Aleph Alpha, Germany’s great frontier hope, was celebrated in 2023 for what was billed as a half-billion-dollar round; this spring it was absorbed into Canada’s Cohere, its shareholders left holding roughly ten percent of the combined entity. Mistral, once hailed as Europe’s answer to OpenAI and Anthropic, tells a similar story: its latest big launch, Mistral Medium 3.5, arrived in April and is not even competitive with Qwen3.8 27B, an open-weight model from Alibaba a fraction of its size that I run for free on my laptop.
The reflex is to read this as another chapter in the long book of European tech decline. I read it as a turning point in our trajectory.
I have spent my career inside the European software economy: commercializing AI at SAP years before anyone had heard of a large language model, scaling an EU SaaS start-up, and now investing in the companies building on this wave. So I will not pretend the old criticism was wrong. Europe was slow. Too much fear of what innovation might break, too little expertise among start-up CEOs in how to commercially scale a business, and a decade of watching American platforms take markets we considered ours. But the structure of the AI game has shifted in the past eighteen months. Europe cannot fund the frontier model race. That is not the question that matters. The question that matters is whether the race is worth being in at all. The prize at that finish line is depreciating faster than almost any asset in technology history, and the durable value is accumulating one layer up, in a place Europe is unusually well built for.
A race Europe could not buy into
Start with the numbers. The Financial Times’ compilation of Q1 earnings puts the four largest American hyperscalers at roughly $725 billion of planned capital expenditure for 2026, up around 77 percent from about $410 billion the year before. OpenAI alone has announced compute commitments of about $1.4 trillion over eight years, though it has since steered investors toward some $600 billion in actual spending through 2030. Europe’s flagship response, the InvestAI program, is a €200 billion mobilization target of which around €50 billion is pledged public money, much of it redirected from existing EU programs. The first “AI gigafactories” were promised €20 billion of it. When the formal call finally went out on 30 July, it offered up to €10 billion in EU and national funding for as many as seven sites, with a further €20 billion sought from private investors. Awards are due in early 2027, with first construction expected later that year.
If every euro of the €200 billion target materialized tomorrow, in cash, it would fund a bit more than three months of American hyperscaler capex at the 2026 run rate. The gigafactory money that actually reached a call, €10 billion over half a decade, is what a single US hyperscaler now spends in about three weeks.
- $725bn Planned 2026 capex, the four largest American hyperscalers
- $600bn Actual spending OpenAI now steers investors toward, through 2030
- €200bn InvestAI mobilization target, of which €50bn is pledged public money
- €10bn Actually offered in the first gigafactory call, for up to seven sites
The deeper problem is that the gap was never really about willpower or vision. Frontier training runs on two inputs, energy and chips, and Europe is structurally short of both. European industry pays roughly twice the American price for electricity, and Germany’s new subsidized industrial rate leaves data centers out. The one cheap region, the Nordic hydropower belt at around three cents per kilowatt hour, realistically supports single-digit gigawatts in total, far below what individual American campuses now draw. On chips, Europe designs almost none of the accelerators and deploys under five percent of global AI compute. The bitter irony is that Europe owns the single most concentrated chokepoint in the entire supply chain, ASML’s one hundred percent monopoly on EUV lithography, and exports the advantage to Arizona and Taiwan. Europe’s largest lab, Mistral, is worth a fraction of OpenAI or Anthropic.
Let me be precise about what Europe was priced out of. Not AI as such: wherever a frontier can be reached with talent rather than a hyperscaler’s balance sheet, European teams keep reaching it, with the FLUX image models shipping from Freiburg’s Black Forest Labs, ElevenLabs setting the pace in voice from London, and some of the most credible world-model teams forming here. The deficit was never ability. The race Europe could not buy into is the general-purpose one, where a single training run is a billion-dollar event, and there it was priced out. The macroeconomics of energy, compute, and capital made participation impossible.
