Sovereignty Runs on Time-to-Power

Commentary on “How can Europe build sovereign AI computing capacity?”  ·  Bruegel

Every sovereign-AI debate eventually reaches for the same tools: local chips, local ownership, local-content rules. A European economist has just published the most useful thing said on the subject in a while, and its usefulness is that it points somewhere else entirely. Europe's sovereignty problem, on this diagnosis, is not a shortage of chips, capital, or resolve. It is a shortage of speed.

Writing for Bruegel, Bertin Martens projects that Europe's share of global AI infrastructure will barely exceed 5 per cent by 2031, against more than 70 per cent for the United States and 15 per cent for China. Capital is not the reason: more than 70 per cent of Europe's planned compute investment is already privately financed. The binding constraint, he argues, is operational speed and grid connectivity, the ability to permit a facility and connect it to power. His recommendations follow, and they are notable for what they reject. Rather than subsidise private ventures or mandate domestic chip production, he would put public money into anticipatory grid investment, faster connections, and "competitive time-to-permit grants" that reward the regions that move quickest. And rather than write local-content rules into the EU's proposed Cloud and AI Development Act, he would drop them, let US hyperscalers install more chips on EU soil, and settle for output-side signals: voluntary labels certifying that a service was computed in the EU, plus mandatory disclosure of what hardware is installed and what runs on it.

This is the series' thesis, arriving as European policy

Readers here will recognise the diagnosis, because it is the one this series keeps reaching. The constraint is time, not capital. The queue is the policy. When capital is abundant and the bottleneck is a connection date, the scarce asset is a permit and a place in the interconnection line, and whoever shortens that line decides who builds.

Martens has taken that observation and turned it into an allocation rule: stop paying developers to want to build, and start paying to make building fast. A competitive time-to-permit grant is the queue problem addressed head-on, funding aimed not at the project but at the speed of approving it. It is close to the most sensible thing a government can do once it accepts that the money was never the obstacle.

Protectionism is the wrong instrument, and he is right about that

The reflex in every sovereignty panic is to wall off inputs: build the chips here, own the stack here, require local content. Martens's argument against this is worth internalising. Input restrictions raise costs, reduce competition, and can suppress the very investment they are meant to attract, while doing little for the thing sovereignty is actually about. A local-content mandate that adds two years and thirty per cent to a project does not produce sovereign compute. It produces less compute, later.

This series made a version of the point when the equipment-dependency story broke: you cannot reshore a transformer by decree, and pretending otherwise only slows the build. Sovereignty pursued as protectionism tends to buy less of the thing it wanted.

But "computed in the EU" has a ceiling

Here the piece should be read with one eye open. Martens's pragmatism, welcome the hyperscalers' chips, label the output, require disclosure, is a real improvement on protectionism and honest about its trade-offs. But by the fuller definition of sovereignty this series has used, it is a floor, not the whole building.

A label certifying that a workload was computed on EU soil tells you where the electrons flowed. It does not tell you who owns the hardware, who can switch it off, or whose law prevails if the vendor changes the terms, the questions the OpenAI pause and the sovereign-stack debates put front and centre. A workload run in Frankfurt on a US hyperscaler's chips, under that hyperscaler's control, is more sovereign than one run in Virginia, and less sovereign than one run on infrastructure a European institution actually operates. Disclosure is the right first step precisely because it makes that gradient visible. It is not the same as closing it.

The number is a grid number

Strip the geopolitics and 5 per cent by 2031 is not a verdict on European ambition or European capital. It is a verdict on European interconnection queues and permitting timelines, the same machinery that in North America has produced multi-year waits, behind-the-meter gas, and moratoriums. Every jurisdiction that wants a larger share of this build is discovering the same thing in its own accent: the ceiling is not desire or money, it is the speed at which power and permits arrive.

The Canadian read

Canada should read this as a mirror, because it faces the same choice. The reflex will be to reach for the visible tools, ownership rules, content requirements, a national champion. The more effective move, on this analysis, is the invisible one: anticipatory grid investment, faster interconnection, and permitting a serious developer can plan around.

A country that gets time-to-power down will build sovereign capacity whether or not it legislates the word. A country that legislates the word and leaves the queue at three years will have neither the capacity nor the sovereignty, only the statute. And the fuller sovereignty, ownership, control, jurisdiction, accrues to whoever builds and operates the infrastructure, which is the reason to make sure the fast builders include some the country actually controls.

What this means going forward

The value of the Bruegel piece is that it moves the sovereign-AI conversation off the part that makes speeches and onto the part that determines outcomes. The chips and the ownership rules get the headlines. The connection date decides who has capacity in 2031. Europe's 5 per cent is a warning that you can have the capital, the ambition, and the legislation and still lose the build to a permitting office.

Sovereign compute is not first a question of who owns it. It is first a question of who can build and power it before the decade is out. Answer that, and the ownership question is worth having. Fail it, and the ownership question is academic.

VOLTEDGE

Reference: “How can Europe build sovereign AI computing capacity?” · Bruegel · read the article