Subsidising the Build, Losing on the Bill: Why Europe's AI Catch-Up Targets the Wrong Cost Line

In the week of 18 May 2026, the same article appeared in roughly a dozen places at once. CNBC ran it as "Why Europe's electricity prices threaten its AI ambitions." The World Economic Forum framed it as a race that "is becoming a race to power." The think-tank wave was wider: the European Council on Foreign Relations, the Brussels foreign-policy shop, called for an "emergency-level" fast-energy programme; the Kiel Institute for the World Economy, Germany's flagship macroeconomic research institute, described a "strategic gap"; and Bruegel, the Brussels-based economics think tank widely read inside the Commission, published "Europe needs a strategy to close the artificial intelligence compute gap." Even the trade press joined: OilPrice, the industry news outlet, ran the blunt version, "Europe Is Losing The AI Race as Energy Costs Soar."¹²³ The diagnosis was identical across all of them, and it was correct: European power is expensive, the AI build is going elsewhere, and the gap to the United States is widening.

The prescription was also identical, and that is the part worth stopping on. Build faster. Mobilise more public money. Cut permitting time. Stand up an emergency energy programme so the gigafactories have something to plug into. The consensus diagnoses an operating-cost problem and then reaches, almost reflexively, for a construction subsidy. That instinct is the inheritance of fifty years of industrial policy, and it is aimed at the wrong line of the income statement. The EU's two flagship instruments for this build, the EU Chips Act (the bloc's 2023 framework for European semiconductor manufacturing) and InvestAI (the €20 billion AI gigafactory programme announced in February 2025), are both capital-expenditure tools. The race is being lost on operating expenditure. A capex instrument cannot close an opex gap that compounds over a ten-year asset life, and no amount of building faster changes that arithmetic.

What the Consensus Gets Right

The gap is real and it is structural, not a 2022 spike that has since normalised. Bruegel's Policy Brief 01/2024 put industrial electricity prices in the EU at 158 percent above the United States in 2023, roughly two and a half times the US level, and its later work treats the gap as an established baseline rather than a transient shock.⁴⁵ The May 2026 snapshot that the consensus pieces quoted tells the same story in monthly terms: industrial electricity at roughly $89 per MWh in Germany and $112 in the United Kingdom against about $28 in the United States, with France an outlier at $44 on the strength of its nuclear fleet.¹

For the workload that actually matters here, a flagship AI cluster signing a multi-year off-take contract, the honest comparison is narrower than the grid-average headlines but still wide. German industrial electricity runs at roughly €105 per MWh for the energy-intensive tier that qualifies for state relief on carbon costs and grid fees, and about €168 per MWh for the unreduced rate that applies to most consumers, on 2024 modelling by the German industry association BDEW.⁶ A US hyperscaler procuring through a long-term power-purchase agreement pays something closer to €35 to €65 per MWh: on the Texas grid (ERCOT), solar PPAs clear around $35 to $45 per MWh; on PJM, the mid-Atlantic grid covering Northern Virginia's data-center belt, PPAs clear around $65 to $75.⁷⁸ Call it a two-to-three-times gap on the cost line that dominates the asset's life. The Nordic alternative, where large-consumer wind PPAs in Sweden and Finland close at €40 to €60 per MWh, narrows that to roughly one-to-1.4-times, an interesting number the consensus coverage tends to mention and then walk past.⁹

Capex Money for an Opex Problem

Here is where the diagnosis and the prescription come apart. The European Chips Act mobilises up to €43 billion of public and private investment through 2030, of which the directly committed EU and Member State public money is about €8.6 billion, the rest leveraged private co-investment and anticipated IPCEI (Important Projects of Common European Interest) spending.¹⁰ InvestAI, announced by Commission President Ursula von der Leyen at the Paris AI Action Summit on 13 February 2025, adds €20 billion specifically for up to five AI gigafactories, each specced for roughly 100,000 advanced AI chips, on a 70 percent private, 30 percent public funding split.¹¹ In the program's own design language, the InvestAI mechanism covers "up to 35 percent of capital expenditure" for approved projects.¹² That phrasing is significant. Every euro in these instruments is a one-time grant or loan against the cost of building. Not one of them is a recurring subsidy against the cost of running.

