Europe's AI Sovereignty Layer

The accepted defeat

Europe has been told it lost the AI race.

By the visible scoreboard, that judgment is hard to dispute. The frontier models are mostly American. The hyperscale clouds are American and Chinese. The largest pools of AI venture capital sit outside Europe, with the United States taking roughly 66% of global AI venture funding between 2023 and mid-2025 against Europe's ~12% (CEPS). Last year American labs produced the majority of the world's notable AI models while Europe produced a handful (Stanford AI Index). More than 90% of the world's AI data-centre capacity sits in the US and China (Oxford). To take the measure of the gap, the US "Magnificent Seven" tech firms are worth on the order of $17tn, roughly the annual GDP of the European Union (General Catalyst). The companies setting the global AI narrative are not European.

That has hardened into a familiar conclusion: Europe will regulate AI, adopt AI, and subsidise AI, but it will not own the core value of AI. It is a reasonable reading of the evidence, and this report does not pretend otherwise. Europe is structurally disadvantaged in the race everyone is watching.

This report argues only that the conclusion is incomplete, because it assumes there is only one race.

The wrong scoreboard

The scoreboard the market is using measures a single contest: the race to build larger models, train them on more compute, distribute them through consumer and enterprise software, and turn them into the default interface for digital work. Models, chips, cloud, capital, developer mindshare. On that board, the winners are already clear, and they are not in Europe.

But a scoreboard is a choice about what to count. This one counts the visible layer of AI, the layer that faces the user, generates the headlines, and attracts the venture capital. It does not yet count the layer where AI meets the physical economy, because that layer is early, unglamorous, and largely private. The question worth asking is not who is winning the race on the board. It is whether the board is measuring the race that will matter most.

A second race is beginning underneath the first.

The second race

This race is not about which model can write better text or answer more questions. It is about which intelligence layer can operate the physical economy: factories, power grids, robots, logistics systems, industrial equipment, defence platforms, and regulated infrastructure. It is AI that does not talk, but acts, that optimises production, predicts equipment failure, inspects hazardous sites, balances a strained grid, and guides an autonomous system under contested conditions.

The bottlenecks in this race are different from the ones in the model race. In consumer AI, the scarce inputs are compute, model talent, and distribution. In industrial AI, the model matters but is not enough. What decides the winner is the ability to connect intelligence to machines: contextualised operational data, certified hardware, safety-critical deployment, customer trust, energy access, procurement relationships, and the ability to embed intelligence inside real-world workflows where a wrong answer breaks something physical. The winning company is rarely the one with the best demonstration. It is the one embedded deeply enough in operations that removing it would stop the plant.

Those bottlenecks are the plot. And they are the reason Europe's position looks different in this race than in the last one.

Europe's hidden substrate

Here is the inversion at the centre of this thesis: the same industrial base that made Europe look slow in the software era may become strategic in the industrial-AI era.

Europe was a laggard in consumer software because the value there accrued to platforms, network effects, and scale, none of which favoured a fragmented continent of industrial incumbents. But industrial intelligence accrues value to different things, and Europe holds an unusual concentration of them. It has the deepest manufacturing base in the developed world (manufacturing is ~15.9% of EU gross value added, ~19.9% in Germany, per Eurostat/Destatis). It has the densest robot fleets on earth (Western Europe leads at 267 robots per 10,000 manufacturing workers; Germany at 449, and Europe installed 17% of the world's industrial robots in 2023 against 11% in the Americas, per IFR). It has industrial champions and a Mittelstand of hidden global segment leaders, decades of proprietary operational data locked inside those firms, the most complex energy system in the world to manage, regulated and safety-critical environments by the thousand, an acute and newly funded defence urgency, and customers who, precisely because they are regulated, need trust, compliance, and local deployment rather than a public API.

These assets are not glamorous, not consumer-facing, and nothing like the model race. That is exactly why the market discounts them. But they are the terrain through which the next phase of AI must pass. Europe's industrial base is not a legacy weakness if AI value moves into the physical economy. It becomes the ground the winners have to stand on.

The non-obvious insight

The obvious AI story is models, chips, cloud, and frontier labs. The deeper industrial-intelligence story is operational data, certified deployment, physical execution, sovereign infrastructure, power, procurement, and workflow ownership.

