Infrastructure Capture

Every major technology needs a foundation: roads for cars, power lines for electricity, data centers for artificial intelligence. Infrastructure capture means that a few companies control this foundation and everyone else has to rent it. Whoever owns the pipelines determines the prices, the rules, and who is even allowed to participate. The English term “capture” here means to “seize” or “take possession of.” It doesn’t refer to a single malicious act, but to a state of affairs that develops over years. In AI, this is discussed especially heatedly, because today already very few companies provide the underlying infrastructure.

Why a handful of data centers decide the entire market

AI systems aren’t built on a laptop. They require thousands of specialized chips, massive facilities, and enormous amounts of power and cooling. A single large training project can cost hundreds of millions of dollars. Only a handful of corporations can raise such sums. Everyone else – start-ups, universities, government agencies – rents computing time from exactly these corporations.

This shifts power. The provider sees which customers are growing and can raise prices once a customer can no longer switch away. It can also give preferential access to capacity to its own projects when chips are scarce. For a competitor, this is a double problem: they pay rent to the very company they are competing against.

It’s important to distinguish this from a classic monopoly. A monopoly means: one provider, one product, no alternative. Infrastructure capture can also exist with multiple providers, if they all sit on the same scarce foundation and no newcomer can enter. The market then looks lively from the outside, but hangs by very few threads.

How dependency builds up layer by layer

You can picture AI as a stack of several layers. At the bottom are the chips, above them the data centers, and above that the large foundation models. Foundation models are the elaborately trained base systems from which other companies build their apps. At the very top sit the finished products that users interact with. Whoever controls one of the lower layers has influence over everything above it.

This is reinforced by switching costs. Once a company has set up its data, its programs, and its workflows with one provider, moving away costs months of work. Experts call this lock-in, meaning an enclosure effect. Even as prices rise, the customer usually stays, because leaving would be more expensive than staying.

Another building block is equity stakes and in-kind payments. Large cloud providers invest in AI companies and pay out part of that investment not in cash but in computing time. The money thus flows back into their own house. Such arrangements are legal and economically sensible, but they further cement the stack. Incidentally, a common misconception is that only software matters: the real bottleneck is often power and grid connections.

The term in antitrust proceedings and quarterly earnings

In the news, this topic usually appears without the English technical term. When antitrust authorities in the EU or the US examine whether a cloud provider’s stake in an AI lab harms competition, this is exactly what’s at issue. Debates about chip export controls or about state-funded data centers in Europe also belong here.

For investors, the term is interesting because it explains why chipmakers and cloud companies often benefit more from the AI boom than the application companies do. Whoever sells the infrastructure earns money from every provider – regardless of which app ultimately succeeds. In analyses, you read sentences about “bottlenecks” or “chokepoint power.”

In everyday life, the consequences are felt indirectly. When the price of computing time rises, AI features in apps become more expensive or disappear from the free version. Schools, hospitals, and government agencies also feel it when their digital tools depend on a handful of foreign providers. This is exactly why some governments are demanding their own computing capacity, often under the banner of digital sovereignty.