But the past two years have revealed that being priced out of this particular race is a stroke of luck, because the lead you might achieve can vanish in an eyeblink. The price of a fixed level of AI capability is collapsing at roughly ten times per year. What GPT-4 charged in early 2023, about $30 per million input tokens, is available today for under fifty cents, often from open-weight models you can run on your own hardware. And the pursuit has never been closer than this month. On July 17, China’s Moonshot AI unveiled Kimi K3: at 2.8 trillion parameters the largest open-weight model ever built, ranked by independent evaluators just behind the two newest American flagships on aggregate capability, and first in the world in blind preference tests for frontend coding. One tier below it, models like GLM-5.2 deliver near-flagship coding at roughly a third of closed-model prices.
To be fair, and this matters, the closed frontier keeps its lead where it counts most. On the hardest long-horizon work the gap remains wide, it grows with task difficulty, and it is probably larger than the leaderboards show, since the American labs hold their strongest models back.
But measure the lag however you like, at Epoch AI’s central four months or anywhere in the two-to-six-month range analysts now argue over, and the strategic shape is the same. Each generation costs multiples of the last, whilst the pursuer arrives within one or two financial quarters and sells the result at commodity prices. Hundreds of billions of dollars, spent annually, to hold a lead that melts away between product cycles.
Involuntary or not, Europe has been spared the most expensive commodity purchase in industrial history. And value never disappears when a layer commoditizes; it migrates to the layers around it.
In January I published a white paper, The Harness Thesis, arguing exactly this: that foundation models are becoming frozen, interchangeable infrastructure, and that the durable value accumulates in the harness built around them. In essence the harness is three things: a routing brain that sends each task to the right model on the right hardware, a memory of how a specific organization actually works, and a checker that tests output against reality. Six months on, the market data reads like a progress report on that argument.
The routing layer
The squeeze is happening exactly where the margin compression logic said it would: at the model layer. The labs run heavily negative net income whilst selling the dream of AGI, and enterprises treat models as swappable suppliers and are measurably swapping them. On OpenRouter, the largest neutral routing marketplace, American models’ share of routed tokens fell from roughly 70 percent to 30 percent in a single year, and at its peak 58 percent of the tokens routed for US companies went to Chinese open-weight models. Named Western companies, Coinbase among them, have cut their AI bills roughly in half by rerouting agent fleets to cheaper open models. Meanwhile, Menlo Ventures’ survey of nearly five hundred enterprise buyers found generative AI spend reached $37 billion in 2025, more than tripling year over year, with over half of it going to the layer above the models.
Europe’s card in this first layer is geopolitical as much as technical. The model market is now genuinely heterogeneous: different models lead at different tasks, and they run best on different silicon, from NVIDIA to Cerebras to Amazon’s custom chips. Whoever routes across all of them can deliver the same output at a fraction of the cost, and neither the American hyperscalers nor the frontier labs can credibly offer that, because each is structurally committed to its own models, its own chips, or both. Neutrality is the one feature the giants cannot bundle, and Europe’s position between the American and Chinese AI blocs converts it into a business model: data-retention terms an enterprise can believe, freedom to send every workload to the cheapest capable model on the cheapest capable chip, and no strategic agenda beneath the routing decision. That is the position companies like Callosum, a London-based platform orchestrating workloads across models and chips, are converting into enterprise contracts. (Disclosure: I am a partner of InThisTogether.vc, and facilitated an investment in Callosum.)
Routing decides where the work goes. The second layer decides whether the work is any good, because an agent without knowledge of the company it serves is a very fast intern.
The knowledge layer
Every European strategy essay eventually recommends “specialization,” so let me first concede what that advice ignores. Europe has bet on the complementary layer before and lost badly. The sovereign cloud initiatives of the last decade produced no European hyperscaler whilst AWS, Azure, and Google took seventy percent of Europe’s own market; Nokia owned mobile and ceded the platform, and with it the entire app economy. Both failures share that Europe held a position with no proprietary substance beneath it, the software equivalent of what we today call a GPT wrapper, and the platform owners simply rolled over it. So the claims in this layer and the next have to pass a test, and Europe’s industrial position suggests it has three parts. A data position is defensible when three things hold: a large, long-lived installed base of your machines or software operates around the world; a service or regulatory relationship generates continuous, proprietary data from that base; and switching costs plus deep domain know-how resist reverse-engineering.