Now put the operating cost next to the subsidy. Take a 100,000-GPU cluster built on Nvidia's H100, the workhorse AI training chip of the 2024-2026 build-out. The closest real-world reference is xAI's Memphis deployment, branded Colossus: the 100,000-H100 supercomputer Elon Musk's xAI stood up in 2024, drawing about 150 MW continuous including cooling and overhead.¹³ Run at scale, that is roughly 1.17 million MWh per year. At the German unreduced rate the annual electricity bill is about €198 million; at the energy-intensive tier about €122 million; on an ERCOT PPA about €42 million. The delta against Texas is therefore between €80 million and €156 million every year, or €800 million to €1.56 billion over a ten-year asset life, per single cluster.¹³ The public subsidy per InvestAI gigafactory, on the €20 billion across five sites at a 70-30 split, is about €1.2 billion. That subsidy is paid once; the disadvantage is paid every year. At the unreduced German rate the cumulative electricity gap overtakes the entire public subsidy at around year eight, before the cluster has finished a single ten-year asset life, and at the energy-intensive relief rate the subsidy covers roughly one asset-life of the gap and no more. Either way the grant neutralises at most one generation of the disadvantage on an asset that refreshes every few years, while the operating gap recurs for the next cluster and the one after that. A capex instrument can pre-pay a slice of an opex problem once. It cannot stop it from recurring.

This is why the 1970s toolkit misfires. A one-time construction grant works for a steel mill or a chip fab because those are capex-dominated assets: the dominant cost is incurred once, at the build, and a subsidy that defrays the build changes the project's lifetime economics. AI compute is the first industrial workload in decades whose lifetime cost is dominated by opex, and specifically by electricity. Inference, not training, accounts for an estimated 80 to 90 percent of total AI energy consumption, and inference is a continuous, 24-hour load that scales with traffic and never falls to zero.¹⁴ That is the textbook profile of an opex-dominated industrial process, and it is precisely the cost structure a capex grant cannot touch.

The cost-structure difference shows up directly in the size of what each region actually builds. The largest AI training facility on the public European record is the Paris cluster announced in December 2025 by Mistral, France's flagship AI lab: 13,800 GPUs drawing 44 MW. The American equivalent is the Stargate site in Abilene, Texas, the centerpiece of the $500 billion AI-infrastructure consortium of OpenAI, Oracle and SoftBank announced in January 2025: 450,000 GPUs drawing 1.2 GW. The ratio is roughly one to twenty-seven.¹³ The European facility is not a smaller version of the American one; it is a different category of object, and the difference is downstream of the cost of the electricity each one has to buy for a decade.

Magdeburg as the Tell

The arithmetic is abstract until a project cancels on it. One already has, and it makes a useful case study because everything else about the deal had already been agreed. Intel announced its Magdeburg "mega-fab," a planned chip plant in Saxony-Anhalt, in March 2022; by June 2023 the German subsidy had been negotiated up to €9.9 billion, close to a third of the roughly €30 billion project cost. Construction was postponed for two years in September 2024, and the German and Polish projects were cancelled outright by CEO Lip-Bu Tan in July 2025.¹⁵ Intel's official reasons were internal: a $1.6 billion quarterly loss, a $10 billion company-wide cost-reduction plan, and soft demand in the foundry segment the fab was built to serve. No Intel statement reviewed cited German electricity prices.

So the opex reading is an inference, and it should be labelled as one. What Intel did not say, but what any CFO comparing a 100-to-200 MW German fab against an Arizona alternative would have computed, is that the lifetime electricity differential at German industrial rates is the same order of magnitude as the entire German subsidy package. A subsidy worth a third of project capex is the kind of intervention only required when the operating economics do not carry the project on their own. Magdeburg does not prove that every European fab fails; TSMC's Dresden plant broke ground in 2024 and is on schedule. It proves something narrower and more useful: a capex subsidy, however large, gets cancelled when the operating math turns, because the subsidy was never addressing the operating math.

The Captive-Demand Trap

Magdeburg was the case where the project could walk away, because no one had to build it in Europe. The standard reassurance against any broader pessimism is the other side of the ledger: the projects that cannot walk away are happening anyway. The hyperscalers are committing in Germany regardless, so, the argument runs, the cost disadvantage cannot be binding. Microsoft committed €3.2 billion in February 2024 to double its German capacity, AWS has pledged €8.8 billion to the Frankfurt region, and Frankfurt's operational IT load is on a path toward 3.3 GW by 2029.¹⁶ The investment is real. What it is not is evidence that the economics work, because most of it is not competing for the cheapest electricity. It is compliance-driven capacity: inference, cloud and colocation for the European market, pinned to European soil by a thickening stack of data-residency rules: the GDPR for personal data, the NIS2 cybersecurity directive, DORA (the Digital Operational Resilience Act, governing financial services) and, from August 2026 for high-risk systems, the AI Act.¹⁷ These workloads have to run in the EU regardless of what the power costs, which is precisely the problem.

