Over the past twenty-four months, the European Union has pledged more public capital to AI infrastructure than in the entire prior decade combined. Its share of globally available AI-optimized accelerator capacity has, by most credible estimates, barely moved.
Margrethe Vestager — the EU's long-serving competition chief, now in the closing chapter of her tenure — used one of her final public interventions to call for a "balanced funding model" to close the compute gap with the United States and China. That single phrase is doing an enormous amount of work. Balanced between whom? Funded how? Delivered by when?
The item is thin: no mechanism, no number, no timeline. Which is precisely why it rewards reading. Policy language at the end of a term tells you what an institution already knows but will not yet say out loud. The headline is about money. The arithmetic is about settlement — where the money lands after it leaves the treasury, and whose balance sheet it credits.
Here's the tell: Europe is funding demand and importing supply.
Context
To grasp why Vestager's phrasing matters, you need the baseline she is quietly conceding.
Europe does not lack supercomputers. EuroHPC runs LUMI in Finland, Leonardo in Italy, MareNostrum 5 in Barcelona, and JUPITER in Germany — the bloc's first exascale-class machine. On paper, that is a respectable fleet. In practice, a public science instrument and a commercial AI training cluster are different animals. One runs on grants and queue allocations. The other runs on utilization rates, power contracts, and margin. Confusing the two is the single most expensive category error in European tech policy.
Then there is the money. InvestAI, unveiled in early 2025, targets roughly €200 billion in European AI investment, with about €20 billion earmarked for "AI gigafactories" — purpose-built facilities several times larger than current AI supercomputers. Layer on the €43 billion European Chips Act and the EuroHPC "AI Factories" amendment, which opens supercomputer access to startups and SMEs, and you have a genuine fiscal mobilization rather than a press release.
It is also, measured against the thing it claims to match, small. US private AI investment has run at roughly $60–70 billion in a single year, overwhelmingly venture- and hyperscaler-funded. China deploys state capital at a scale Europe's fiscal architecture was never designed to replicate. And that is before we get to Draghi.
Mario Draghi's September 2024 competitiveness report put the bloc's additional annual investment need at €750–800 billion — across all sectors, not merely AI — and named the diagnosis with unusual bluntness: Europe excels at invention and exports its scaling. Vestager's call is not a fresh idea. It is an inheritance, restated under time pressure.
What has changed is the framing. "Balanced funding" is a euphemism for a fight the Commission has never resolved internally: how much public money, how much private, and whether subsidies flow to national champions, to open competition, or — Vestager's historical instinct as an antitrust enforcer — to anyone except the giants.
One more signal worth flagging. This story surfaced via a crypto and Web3 vertical outlet, reporting a policy item with zero Web3 content. That is not random. It suggests readers in the decentralized-compute space are already bidding on the possibility that sovereign AI capacity becomes a market, and not only a public utility. Where an outlet's editorial anomaly sits, latent narratives usually follow.
Core
Run the arithmetic on where €20 billion of gigafactory capital actually settles.
A gigafactory is, in balance-sheet terms, a GPU order with a building attached. Europe produces no accelerator at NVIDIA's scale, no competitor at AMD's, and no merchant silicon ecosystem that can be substituted at the top of the stack. SiPearl, the French hope, has been late and still does not ship volume silicon at competitive performance-per-watt. Graphcore, a genuine architectural bet, was absorbed into SoftBank after failing to reach escape velocity. So if Europe buys the fastest available path to closing the gap, it wires the money to Santa Clara and Seoul. The sovereignty in "sovereign AI" is purchased, not built. That is not cynicism. It is a purchase order.
Then energy. European industrial electricity has traded at roughly two to three times US benchmarks through the recent cycle. Training and inference are power-conversion businesses with a thin intellectual margin layered on top. A datacenter in Virginia or Texas clears a return at power prices that leave a Frankfurt or Dublin operator underwater. If-then, stated plainly: if your input cost runs 2.5x and your revenue per FLOP is set by a global market, you do not have a compute business — you have a subsidy-dependent one. Public capital can cover that wedge for a while. It cannot cover it indefinitely without mutating into an operating subsidy, which is politically radioactive under the Stability and Growth Pact's deficit ceilings.
Third constraint: operators. Europe has no AWS, no Azure, no GCP, no Alibaba Cloud at the relevant scale. Gaia-X and the EuroStack discourse are attempts to will one into existence from the standards layer downward. Standards do not provision racks. Hyperscalers do. So Europe builds capacity and rents it back through American control planes — a settlement path that routes margin offshore even when the concrete is poured in Brandenburg and the power is drawn from a subsidized grid.
