In July, Sam Altman floated handing 5 per cent of each leading US AI lab’s equity to a public fund. Bernie Sanders went further with a bill claiming 50 per cent ownership of the AI giants for American citizens. Evgeny Morozov’s response in the FT is blunt: an AI sovereign wealth fund is not progressive, it is techno-imperialism, because the data centres and the power go abroad while the profits come home. Strip away the politics and you have a question every plant manager should recognise. Who captures the value?

That question matters long before Brussels writes any rules. When you sign a three-year contract for vision inspection or predictive maintenance, you are deciding who owns the data, the models, and the gains. Here is how to check before you sign.

Europe Hosts the Data Centres. America Books the Profit.

Morozov’s structural point is the useful one. Most of AI’s power goes into training models, and training can happen wherever electricity is cheap. The physical build goes where the grid allows. The profits stay where the equity sits.

For a plant in Eindhoven or Rotterdam, that is a description of your own supply chain. You supply grid capacity, industrial land, planning approval, and a recurring subscription line in your operating budget. What you do not get is a share of the asset those inputs create. Trump’s warning that companies will “be forced to go overseas” is not really a threat to Europe. It is a forecast.

Notice who is absent from the fairness argument. The entire fund debate concerns what American citizens are owed. European industrial buyers, the customers paying for the output, are not in the room.

What Altman, Sanders and Trump Are Actually Proposing

Morozov’s starting observation is worth keeping in view: a growing number of Americans dislike AI, so Washington is trying to change their minds the old-fashioned way, with a stick and a carrot. Three public positions, two instruments. It is worth reading them as mechanisms rather than slogans.

The carrot: equity stakes as a domestic political settlement

Altman’s July proposal is voluntary and modest in scale. Each leading US lab hands five per cent of its equity to a public fund. Nothing changes about how models get trained, where the compute sits, or how subscriptions are priced. It is a claim on future profit, not on control.

Sanders’ bill is legislation rather than a suggestion, and the figure is ten times larger: a national fund holding 50 per cent of the AI giants, with citizens entitled to a stake in their profits. Different magnitude, identical mechanism. Both route value through shareholding, and both treat the question as one of distribution inside a single country. Nobody outside the US holds a unit in either fund.

The stick: the China race argument and why Morozov calls it hollow

On Tuesday, with tech executives at his back, Trump told voters that if they refuse data centres, companies will “be forced to go overseas.” He added that “whoever wins superintelligence wins . . . you’re probably not gonna have a second place.” The implication is direct: block the build at home and you hand China the race.

Morozov’s response is that the stick, on inspection, is flimsy. The threat and the promise are not alternatives. As he puts it, they are “perfectly compatible”: US-aligned data centres can go overseas while AI profits still accrue to Americans.

That is the part worth underlining, because it removes the supposed trade-off entirely. There is no scenario in which hosting the infrastructure earns a seat at the dividend table. The build and the equity were never linked in the first place, and the AI sovereign wealth fund debate never pretended otherwise.

Why ‘Income at Home, Land and Power Abroad’ Is the Whole Argument

Morozov’s subtitle does the heavy lifting: “Accruing income at home from land and power abroad has an old name: empire.” That is not a rhetorical flourish. It is a description of how the cost base and the revenue base of frontier AI have been deliberately pulled apart.

The mechanics are dull and that is why they work. Land, water, substations, transformers, cooling and the local planning fight are all physical and all local. Equity is paper and sits wherever it was incorporated. Nothing in the technology requires those two things to be in the same jurisdiction, so they are not.

Accruing income at home from land and power abroad has an old name: empire

European operations leaders already live with the exported half of that arrangement. Grid connection queues that run for years. Industrial electricity prices that make the business case for a new line marginal. Permitting hearings where a hyperscale campus and the factory on the next plot are bidding for the same megawatts from the same distribution network.