Guaranteed demand removes the price discipline that would otherwise push a workload to where energy is cheap. A hyperscaler serving regulated European data has no choice but to absorb the German or Irish power premium, and a cost that cannot be avoided is a cost that gets passed on. The premium is embedded in the price of every sovereign-cloud contract and every compliant inference call, and it is paid, in the end, by the European banks, hospitals, insurers and manufacturers that consume those services. That is not a niche line item. It is an input-cost tax on the entire downstream economy, and it is the same mechanism that Mario Draghi, the former ECB president and Italian prime minister, identified in his September 2024 EU competitiveness report when he put EU electricity prices at two to three times US and Chinese levels and named energy cost as a central drag on European competitiveness.¹⁸ The regulation that guarantees the demand also guarantees that the premium is never disciplined away. Captive demand is not a sign that the cost is bearable. It is the channel through which the cost spreads.

The Frontier Is Moving Toward Efficiency

If the diagnosis stopped at captive demand it would be a counsel of despair: an opex gap the subsidies cannot reach, channeled by regulation into a premium on every European data load, with the twenty-seven-to-one scale gap as the visible monument to how far behind a capex-only strategy leaves the region. The harder and more interesting question is whether there is a positive move from the same arithmetic. There is, and it starts from accepting that the scale gap is not the frontier worth chasing.

The scale gap itself is not one Europe can close by spending, and no subsidy moves it. But there is a more hopeful reading, and it starts from the same arithmetic. If electricity is the binding constraint, then the competitive frontier is not scale, it is capability per watt, and the evidence is that the industry is being pushed toward that frontier anyway. The scaling laws that drove the 2010s are showing diminishing returns: each doubling of pre-training compute now buys a smaller increment of capability than the last, and the gains have increasingly come from inference-time reasoning rather than from larger pre-trained models. Ilya Sutskever, the OpenAI co-founder who has been one of the loudest in-house voices for scaling, put it at NeurIPS, the field's main machine-learning conference, in December 2024, in his own words: "the 2010s were the age of scaling, now we're back in the age of wonder and discovery."¹⁹ A growing research literature, including work explicitly titled around "the race to efficiency," now reframes the scaling question as a question about cost per unit of capability rather than raw parameter count.¹⁹

DeepSeek, the Chinese AI lab whose V3 release in early 2025 stunned the industry with frontier-level performance at a fraction of the expected cost, is the proof of concept that questioning the standard pays. The V3 model was trained at 8-bit floating point precision (FP8, half the standard 16-bit format that frontier labs had treated as the floor) on a cluster of H800 chips, the deliberately throttled variant Nvidia ships to China under US export controls. Hardware-software co-design and fine-grained quantisation held accuracy at FP8 without loss, an approach DeepSeek's authors describe as levelling the playing field for teams without access to the most powerful hardware.²⁰ The lesson is not that DeepSeek is European. It is that the conventions everyone treated as fixed, full numerical precision and the biggest available chip, were not fixed, and that the meaningful cost reductions came from changing them rather than from buying more.

This is the part of the map where Europe has a record. The region's notable contributions to computing have come not from outspending the field but from setting a standard others adopted: GSM, developed by European telecoms standards bodies between 1982 and 1990, became the blueprint for global mobile communications and made European manufacturers world leaders for a decade; MP3 came out of Germany's Fraunhofer Institute and became the default way the world stored audio for two more.²¹ Neither was a brute-force capital play. Both were constraint-driven engineering that hardened into a standard. An energy-constrained region has a particularly clear incentive to invent the energy-efficient frontier, because it is the region paying a premium for every wasted watt.

Getting Creative Means Getting Creative About Opex

The May 2026 consensus pieces all closed with versions of "Europe needs to catch up." If "the train has left" is the right way to read what they were describing, then the EU is running after it on the wrong platform. Catching up does not mean building faster on the same cost base. It means attacking the operating cost directly. Two narrow patches help at the margin, and Europe should use both. The move with leverage sits on a different cost-line entirely.

The first patch is geography. A pure-training facility in Sweden or Finland pays within 30 to 40 percent of the Texas rate, close enough that capex differences dominate again and the opex gap stops being decisive.⁹ InvestAI siting is Member-State-driven, which actively invites Germany and France to compete for sites their cost structures cannot justify; the easy creative move is to mandate Nordic or nuclear-French siting for any pure-training gigafactory and leave compliance-driven inference capacity to the distributed model where data-residency rules already pay the premium.

