The International Federation of Robotics counted 542,000 industrial robot installations worldwide in 2024. China took 295,000 of them, more than half. Europe installed about 85,000, and the United States 34,200.
The country whose frontier models usually sit at the top of the benchmarks performs last.
The so what here is simple. Read the table as an Olympic arena. You cannot compare a 100-metre runner with a swimmer or an equestrian rider. Many athletes compete in the same arena at once; the events sometimes overlap, and from the stands you can see them almost in the same place. They are competing in different sports. So the first job is to understand which race you are watching before you evaluate anyone in it.
AI is scored from the stands. Stanford’s AI Index measures the model race, and there the United States leads, with China 2.7 percent behind. The IFR table measures the machines race, and there China leads and the United States trails. Europe runs a third event, on the factory floor and in the equipment the other two depend on. For a chief executive, the useful question is smaller than who is winning: which of those races does the company’s own AI depend on, and where could it be cut?
Every region leads one layer and buys the rest.
Each region leads one layer and buys the others from somewhere else. The United States designs the leading chips, and TSMC in Taiwan manufactures almost all of them, on Stanford’s account. TSMC’s most advanced lines run on chipmaking machines from ASML in the Netherlands. Europe holds that step and rents most of its cloud from three American companies. China installs more industrial robots than the rest of the world combined and still needs a United States licence, granted case by case, to import top-end American chips such as Nvidia’s H200.
Any company’s AI runs on five things it does not make: the power, the chips, the cloud it rents, the model it calls, and the machines that act on the answer. Each comes from a different place, and each can be cut by a different hand, a regulator, an export office, or a supplier changing its terms with no government in the room. A company whose five supplies all sit inside one bloc has taken a position on geopolitics without deciding to.
Intelligence is becoming something a company rents.
The price of a unit of intelligence keeps falling. AI is sold by the token, a fragment of a word, and Mistral Large 3 charges $0.50 for a million of them where Claude Opus 5 charges $5 and reads as the premium product because it is one. The bill moves the other way. The number of tokens a single task consumes has grown by orders of magnitude, because today’s models think before they answer and today’s AI agents, which work through a task on their own, read and re-read whatever they are handed, so the cheaper unit price arrives on a larger invoice. Chen and colleagues named it the price reversal: the model that is cheaper per unit finishes the job at a higher total cost. Both numbers are real.
The consequence for the buyer is that this year’s model is a temporary choice. The gap between the leading American and Chinese models is 2.7 percent. The gap between the best model a company can only rent and the best one it can download and run itself is about 3.3 percent. At those distances, the supplier a company picks today has near-equals today, and anything expensive built around that one supplier, a custom integration, a long contract, is a liability the company chose to buy. Rent the intelligence. Refuse to marry the supplier.
What compounds is what no supplier can hand back.
The rented layer gets better every quarter and belongs to someone else. What a company owns is its accumulated knowledge: its customer and operational history, and the way its work actually gets done. No supplier holds that, and none can hand it back once it has been neglected. It also gains value every time the rented intelligence improves, because a stronger model is only as useful as what it is pointed at. An insurer that feeds a better model the same thin claims file gets a more fluent version of the same thin answer. The insurer that has kept its claims history and its policy interpretations in order gets a better decision. The model is the same in both cases. The asset is the difference.
Europe is running a different race.
Europe rents most of its cloud. AWS, Microsoft and Google hold about 70 percent of the European cloud market, and European-headquartered providers hold roughly 15 percent of their own home market. A European data centre run by an American provider answers the question of where the data sits and leaves open who can be ordered to hand it over, and procurement language routinely files the first answer as if it settled the second. That dependency is real.
Europe’s strength sits in a different event. It holds the chipmaking machines that leading-edge chip factories run on. By the IFR’s measure, Germany has the most automated factories of the three regions, and ABB, KUKA and Universal Robots are long-standing champions in certified industrial automation. The chips inside those machines are European too, from STMicroelectronics, Infineon and Axelera. Europe’s race is the machine that carries its own intelligence onto the factory floor under a safety certification the buyer already trusts. Model counts and cloud share score a different sport. On its own track, Europe has the machines and the chips in one place. That is a different event, and for a chief executive whose business moves physical things, it is the one that matters.
The Chief Executive keeps three decisions and delegates the rest.
The worksheet belongs to the technology team. The judgment stays upstairs, and it comes down to three decisions.
Rent the intelligence, and refuse to marry the supplier. The question to put to the team: how long would it take to swap our main model for a rival, and what would it cost? If the answer is a project, the company has married a supplier at the one layer where the alternatives sit within a few points. In the stacks I have mapped, the model was the only layer anyone had treated as a decision, and it is the cheapest of the five to change.
Own the knowledge, and carry it as an asset. The question: where does our accumulated knowledge live, who is responsible for it, and could a new model use it next quarter without a rebuild? If nobody owns the answer, the company is renting intelligence and pointing it at nothing in particular. Put the knowledge on the books the way a factory is on the books, with an owner and a budget.
Keep a second option open on whoever can switch you off. The question: which supplier could change our terms or cut us off tomorrow, and what would we run instead? For the cloud, the answer is a second provider or a contract that lets the company leave with its data. For the model, one answer is a model the company can download and run on its own machines. Mistral Large 3 is one. It cannot be withdrawn from the company running it. One failure, removed.
The robot table will be published again next year, and from the stands it will look like a scoreboard again. The company reading it is running its own race: renting intelligence that gets cheaper by the unit and dearer by the job, owning knowledge that compounds, and holding a second option at every point where a supplier or a government could cut the line. Whichever region the vendor comes from, that is the sport the company is in.
Before evaluating the winner, we need to understand what sport we are watching.
Victor Coimbra, Partner & CTO Americas, Artefact | victorcoimbra.life
