ECB: in Europe, AI capex is paid for with cash

The precedent: when investment is paid for in-house

In West Germany in the 1950s, industrial reconstruction was paid for largely out of retained earnings. The mechanism was simple: high margins, low dividends, a tax system friendly to reinvestment. Banks were useful. The corporate balance sheet was more useful.

In Japan in the decade after 1990 the same pattern returned, inverted. Firms used cash to cut debt and covered investment with their own resources, while the central bank held the cost of money near zero.

In both cases the monetary authority discovered the same thing. Rates move spending when spending runs through credit. Two precedents sketch a rule; the third turns it into a pattern.

The third case is now, in Europe, and it concerns the technology that will absorb capital for the rest of the decade.

The number out of Frankfurt

On 2 October 2026 the ECB blog published the first broad measure of how European firms intend to pay for artificial intelligence. The source is the Survey on the Access to Finance of Enterprises: around 5,000 euro area firms, questions covering the twelve months ahead.

72% of those planning an AI investment are counting on internal funds^[1], current cash and retained earnings. The authors are Annalisa Ferrando, Sara Lamboglia, Judit Rariga and Maurice Schmidt.

The rest of the picture confirms the direction. 49% allocate the budget to technology and tools, 46% to staff training, 40% to data infrastructure, 12% to hiring specialists.

38% ruled out every line.

Large firms invest across all categories more than small ones, with the same hierarchy of priorities. The financial press picked up the number as a curiosity about credit (Crypto Briefing^[2]). It is in fact a piece of monetary policy information.

The mechanism: transmission loses a channel

Monetary policy works through several channels; the most studied remains bank lending. The central bank raises rates and the cost of borrowing rises. Marginal projects fall away and demand slows. The model assumes investment originates on a bank balance sheet.

When three firms in four pay with their own resources, that assumption weakens. The relevant cost becomes the alternative return on cash: the short-dated government bond, the deposit, the share buyback.

The brake stays live, with a different lag and a different force. A hike makes cash more precious and slows discretionary spending. A cut, conversely, does little to those who have already decided to pay in-house.

We are beyond the cycle here. This is a regime change in transmission, and it holds on a decade-long scale. The credit channel retreats as investment shifts toward the intangible.

The transatlantic divergence

In the United States the AI machine runs on debt. Chipmakers, labs and new-generation cloud operators pledge assets, issue bonds and draw on private credit to build data centres.

In Europe the same capacity is being built with the cash of the firms that will use it. Two different financial architectures produce two different risk profiles. The market still prices the second as if it were the first.

The divergence between American speed and European solidity always resolves. The question is how. In debt capitalism, an estimation error becomes credit stress, with spreads widening. In cash capitalism, the same error becomes a loss of earnings, absorbed by shareholders.

The first regime runs faster on the way up. The second holds better on the way down. The British railway crisis of 1847 and the European Monetary System crisis of 1992 tell the same lesson: long-term infrastructure paid for with short-term money exposes the balance sheet at the moment of refinancing.

Who can afford artificial intelligence

The distributional corollary is harsh. Under a self-funding regime, AI capacity becomes a function of balance sheets that are already rich. The ranking of European firms crystallises around those that have accumulated cash over the past five years.

The 12% planning to hire specialists is the most telling figure in the series. Buying a tool costs one budget line. Building internal expertise costs years of the income statement, and the continent has chosen to train the staff it already has.

This reduces the need for credit and lowers the ceiling of ambition.

An SME with thin margins will stay outside the adoption cycle. Its exclusion runs through cash, far from the bank counter, so it stays invisible to credit statistics. A support policy built on the cost of borrowing will miss the target.

The position, and what would refute it

My position is explicit: in Europe, AI capex forms outside the credit channel. The continent's technological capacity becomes a function of balance sheets rather than of the cost of money.

Three facts would change it. First: a later edition of the SAFE with the internal funds share falling toward 60%. Second: a wave of European bonds issued for data centres. Third: a public guarantee programme that shifts AI investment onto bank balance sheets.

The third is the most likely, and it should be looked for in the documents of the European Investment Bank.

Anyone chairing a board should read the figure as a balance sheet constraint rather than a sector data point. Anyone overseeing risk should add a scenario in which a rate cut leaves technology spending exactly where it is. Anyone speaking to investors should revisit the narrative linking falling rates to rising investment: in eighteen months it could prove wrong.

Three implications for capital:

1. Reallocate toward cash-rich firms, 36-month horizon

A family office seeking exposure to European AI should look for it in the deep balance sheets of manufacturing and services rather than among infrastructure suppliers. Value accumulates with those who adopt and generate cash. Selection rewards those able to self-fund for several years in a row.

2. Price American and European AI risk separately, 24-month horizon

A model treating AI as a single factor gets the continent wrong. American exposure reacts to private credit spreads; European exposure reacts to operating margins. These are two distinct factors and they call for two distinct hedges.

3. Revisit the rate elasticity of investment in VAR models, 18-month horizon

Models calibrated on historical variance link investment and rates with a coefficient estimated on a physical capital economy. Intangible investment paid for in-house responds less. The model will keep forecasting a recovery that arrives late.

The forecast

Expected event: the next edition of the SAFE measuring AI investment plans. Publication by 31 December 2027. The share of euro area firms counting on internal funds will remain at or above 70%.

Confidence: 75 out of 100. Horizon: 31 December 2027. Verification: the table on expected funding sources in the ECB SAFE report.

Kill signal: a SAFE edition published by that date recording the share below 60%.

What to watch:

Three indicators will signal early how this ends.

• The volume of European bonds issued to finance data centres
• The share of SMEs stating they are deferring adoption because of a cash constraint
• Any ECB document treating AI self-funding as a monetary transmission issue