The non-obvious claim is about durability. The visible AI layer may commoditise faster than almost anyone building it expects, as frontier capability converges, open weights chase the leaders, and inference costs collapse. A model that is a moat today can be a feature in eighteen months. The physical control layer commoditises far more slowly, because it is embedded in operations, regulated, safety-critical, data-rich, local, operationally complex, and expensive to rip out. You can swap a model with an API call. You cannot swap the contextualised data foundation running a refinery, the explosion-proof robot certified for a live gas plant, or the autonomy software a ministry has accredited for a weapon, not without risk, downtime, and years.

Value tends to concentrate where replacement is hardest. In the physical economy, that is not the model. It is the layer beneath it.

The future state

If the thesis is right, the physical economy begins to run on an industrial-intelligence control layer, and the change is concrete rather than abstract. The next phase of AI will not only generate answers. It will operate assets, reduce downtime, move machines, balance energy, guide weapons, and automate regulated workflows.

In that future, factories become more autonomous and self-optimising rather than merely monitored. Industrial data stops being exhaust and becomes a strategic, defended asset. Maintenance moves from reactive to predictive as a default. Robots move out of caged, controlled cells and into hazardous and complex sites humans should not enter. Energy becomes both a hard constraint on where AI can run and a control point worth owning. Defence becomes software-defined, autonomous, and, where sovereignty is at stake, deliberately domestic. And across all of it, the companies that own the data, the workflow, the certified deployment, and the physical execution become harder to replace as the system matures, not easier. What becomes less valuable is the undifferentiated model, the commodity sensor, and any tool that sits at the edge of work rather than inside it.

The thesis, in one sentence

We believe the physical economy will reorganise around an industrial-intelligence control layer, AI embedded in factories, energy, robots and defence, because foreign dependency, energy constraint, labour scarcity and re-armament leave no alternative, creating durable value for the bottleneck layers this AI must run through (industrial data, certified physical execution, autonomy software, power and chips), layers Europe disproportionately owns and a market fixated on the model race currently underestimates.

The investment question

The market is still pricing the visible layer, the frontier labs, the cloud platforms, the companies with the clearest AI narrative. This thesis is a wager on the layer beneath it.

So the question that drives the rest of this report is not whether Europe can build the next OpenAI. It cannot, and it does not need to. The question is this: if AI moves from the chat window into the physical economy, what becomes scarce, strategic, and hard to replace, and who owns it? Europe may not own the visible layer. It may own the terrain the next layer has to run through. What follows tests that claim, the timing, the structural forces, the falsifiable convictions, where value accrues, what would make the thesis wrong, and the specific assets that would let an investor own it.

This future has been discussed for a decade. Five things changed the timing, and they changed it together.

The AI itself crossed the threshold. Foundation models, vision-language-action models and cheap edge compute turned industrial and physical AI from aspiration into deployment. NVIDIA's GR00T, Cosmos and Jetson Thor and Google DeepMind's Gemini Robotics are productising the control layer now, not in a research horizon.

The model layer began to commoditise. As frontier capability converges and cheapens, durable value migrates down to the data and the deployment, and in industry that wallet is barely open: industrial-AI spend is still only ~0.1% of industrial revenue (IoT Analytics). The money has not been spent yet, and it will be spent below the model.

The energy constraint turned binding. The IEA projects data-centre electricity demand roughly doubling to ~945 TWh by 2030, with around a fifth of planned capacity already facing grid-connection constraints, and strain running down to transformers and power electronics. Power is now the rate-limiter on the entire build-out, which turns the power-and-cooling layer from plumbing into a control point.

Ukraine and re-armament turned defence AI into procurement. European defence-technology venture funding rose roughly thirteenfold, from ~€200m in 2021 to ~€2.6bn in 2025 (McKinsey), against a backdrop of up to €800bn under ReArm Europe and a NATO 5%-of-GDP target. Doctrine now accepts that software-defined autonomy offsets conventional mass, and budgets are following doctrine.

Capital is reallocating into the physical stack. After a decade flowing into consumer software, capital is rotating toward industrial, defence, energy and robotics, the layers this thesis is about.

Five years ago the AI was not good enough and the urgency did not exist. Five years from now the control-layer winners will already be entrenched, their data compounding and their switching costs set. The window to own the layer, rather than rent it, is open now.

The single analytical claim this thesis stands on is that value in the physical economy will accrue to the control layer rather than the model layer, so it is worth stating precisely what the market is focused on, what it is missing, and why the miss persists.

