Europe's AI Bill: A 19 GW Gap and €600 Billion

The number that matters is $38 billion per gigawatt

The figure that decides Europe's industrial fate is $38 billion per gigawatt. Almost everyone is arguing about €600 billion instead.

On 14 September 2026, in Vienna, ECB President Christine Lagarde laid out Europe's arithmetic on artificial intelligence in a speech that stands as a primary source^[1]. The Commission's impact assessment for the Cloud and AI Development Act, SWD(2026) 502, points to a data centre capacity gap widening from 3 gigawatts in 2025 to roughly 19-20 gigawatts in 2036. Epoch AI, on 14 May 2026, measures the capital required for a one-gigawatt AI data centre at around $38 billion, servers included.

Nineteen gigawatts at $38 billion each comes to roughly $720 billion, in other words close to €600 billion. The speech presents that sum as an upper bound, in so many words.

The consensus has the wrong frame

The debate reads the €600 billion as an unaffordable invoice, and therefore as a verdict. It is in fact the product of two moving quantities: a capacity expected in 2036 and a unit price measured in May 2026. The price per gigawatt remains the variable that truly counts, because it changes every quarter.

A trajectory needs at least three points with a date and a source attached. Here they are, and they should be read together.

• Missing capacity in Europe: 3 GW in 2025 (Commission impact assessment, SWD(2026) 502).
• Missing capacity in 2036: roughly 19-20 GW (same source).
• Price of one AI gigawatt, servers included: around $38 billion (Epoch AI, 14 May 2026).
• Total bill: roughly $720 billion, about €600 billion, flagged as a ceiling (ECB, 14 September 2026).

Eleven years separate the two ends of the capacity range: sixteen gigawatts to be built, about one and a half gigawatts a year, simple arithmetic on the source's two data points. That pace makes the project industrial, and therefore governable. The spending arrives spread out, and the unit price that governs it moves along the way.

Anyone signing multi-year contracts today is buying this cycle's scarcity at this cycle's price. The bottleneck always migrates: from silicon to packaging, from memory to power, from compute to human judgement.

European capital already exists, with an American address

Euro area households hold around €440 billion in US technology equities.

European savings are therefore funding the technology of the moment, just as they did during Austria's Gründerzeit. The Vienna speech recalls that French savers paid for the Suez Canal, while larger British and German sums went into American railways. In 1873 the Vienna stock exchange collapsed, Wall Street followed four months later, and what that money had built stayed in the United States.

Then comes the part that weighs most: the growth stayed there too. By the end of the century the American economy was the world's largest, while Europe was living through two slow decades.

The second channel is more recent. American hyperscalers have issued more than $100 billion in bonds over the past year and account for almost a tenth of new euro-denominated issuance by European non-financial corporates. The continent's capital market is already funding servers, in Seattle and in Virginia.

European demand is already running

Europe's scarcity is on the supply side of compute. Demand is moving fast.

Euro area firms will devote around 10% of total investment to AI in 2026, according to the figures presented in Vienna. AI-linked credit accounted for roughly a quarter of the growth in corporate lending in the first quarter of this year. The share of euro area workers using AI at work has doubled in two years and now exceeds 50%.

The historical comparison makes the speed visible: the internet took around a decade to reach that threshold. This is an adoption curve that jumps, rather than climbing gently.

That changes the nature of the problem. A continent that adopts quickly and builds slowly imports capacity, pays in foreign currency, and leaves the margin to whoever owns the servers.

Cliff event: the instrument that turns savings into gigawatts

The jump comes when a public instrument converts domestic savings into domestic compute.

The first IPCEI dedicated to cloud and artificial intelligence is that instrument. An Important Project of Common European Interest works like a long contract between member states, the Commission and companies: it coordinates aid, sets capacity commitments and makes demand predictable. For those doing the building, predictability is worth as much as capital.

Adoption, in cases like this, climbs in steps. Europe's step coincides with the signing of the first common project, because from that moment the price per gigawatt becomes a public, verifiable budget line. Before that signature, the continent buys capacity elsewhere, at the prices of the cycle under way.

Three categories that will change shape by 2030

A thesis holds up when it names who changes shape. Three European categories come out of this curve looking different.

• Data centre and colocation operators: from sellers of space to sellers of power with the compute thrown in.
• Utilities and power generators: the price per kilowatt-hour becomes the real price of compute.
• Technology procurement teams: contract duration becomes the main risk.

European operators will watch the bill shift from the building to the power socket, because energy cost dominates a facility's life cycle. Utilities will decide where the gigawatt is born, and will therefore negotiate from strength with anyone arriving with a ten-year plan. Procurement teams will discover that a five-year cloud contract signed in 2026 locks in today's scarcity price for the whole period.

This is a regime change, not a seasonal fashion. The right question to put to a supplier concerns how the price is indexed to the curve, rather than the depth of the opening discount.

My position, and what would change my mind

My position: Europe should index its AI plan to the price per gigawatt and the price per kilowatt-hour, treat the €600 billion as a theoretical ceiling destined to fall, and fund it with the stock of savings already exposed to American technology. The money exists; the address needs changing.

The reasoning rests on three facts cited in Vienna: €440 billion of European household money already invested across the Atlantic, a capacity gap measured by the Commission, a unit price measured by Epoch AI. A stock of savings, a physical requirement and a price together make an industrial plan. What is missing is the instrument that connects them.

What would change my mind: a price per gigawatt that holds steady or rises for six consecutive quarters. At that point the €600 billion ceiling would become a floor, and buying foreign capacity would remain the rational choice.

Prediction, horizon, kill signal

Prediction: by 31 December 2027 the European Union launches the first IPCEI dedicated to cloud and artificial intelligence, with announced public funding below one tenth of the €600 billion given as the upper bound.

Confidence: 70 out of 100. Horizon: 466 days from today.

Kill signal: the announcement, by that same date, of a cloud and AI IPCEI with public funding of €60 billion or more.