The first place Europe passes the test is inside its own enterprise software. Ask anyone who has deployed AI inside a large organization what limits the results, and the answer is rarely the model. It is context: how this company quotes, approves, escalates, and handles the exceptions nobody wrote down. In the language of my paper this is the preference layer, the accumulated memory of how a specific organization works, and it compounds with every workflow it captures. It is also the layer Europe has spent half a century accidentally preparing for. By SAP’s own count, 77 percent of the world’s transaction revenue touches an SAP system, and more than 400,000 companies run their operations through software designed in Walldorf. The installed base is there, the switching costs are prohibitive, and the process logs flowing off that base are exactly the continuous proprietary data the test demands. The diligent, exhaustive process documentation that consultants mocked as bureaucracy for a generation turns out to be the raw material of the agentic era, because an AI agent is only as good as the process knowledge it is wired into.
The scaled proof that this converts is Celonis. Founded in Munich in 2011, it built the process mining category on exactly this move, reading how an organization actually works out of the logs its systems already produce, and today serves more than 1,400 enterprise customers, on a $13 billion valuation from 2022. Its product line has quietly become an agent-feeding machine: the Process Intelligence Graph, a digital twin of a company’s operations, tripled its customer count in a year to over 300 Global 2000 enterprises, and Celonis shipped the first process intelligence MCP server, a standard plug that feeds this organizational memory directly into AI agents. The company’s own survey of 1,600 business leaders this February found 85 percent of organizations want to be an “agentic enterprise” within three years, whilst 76 percent admit their current processes are holding them back. Discount the self-interest in that survey and the gap it describes is still the market for this entire layer. A generation behind Celonis, Interloom in Munich raised a $16.5 million seed to build the context graphs that hand agents this memory natively, and Berlin’s n8n more than doubled its valuation to $5.2 billion this year as SAP joined as a strategic investor, a measure of how much pull the layer above the models now has.
And this layer has a clock, ticking louder and louder by the day. The employer-affiliated German Economic Institute calculated in June that Germany’s labor force will be short 4.3 million workers by 2036, a projection revised up from three million only two years ago. Of the 14.1 million baby boomers yet to reach retirement age, only 7.6 million will remain below it in 2030. The rest retire over the six years that follow, at an average pace of roughly 1.3 million people a year. Most of what they know was rarely written down. Every one of those retirements takes uncaptured process knowledge with it, which makes this decade a one-time window to move Europe’s organizational memory out of heads and into systems that agents can use.
Which raises the question German software has to answer this decade, and the risk if it hesitates: SAP owns more business-process ground truth than any company on earth, and it will either convert that into the knowledge layer of enterprise AI or become the system being orchestrated by someone else’s agents. The American platforms want this layer too; Microsoft is wiring agent memory into every Office seat. The difference is where the knowledge already sits, and for the world’s industrial processes it sits disproportionately in European systems.
But a company’s memory of itself, however deep, only tells an agent what should happen. The third layer is where AI output meets the question that decides everything in the physical economy: is it actually true?
The ground truth layer
Apply the same three-part test to the physical world, and you find that Europe’s most defensible AI assets have been hiding inside its most boring companies all along.
Rolls-Royce does not really sell engines. It sells flight hours, and its Trent fleet has logged more than 150 million of them. The telemetry from every one of those hours feeds the predictive maintenance behind power-by-the-hour, an entire business model built on data no competitor possesses. It is also why China’s COMAC still cannot ship a jet without Western engines.
The European elevator majors run the same play. Kone, Schindler, and TK hold decades-long service contracts across millions of units in Europe alone. Every ride is telemetry. The service annuity is the profit engine and the data moat at the same time.
And ASML turns out to be the purest example of the pattern. Its real fortress is a three-decade feedback loop of manufacturing data with TSMC, Samsung, and Intel. China has copied pieces of the hardware. What it cannot copy, analysts at TrendForce conclude, is that data loop, because no amount of money can shortcut thirty years of production runs.
The startup proof that this transfers to AI is PhysicsX, founded in London by Formula 1 engineers, which raised at a $2.4 billion valuation this June; IEEE Spectrum reported that at General Motors an in-house large physics model cuts a two-week aerodynamics cycle down to minutes, and behind that speed sits years of simulation and physical test data no lab can download. A frontier model can imitate the style of an aerodynamics report. It cannot know whether the wing works, because knowing requires the record of the last thousand wings; the ones that worked and the ones that didn’t.