What the market is focused on. The frontier-model race and its inputs: parameter counts and benchmarks, hyperscale compute, cloud platforms, developer ecosystems, and the capital pools funding them. It is a real contest with real winners, and it is measured obsessively because it is visible and consumer-facing.

What the market is missing. That the layer where AI enters the physical economy has a different economic structure. There, the model is a component, not the product; the product is the system that connects the model to a machine, a dataset, a regulated customer, and a real-world decision, and takes responsibility when something physical is at stake.

Where value will actually accrue. To the layers that are hardest to replace: contextualised operational data, certified hardware, safety-critical and accredited deployment, sovereign infrastructure, energy, procurement access, and owned workflows. These are defensible for reasons that have nothing to do with model quality. They are embedded in live operations, regulated, safety-critical, data-rich, local, operationally complex, and costly and slow to rip out. A model can be swapped with an API call; a certified data foundation running a plant cannot.

Why the insight is not obvious yet. Because the visible AI story is fast, consumer-facing and US-centric, while the control layer is slow, private and industrial. The mismatch, structurally critical but narratively invisible, is precisely the condition under which assets are mispriced. The market is not wrong that Europe is losing the model race. It is looking at the wrong layer. A major European venture investor reaches the same conclusion from the other side of the table: model performance that once required massive training budgets is now matched at a fraction of the cost, and inference pricing has fallen more than two orders of magnitude in two years (General Catalyst), the clearest sign that durable value is migrating below the model.

The thesis is not a bet on imagination. It emerges because four structural forces bear on the physical economy at once, and none resolves on its own.

Strategic dependency. Europe runs an industrial economy of global importance on an AI, compute and chip substrate it does not own. Dependency is tolerable in peacetime and intolerable when the dependent layer becomes the source of competitive advantage and the supplier is a strategic rival. The AI gap, the venture split, the model-production gap, the concentration of data-centre capacity abroad, stops being an abstraction the moment AI becomes the input that decides industrial competitiveness. A major European venture investor concedes the point directly: despite strong networking expertise, European providers "lack operational sovereign cloud solutions," forcing continued reliance on US providers for critical infrastructure (General Catalyst).

Energy. AI's appetite is running into physical limits. Data-centre electricity demand roughly doubles by 2030, AI-optimised demand more than quadruples, and around 20% of planned capacity faces grid risk (IEA). The strain runs down to transformers and power electronics, where lead times have blown out, and to critical inputs, where China refines an estimated ~99% of the world's gallium. Power has become the rate-limiter on the build-out, which makes whoever owns power and cooling strategic rather than incidental.

Labour and automation. The industrial workforce is shrinking and ageing while the work becomes more automatable: Europe's working-age population is projected to fall by nearly 2 million a year to 2040, enough to subtract about 0.4% from annual economic growth (McKinsey). Global robot installations reached 542,000 units in 2024 on an installed stock of 4.66M (IFR), and physical AI is making robots capable of inspection and manipulation that fixed automation never could. The constraint shifts from "can a robot do this?" to "who owns the intelligence and the data that makes it useful?"

Re-armament. The largest reallocation of European public capital in a generation is underway: up to €800bn under ReArm Europe / Readiness 2030, and a NATO 5%-of-GDP target implying on the order of €288bn a year in additional spending. Ukraine reset doctrine, with software-defined autonomy now offsetting mass, and the money is increasingly directed at the software and autonomy layer rather than legacy platforms.

Six testable claims follow from the forces above. Each is stated with the strongest counterargument, taken seriously, and the rebuttal. They are the load-bearing beliefs; the section on what would make the thesis wrong names what would snap them.

Conviction 1 — The beachhead is downtime. Predictive maintenance is where industrial AI first pays for itself. Unplanned downtime costs the world's largest firms an estimated ~$1.4tn a year, around 11% of revenue (Siemens/Senseye, True Cost of Downtime 2024); AI-driven predictive maintenance cuts it materially, and the WEF Global Lighthouse Network's leading sites report outsized gains. Counterargument: most industrial AI dies in "pilot purgatory" and never scales (McKinsey). Rebuttal: which is exactly why the investable point is the data-foundation layer beneath the pilots, since the firms that own contextualised operational data are the ones that escape purgatory.