Now the leverage problem, which is where my day job intrudes.
In 2025 I spent several months working with legal-tech teams mapping MiCA-era arbitrage for cross-border payment firms — an eighteen-column matrix comparing compliance cost against liquidity access across seven jurisdictions. That exercise taught me a durable rule of capital behavior: capital does not move to where subsidies are announced; it moves to where exit liquidity exists. Europe's binding constraint is not the size of the check. It is the absence of a late-stage capital market that lets a founder convert a winning bet into a realized return without relocating to Delaware. Subsidize the entry, ignore the exit, and you fund a pipeline that terminates in an American acquisition — after the European taxpayer financed the R&D.
A parallel from my macro work is worth drawing. In 2022 I built a correlation model linking stablecoin dominance to global M2, and found that stablecoin inflows into emerging markets led local currency depreciation by roughly fourteen days. Official flows are lagging indicators. Private capital moves first and reports later. Apply that lens here and the issue sharpens: by the time a gigafactory breaks ground, the private capital that should have been co-investing has already priced the opportunity and chosen a different jurisdiction.
Historical precedent is unkind to the optimistic case. The Chips Act's €43 billion and the EuroHPC program have produced real infrastructure and contested commercialization. Public compute has a specific failure mode: built for ribbon-cuttings, underutilized for want of MLOps talent and paying customers, then defended on strategic grounds when the utilization numbers come in. When I tracked 500 autonomous AI trading agents across six months in 2026, the finding that unsettled me was not the herding — it was how quickly well-capitalized actors colonized any mispriced resource. Allocate public compute below market-clearing and you will discover, within two quarters, that the allocation has been farmed by the best-capitalized incumbents. The same dynamic that turned early DeFi liquidity mining into a bot harvest turns "AI Factories" access into a queue managed by whoever fields the sharpest procurement team. This is exactly why Vestager reached for the word "balanced." It is an antitrust instinct wearing a fiscal costume.
Read charitably, she is warning against subsidy capture. Read structurally, the word is a placeholder for mechanisms that do not yet exist: a common debt instrument, EIB guarantees, local-content conditions, or procurement commitments that convert state demand into private revenue certainty. Absent those, "balanced funding" is a slogan with a budget line attached.
Contrarian
Here is where I part company with almost everyone in this debate, including people whose diagnosis I share.
The consensus holds that Europe must close the frontier-training compute gap. I think that is the wrong target, and that pursuing it is the most reliable way to waste the money.
Frontier training is a winner-take-most market with brutal capital intensity and brutal iteration speed. Europe sits two full product cycles behind and is structurally slower to redeploy capital — fiscal rules, procurement cycles, and coalition politics all add latency. If-then: if parity in frontier training is the objective, the correct strategy is to lose slowly and expensively. There is no configuration of a balanced funding model that closes a two-cycle gap against counterparties spending more per quarter than your annual envelope. Pretending otherwise converts a strategic choice into a financial sinkhole.
The gap that actually matters is inference. Model efficiency — quantization, distillation, small specialized systems — is steadily collapsing the compute premium required to serve a given capability. Europe's comparative advantage was never raw FLOPs. It is regulation-as-product, since the AI Act is now the global reference text; industrial verticals with proprietary data and defensible distribution; and a payments and compliance stack that already knows how to settle cross-border value. The AI Act manufactures demand for compliant, localized AI services. That demand is real, near-term, and defensible. A gigafactory is none of those things.
Watch the second-order effect of the export-control regime as well. Europe is wedged between American hardware dependence and Chinese market access, so every procurement decision becomes geopolitical. Sovereign compute running on imported accelerators is sovereignty on license — revocable, denominated in dollars, and subject to someone else's foreign policy.
Takeaway
So watch the contracts, not the communiqués. Watch whether gigafactory capital carries local-content strings, whether EIB guarantees create genuine private risk-sharing rather than one-sided public risk-bearing, and whether the promised capacity actually arrives on schedule in 2027.
By then, ask the only question that settles the argument: if serving a given capability requires a fraction of the compute you budgeted, and the accelerator you bought was fabricated somewhere else — what, precisely, did you buy?
The answer won't be recorded in a white paper. It will be recorded in a settlement ledger. And ledgers, unlike policy briefs, do not do euphemism.