This is where founders and Mittelstand owners should sit up. Across European machinery, pumps, compressors, and marine equipment, the installed base exists and the data is being generated every hour, but much of it is never captured; the moat is real and uninstrumented. That is the opportunity, and it is also the clock, because every plant that closes and every fleet that retires takes its future data with it, and competitors in the US and China are building instrumented installed bases of their own from scratch.
I will not oversell it, because the test cuts both ways and Europe has already failed it twice this decade. In wind turbines, Chinese manufacturers swept the global top three in 2024; the installed-base advantage now grows in China, not Denmark. In premium automotive the reversal is more brutal still: the presumed moat was combustion-era engineering data, the EV transition devalued it, and German carmakers now license their driving software from Chinese firms. The dependency flipped. Where the underlying data is public, synthesizable, or accumulating faster elsewhere, frontier models and Chinese industry will swallow the niche. The durable positions are the ones that resist this, and the open challenge there is no longer whether such companies can be built in Europe. It is whether they can commercially scale into the American market without losing momentum: a problem many European founders face, and one they tend to solve faster with builders who have been there before.
The bottom line: America owns the models, and Europe owns an outsized share of the ground truth they must answer to. Verified output is the one product in AI whose price is unlikely to collapse.
Dependency kills
There is one warranted objection to all of this: if Europe runs its economy on models it does not control, it is exposed, and June removed any doubt about how exposed.
On June 12, the US Commerce Department applied export controls to Anthropic’s newest models, Fable 5 and Mythos 5, three days after their launch, citing a reported jailbreak of the models’ safeguards. The directive required cutting off every foreign national worldwide, and since Anthropic could not verify nationality in real time, it switched both models off for every customer on the planet. The lesson was priced within days. Days into the blackout, Zhipu, the newly listed Chinese lab behind the GLM models, released GLM-5.2, an open-weight model far stronger than anyone expected. Its shares surged 33 percent in a single session as Wall Street raised its bets on Chinese AI after the Anthropic curbs, and within ten days its market value touched a record HK$1 trillion. Enterprises had realized that closed American intelligence can be switched off by a directive, and open Chinese intelligence cannot. Washington lifted the directive at the end of the month. The demonstration stood.
Nor was it a one-off. In January 2025, the US AI Diffusion Rule sorted even NATO allies into tiers, capping advanced chip access for Poland and the Baltic states; it was rescinded, but it was drafted, signed, and published, which is the part that cannot be unseen.
And this month delivered the pattern’s logical conclusion from both directions at once: Washington is reportedly weighing restrictions on Chinese open-weight models, whilst Beijing’s Ministry of Commerce has discussed export curbs on its own top models with Alibaba, ByteDance, and Zhipu. Both blocs, in the same month, contemplating switches they could throw on the models the world runs on. “Dependency kills,” as the Hague Centre’s Tim Sweijs put it.
So yes, dependence is real. The mistake is concluding that the answer is a European counter-race, because racing and losing produces the same dependency plus a burned decade of capital. The answer the numbers support is optionality: the best open-weight models running on European soil, a moderate sovereign compute base for running inference, and at least one credible lab that keeps the ability to fast-follow alive. That is the strategic case for Mistral, whose industrial contracts with CMA CGM, Stellantis, and Airbus matter less as a bid for the frontier crown than as proof of this “insurance policy”. It is fitting that Mistral’s largest shareholder is now ASML: the one data moat Europe already owns, underwriting the capability it refuses to be without. Nathan Benaich of Air Street Capital put this in one line: “The point is not to win the race. It is to make sure no one else can end it for you.”
Picture the European AI economy this argument implies, five years out. The models come from wherever they are cheapest that quarter, American, Chinese, open-weight, swapped as casually as cloud regions today. Above them, the layer that routes, remembers, and verifies is run from Berlin and London and Stockholm, wired into data based on European companies that produce the one thing no model can generate: proof that the answer is right.
Let them race. Europe should build the harness, because the models are becoming a commodity, and the harness is becoming the layer where value accumulates.