Conviction 2 — Physical AI scales first in the dull, dirty and dangerous. Autonomous inspection, then manipulation, in hazardous heavy industry is the first commercial act of physical AI, with hazardous-environment inspection already a named sub-market, and Goldman Sachs puts the humanoid TAM at ~$38bn by 2035 (revised up sixfold) while naming manipulation software, not hardware, as the bottleneck. Counterargument: China's robotics scale, 54% of 2024 installs, ~2M installed stock, robotics central to the 15th Five-Year Plan, means the West cannot win robotics. Rebuttal: China wins the volume manufacturing of robots; the contested, defensible value in the West is certified deployment in regulated environments, the operational data the fleet generates, and the manipulation-software layer, not unit share.

Conviction 3 — Defence becomes software-defined, and Europe builds it sovereignly. The spending is structural (ReArm, NATO 5%), and in military AI software is already the largest layer with Europe the fastest-growing region, value concentrating in data integration, autonomy control and classified deployment. Counterargument: procurement favours legacy primes, the category leader looks richly valued, and battlefield reports show demonstrations are not combat. Rebuttal: the spend is multi-decade, software margins and combat-data feedback loops compound, and the sovereignty requirement is a moat US and Chinese vendors cannot cross in Europe, but the price and performance risks are real and must be underwritten, not waved away.

Conviction 4 — Energy is the binding constraint, and therefore an ownership layer. Demand doubling, ~20% of projects at grid risk, transformer and power-electronics strain (IEA). Whoever owns grid, power and cooling sells into a constraint that cannot be wished away, with data-centre demand already ~30% of orders at the leading electrical-infrastructure vendor and its fastest-growing driver. Counterargument: efficiency gains and new nuclear could relax the constraint. Rebuttal: even optimistic efficiency leaves grid, transformer and cooling bottlenecks binding this decade, because the fixes are physical and slow, which is precisely what makes the layer scarce.

Conviction 5 — Sovereignty is gated by physical chokepoints, not ambition. Some positions in the stack are effectively un-substitutable on any near horizon: a single European firm is the sole supplier of EUV lithography, and ~99% of refined gallium comes from China. Counterargument: chokepoints get designed around or subsidised away (the various CHIPS Acts). Rebuttal: chokepoints like EUV are decade-scale and capital-prohibitive to replicate; ownership of the chokepoint is the most defensible position in the entire thesis, and subsidies have so far reinforced rather than dislodged it.

Conviction 6 — Value accrues to the control layer, not the model layer. The industrial-AI market runs from ~$43.6bn (2024) to ~$154bn by 2030 at ~23% CAGR, with industrial DataOps the fastest slice at ~49% CAGR (IoT Analytics); the tell that the data layer is infrastructure rather than feature is that the category-leading platform is white-labelled by other software vendors. Switching costs on safety-critical data and the data flywheels underneath compound the position. Counterargument: the hyperscalers and foundation-model labs move down into the layer themselves. Rebuttal: contextualised industrial data, certified deployment and regulatory trust are hard for remote hyperscalers to replicate inside safety-critical, sovereign operations, so the layer is defensible because it is physical and local.

The prize is large: generative AI alone could add on the order of €575bn a year to the European economy by 2030, and enterprises expect full AI transformation to deliver two to three times the value cloud did (McKinsey). But read layer by layer, that value does not distribute evenly across the industrial-intelligence stack. At the top, the models commoditise. Beneath them, value pools in layers that are scarce, embedded and hard to replace, and it is those layers, not the model, that the ownership blueprint targets.

The chip chokepoint sits upstream of everything: whoever controls EUV lithography gates the advanced compute the whole stack depends on, and the position is a decade-scale monopoly. The power-and-cooling layer is scarce by physics rather than strategy, and gates where AI can be built at all. The operating and simulation layer, the systems that already run and model factories, converts incumbency into an AI distribution channel and an installed base that is expensive to displace. The data substrate, the contextualisation layer every industrial application reads from, is the new dependency, carrying switching costs and a data flywheel. Physical execution, certified autonomous hardware in hazardous environments, is gated by regulation and the operational data it generates. Defence autonomy, the sovereign software brain, earns software margins and a combat-data feedback loop competitors cannot buy. And at the frontier, manipulation intelligence, the unsolved control layer for dexterous robotics, is the highest-upside and least-proven layer of all.

