Whoever controls the infrastructure controls the outcome

I build artificial intelligence systems in production. I learned early something no architecture diagram shows: whoever owns the layer below decides what you can do in the layer above.

Whoever pays the cloud bill sets the ceiling of what can be built on it. This holds for the company that outsources inference, for the government department that rents a model, for the researcher tied to an API she cannot open. If the foundation belongs to someone else, your system runs by permission.

For a decade we treated the cloud as neutral infrastructure, a technical good with no political color. Swapping physical servers for elastic instances looked like pure engineering. It wasn't. It was a decision about who keeps the key to the room where the server is running.

When AI becomes basic decision infrastructure, and it is becoming that, depending on a closed model stops being technical debt and becomes a transfer of authority.

You send your most sensitive data to the other side of the API. Over there, someone holds the model weights, the training rules you cannot see, the silicon that does the math and the jurisdiction over what counts as a valid answer.

The fallacy of Model as a Service

SaaS traded control of the code for speed to market. That was livable: the code was yours, only the delivery belonged to someone else. Model as a Service goes deeper. It asks for control of the operation's logic.

Tomorrow the provider reclassifies your most ordinary transaction as a risk and blocks it. There is no bug ticket for that, because the rule is not yours. A monopoly on inference is a monopoly on the core of the business.

The state as hostage

Take a state finance department. Tax auditing lives on cross-checking huge volumes of structured data against revenue rules that change by state, by month, with every new law. When the model behind that audit is hosted and operated outside the country, the sovereignty of tax enforcement hangs on the latency and the terms of use of a foreign company.

And there is the perverse cycle. Every token sent carries context from the state into a private infrastructure, which trains, refines and rents it back to the state itself in the next cycle. Anyone who looks only at quarterly metrics misses the capture happening underneath the indicator.

The open source illusion

Then comes the usual counterargument: open models. Freely circulating weights would be the final proof against monopoly. It is fragile. Treating licensed code as the most valuable asset ignores that it sits on top of a mountain range of private silicon.

Independence at the logical layer through open software only moves the problem. Dependence drains downward, into the inference network and the raw energy cost of GPUs. Running sovereign matters little if the server bill breaks your balance sheet at the end of the quarter.

Why the cloud is worse than the railroad

Every great infrastructure concentrated power. Railroads, power grids and fiber optics all converged on a few owners, because laying track, raising poles and burying cable is expensive for every meter of territory.

The cloud is different, and worse. Track had to cross the map, beat the snow, lay sleepers with steel and dynamite. The cost of expanding was physical and slow. A data center's is not. Once built, it serves one more request at a marginal cost close to zero. The concentration the railroad took decades to build, the cloud builds in months, and the local cycle closes before anyone notices there was a contest.