The second patch is classification. Germany's new Industrieller Strompreis, in force since 1 January 2026, cuts power for qualifying energy-intensive industry to roughly 5 cents per kWh, about €50 per MWh, which would put a German site within striking distance of the Nordic range.²² Data centres do not currently count as "energy-intensive industry" under the standard EU state-aid framework, so the relief that could close most of the German gap is sitting unused for the sector that needs it most. Extending that classification, if EU state-aid approval holds, is an opex instrument hiding in plain sight.

Both patches help at the margin. Neither changes the underlying race, because the underlying race is being run on a cost base Europe cannot make competitive at the megawatt level. The third move is the one with leverage, and the one that turns the constraint into an advantage rather than a handicap to manage. Rather than chase a training-scale race it cannot finance on its cost base, Europe can target the efficiency layer the rest of the industry is migrating toward anyway: low-precision training and inference, model architectures that do more per FLOP, hardware-software co-design, and the energy-per-token conventions that frontier labs will eventually have to adopt. This is the "out-engineer rather than outspend" thesis, and the point is not that it is a consolation prize. If capability per watt is where the returns are migrating, and the prior section's argument is that they are, then setting that standard is an attractive catch-up move, and one a region paying a premium for every wasted watt is structurally motivated to make. It is also, conspicuously, not what the gigafactory program is funding. The program is still buying scale by the megawatt. A region that has set global standards before, on phones and on audio, has the precedent and the constraint to do it again on watts per token. The capex program does not even ask the question.

Where This Could Be Wrong

One serious objection runs against the hopeful half of the argument. Efficiency gains, the sceptic says, get competed away: when compute gets cheaper per unit of capability, the industry simply uses more of it, a Jevons-paradox dynamic in which a more efficient frontier raises total energy demand rather than rewarding the efficient. And setting a standard is not the same as capturing its rents; Europe defined GSM and MP3 and still watched most of the downstream value accrue elsewhere. Both points are fair. The response is that they argue for sharper execution, not a different target: in a world where energy is the binding input, the region that sets the efficiency standard at least competes on the axis where its constraint is an incentive, rather than on the axis, raw capital deployment, where it is structurally behind. Standard-setting without value capture is a real failure mode, and one Europe should be determined not to repeat.

A second objection is that regulation is itself the moat: a sovereignty premium that offsets the cost disadvantage, with the proposed Cloud and AI Development Act (the Commission's planned framework for trusted European cloud and AI infrastructure) and the existing AI Act turning compliance into a non-substitutable category. This is real as a market, but it sits downstream of frontier capability. A sovereign cloud either runs sovereign frontier models, which requires exactly the training-compute investment the opex math says European economics struggle to support, or it runs US and Chinese models under European regulatory wrappers, which is a compliance business rather than a sovereign technology stack. It is a legitimate objective. It is not a competitiveness argument, and it should not be sold as one.

A third objection is that prices will fall and the disadvantage is temporary. Bruegel itself lists four mechanisms that could narrow the gap, more renewables, better transmission, targeted subsidies, demand-side flexibility, and does not project the gap closing meaningfully before 2030.⁵ On a ten-year asset life, a disadvantage that persists most of the decade is not temporary in any sense that matters to the financing decision. And a fair version of the argument concedes that capex is also higher in Europe, on land, labour and permitting; the reason to focus on opex is not that capex is identical but that capex disadvantages are at least addressed by the subsidies, while the opex disadvantage that dominates lifetime cost is not addressed at all. The asymmetry between what is subsidised and what dominates is the whole claim.

Personal View

The body above is analytical. What follows is my own view, kept brief.

In my view the mistake here is not the subsidy. Funding the build is good policy; an industrial base needs the capital behind it, and the EU Chips Act and InvestAI are right to put public money where private money alone will not go early enough. The mistake is treating capex as the whole answer and letting opex drop out of the frame. A grant that pays for the building while the electricity bill runs uncovered for a decade is half a policy, and it is the expensive half to get wrong, because the operating cost is the part that compounds. Fund the capex, by all means. Just do not forget that the bill arrives every year after the ribbon is cut, and design for that bill from the start rather than discovering it on the operator's P&L.

The second thing I have come to think is that running after the leader is rarely the clever move, and almost never the clever move when you are running on a more expensive track. You do not win a race you entered late by sprinting harder on the same course; you change where the finish line sits. The same problem, the cost of intelligence per watt, can be attacked from angles the incumbents have little reason to explore: efficiency rather than scale, siting rather than subsidy, the standard that makes everyone's compute cheaper rather than the cluster that makes only yours bigger. Constraint has a way of forcing the cleverer solution, and Europe has both more constraint and more history of clever solutions than the lament gives it credit for. The honest version of "getting creative", in my view, is not catching the train. It is asking whether that particular train was going somewhere worth being, and then finding a faster way there.

END · 06