What becomes more valuable is anything embedded in operations, regulated, data-rich, physically scarce or sovereign-gated. What becomes less valuable is the model itself, undifferentiated hardware, and any tool that sits at the edge of work rather than inside it. The investable opportunity is not the visible layer. It is the bottleneck layer that becomes unavoidable if the future arrives.

A thesis worth holding can be falsified. These are the conditions that would break it, each observable, each with a signal to watch.

Industrial AI stays in pilot purgatory. If the data-foundation layer never actually lets firms escape proofs-of-concept, the beachhead never becomes a business. Watch: net revenue retention and production-deployment counts at the data-layer companies.

China's scale captures the physical-AI layer. If cheaper hardware plus operational scale lets Chinese robotics take not just unit volume but the software and data layer in Western industry, the defensibility argument collapses. Watch: whether Western certified-and-data moats hold on price against Chinese quadrupeds and arms.

The model layer does not commoditise. If the foundation-model labs capture the industrial control layer rather than commoditising into it, value never migrates down to the assets in this blueprint. Watch: whether NVIDIA and DeepMind move from tooling into owning industrial data and deployment.

European re-armament stalls. If the spending pledges fragment on national politics and procurement reverts to legacy primes, the defence conviction loses its base. Watch: conversion of framework ceilings into firm orders, and national-champion preference in awards.

Energy constraints relax faster than expected. If efficiency and new generation remove the bottleneck quickly, the power-and-cooling layer loses its scarcity premium. Watch: grid, transformer and cooling lead times.

The sovereignty premium never materialises. If European buyers keep choosing American capability over sovereign control, the central reason these particular, European assets are mispriced disappears. Watch: sovereign-cloud and sovereign-defence procurement decisions, and whether they favour domestic providers when it costs more.

The point of this section is not to list companies exposed to the theme. It is to identify which assets may own the bottleneck layers of the future state, what claim each has on value capture, and what would have to be true for each to become a category winner. The blueprint has two tiers: a private tier where the asymmetry sits, and a public tier of control points that already hold parts of the stack and offer the same thesis with far less single-name and access risk.

Each name below is analysed against the same frame: the precise category, why that category is a bottleneck, progress to date with specific evidence, the real competitive structure, the strongest winner's case, the strongest bear case, and a clear judgment. Private-company figures move quickly; several here are reported, contingent, or single-source and are flagged as such.

The blueprint resolves on two different axes, and conflating them is the most common mistake in a thesis like this. One axis is thesis-centrality, how essential the layer is to the story. The other is ownability, whether the asset can actually be bought, at an acceptable price, on acceptable terms. They do not line up.

On centrality, the layers the thesis cannot be told without are the data substrate (Cognite), the compute chokepoint (ASML), the energy layer (Schneider) and the operating layer (Siemens), with defence autonomy (Helsing) close behind. On ownability, the picture inverts: the most central private name (Cognite) is among the hardest to own cleanly, the richly valued one (Helsing) is a right-price call, and the frontier name (Acumino) is a venture sleeve, not a position, while the public control points are all ownable today in listed form.

That argues for a barbell rather than a single basket. Own the layers that are both central and ownable now, the public control points (ASML, Schneider, Siemens) as the liquid core, plus the most straightforwardly ownable private (ANYbotics) and the data substrate (Cognite) where access allows. Price the asset that has already priced in dominance (Helsing) and enter only with discipline. Watch the frontier option (Acumino) and size it as a small, high-upside sleeve. The goal is not to hold every beneficiary of the theme. It is to own the bottleneck layers the future state must run through, weighted by how defensible each is and how cleanly it can be bought.

This thesis is European in logic and, usefully, actionable from Greece. The macro is concrete: Greece has committed to a ~€25bn, twelve-year defence-modernisation plan, and it sits inside the NATO Innovation Fund (a >€1bn, 24-nation vehicle), putting a Greek vantage point directly on the defence-autonomy and dual-use flow this thesis tracks. The frontier name in the blueprint, Acumino, is Greek-founded, which turns an abstract "manipulation is the bottleneck" argument into a tangible, local diligence case with real access.

More broadly, a Greek and diaspora vantage offers something the thesis specifically rewards: proximity to European industrial, energy and defence decision-makers, and a sourcing and convening position at the edge of the ecosystem rather than the centre of the model race. The opportunity is not to build the visible layer from Athens. It is to identify, access and help build the companies that own the terrain beneath it, and to be early where the mispricing lives.