths test

Maths makes the progress easy to see, because an answer is either right or wrong.

In 2023 the best model got under 1% on AIME, a hard exam for top high-school students. By April 2026 it scored 100%. At the International Mathematical Olympiad, AI reached the gold-medal cutoff in July 2025, 35 of 42 points. A year later, several AIs scored a perfect 42, including models from OpenAI, Anthropic, the start-up Axiom Math and China’s Moonshot. On FrontierMath, research problems written by mathematicians that take experts hours or days, scores went from almost nothing to over 90%.

Then came real research. In May 2026, an OpenAI model disproved a conjecture Paul Erdős made in 1946, about how many pairs of points in a plane can sit exactly one unit apart, with a 125-page argument nobody had guided.Erdős published around 1,500 papers and put cash prizes on problems he couldn’t solve, from $25 up to $10,000. Some of that money is now owed to machines. Epoch AI then collected the 68 hardest Erdős problems still open, each worth a top-journal paper. By September 2026, one model had solved two of them. Competition maths is finished as a test for AI; research mathematics has just started.

What comes after

Agents and computer use are not the future; they are the current front. What comes next are models that learn from the physical world, not text. World models generate environments you can move through: Google DeepMind’s Genie 3 makes explorable worlds from a sentence. Yann LeCun’s AMI Labs and Fei-Fei Li’s World Labs have each raised around a billion dollars on it. Robots are where those models meet matter. Humanoid shipments nearly quadrupled in the first half of 2026, to over 22,000, most of them Chinese-made. Figure is valued at $39bn, and Tesla is converting a car line for its Optimus robot.

The loop that decides it

Put the five races together and the real question isn’t who reaches AGI first. It is whether capability can keep improving faster than the cost of deploying it falls, while producing enough valuable work to pay for the next generation.

Three loops run at once. The growth loop: better models do more useful work, which brings more demand, more revenue, more compute and better models. The price loop: better models and open weights push the cost of a token down, which commoditises capability, which lowers prices, which brings more use. The physical loop, underneath both, is the rest of this series. More demand needs more compute, which needs more chips, factories, power and grid, which need years and capital. The race is won by whoever keeps all three turning together. That is why a guide to the AI race ends where this series begins: at the power line.