None of that is framed as an AI cost. It arrives as a grid charge, a delayed connection date, a capacity cap on your electrification plan, or a council that has already allocated its headroom. You absorb it as an operating constraint. Someone else books it as capacity.

This is why the AI sovereign wealth fund debate is worth your attention even though it is an American political argument. The fund is a mechanism for returning rent to the people who host the cost. Europe hosting the cost without that mechanism is simply the same structure with the dividend missing.

Which brings the question down to your own contract. If the compute you depend on runs on a grid you are competing with, and the margin on that compute accrues somewhere you have no claim on, then your AI spend is funding an asset you will never own. That is a procurement problem, not a philosophy problem.

The Procurement Questions This Should Change on Monday Morning

If value accrues upstream by design, your job is to keep as much of it as possible downstream, inside your own four walls. That is a procurement discipline, not a political position. Trump told voters that “whoever wins superintelligence wins . . . you’re probably not gonna have a second place.” Fine. You are not competing for superintelligence. You are competing on scrap rate, first-pass yield and how many hours your quality team spends copying numbers between systems.

Contract terms that protect portability and data ownership

Settle three things before signature, not at renewal. First, data egress: you get your raw and labelled data out, in an open format, on demand, at no cost. Second, model portability: prompts, fine-tunes, embeddings and evaluation sets belong to you and are exportable. Third, no training rights on your process data unless you are paid for it.

Vendors will trade a discount for lock-in every time, because lock-in is worth more to them. Take the reverse trade. A 15 per cent price cut on a three-year contract you cannot exit is the most expensive line in your budget. An AI layer you can move between providers in a quarter is worth more than any discount on the table.

Which workloads genuinely belong on your own hardware

Frontier models are a rental. Accept that and rent them for what they are good at: document handling, drafting, summarising deviations, pulling answers out of a decade of CAPA reports. There is no strategic advantage in hosting those yourself.

Vision inspection, anomaly detection on sensor streams and anything inside a cycle-time budget belong on-premise or at the edge. Latency decides it, and so does IP. Your defect image library, your maintenance logs and your labelled failure modes are the one asset no vendor can reproduce. Keep them local, keep them yours, and price the business case on labour hours recovered and scrap removed. Not on access to next quarter’s model.

Where the Value Accrues Next, and How to Be on the Right Side of It

The equity-stake debate will run for years, and no European manufacturer has a vote in it. Washington will decide whether citizens get a share of OpenAI or Anthropic, and the answer will not change your scrap rate either way. What you control is narrower and far more useful: whether each AI deployment leaves behind something you own.

Two outcomes are possible from the same project budget. One produces cleaned process data, documented workflows, and a quality team that understands what the model is doing and why. The other produces a monthly invoice and a dependency you cannot price until renewal. The spend looks identical on the P&L for the first eighteen months.

The firms that capture value from AI in 2026 will be the ones treating frontier models as commodity inputs, not strategic partners. Commodity means substitutable. If your deviation triage runs on structured data, a defined prompt layer, and explicit acceptance criteria, swapping the underlying model is a procurement decision, not a replatforming project. Vendors hate this framing because partnership language is how switching costs get installed.

Morozov’s point about where income lands is worth carrying into your own planning. The physical and the financial were pulled apart deliberately upstream. Downstream, your job is the opposite: keep the asset and the operating knowledge in the same building.

Three things to watch over the next twelve months, because each one moves your cost base:

• EU compute capacity announcements: more regional inference capacity means real alternatives at contract renewal, not just a better negotiating bluff.
• Industrial electricity pricing: training follows cheap power, and if European industrial rates stay high relative to the US, the asymmetry Morozov describes gets wider, not narrower.
• Whether the equity-stake idea reaches legislation: Sanders introduced a bill. Bills mostly die. If one survives, the political cover for data centre expansion abroad changes, and so does your siting and sourcing calculus.

None of that requires a position on techno-imperialism. It requires knowing what you own when the contract ends.

Source: ft.com