Asymmetric Innovation

On 27 January 2025, the most valuable company on Earth lost more value in a single day than it ever had before — than any company ever had before. Nvidia shed $589 billion in one session, a 17% fall and the largest single-day loss in stock-market history. The cause was not a war, a scandal, or a recession. It was a Chinese AI model that was simply cheaper. In a few hours, the market repriced a belief the entire US AI industry had rested on: that leadership belongs to whoever spends the most. This article is about why that belief is wrong, how China exposed it, and why Europe — written off in nearly every AI ranking — has been reading the lesson backwards. Europe's opportunity is not to spend its way to parity. It is to stop trying.

Sumo versus judo

“Asymmetric warfare” is not a loose metaphor for being outgunned; it is a precise idea. The political scientist Andrew Mack asked in 1975 why great powers lose small wars, and Ivan Arreguín-Toft later sharpened the answer: the weaker side wins when it refuses to fight on the stronger side's terms and adopts the opposite strategy. B.H. Liddell Hart called it the indirect approach — never assault a prepared position head-on; dislocate it first. John Boyd reduced it to tempo: win by operating inside your opponent's decision cycle rather than by overpowering him.

The business translation already exists. In Judo Strategy, Harvard's David Yoffie and Mary Kwak argue that rather than oppose strength to strength, successful challengers use their opponents' size and power to bring them down, through three principles: movement, balance and leverage. The incumbent's instinct is the opposite — sumo: win by mass, by being bigger and pushing harder. The whole drama of AI today is a sumo contest between the United States and China over who can amass the most compute and capital. Europe keeps asking how to enter that ring. The better question is why it would. A judoka never becomes stronger than the giant; a judoka uses the giant's own momentum to put him on the floor. Everything that follows is an argument for Europe to fight judo.

Part I — China: judo against the giant

Return to that January morning. The trigger for the largest market loss in history was not a bigger model — it was a cheaper one, and the panic exposed what the incumbents had missed: their defining strength, the belief that frontier AI requires ever-more capital and compute, had quietly become a liability. Deconstruct DeepSeek's play into three asymmetric moves.

Move one — constraint as catalyst. Barred by US export controls from Nvidia's top chips, DeepSeek trained on the deliberately throttled H800: the same computational power as the H100 but lower network bandwidth, the export-compliant version Nvidia built for the Chinese market. Rather than treat this as a cage, they engineered around it — a Mixture-of-Experts model with 671 billion parameters but only 37 billion active at a time, Multi-Head Latent Attention to cut memory use, and FP8 mixed-precision training, run on 2,048 H800 GPUs over about two months for roughly 2.7 million GPU-hours of pre-training. The scarcity meant to cripple them forced an efficiency edge the well-supplied US labs had no incentive to discover. By one hardware analysis, the result delivered comparable performance for roughly eleven times less hardware cost.

Move two — open weights as leverage. DeepSeek released its models open-weight. This is the purest judo in the story: it commoditises the exact layer — frontier model access — that the US labs monetise at a premium. When the capability you sell at high margin is suddenly free and runs far cheaper, your pricing power evaporates. The $589 billion Nvidia wipeout was that leverage made visible: investors grasped in a single session that efficiency could decouple AI progress from the limitless GPU demand on which the whole valuation rested. Tellingly, even Nvidia conceded the model was an excellent advance built on export-compliant compute.

Move three — free-riding on the frontier, the contested one. OpenAI has said it found evidence of “distillation” by China-based groups seeking to replicate US models, and told the US Congress that DeepSeek-associated accounts developed methods to reach its models through obfuscated third-party routers that masked their source. The White House AI adviser David Sacks called the evidence “substantial,” though without detailing it. DeepSeek itself has acknowledged distilling some of its smaller models from open systems such as Llama, but the distillation-of-OpenAI claim remains an allegation, not a proven fact. Strategically, though, the logic is unmistakable: turning a rival's most expensive asset — its frontier model's outputs — into your own cheap training input is leverage in its rawest form.

Now the correction that makes this case usable rather than embarrassing. The famous “$5.6 million” was only the cost of a single pre-training run at notional rental rates, excluding R&D, salaries, failed runs and hardware. SemiAnalysis estimates DeepSeek's true picture at around $1.3 billion in total server capital expenditure, on a fleet of roughly 50,000 Hopper GPUs. Even DeepMind's Demis Hassabis judged the headline figure “exaggerated and a little bit misleading,” reflecting only the final training round. For scale, OpenAI's GPT-4 reportedly cost more than $100 million to train.

So the lesson is not “a broke startup beat the giants.” It is sharper: a challenger with a fraction of the incumbents' resources won by refusing their terms — efficiency over scale, openness over enclosure, speed over capital. Not a budget story. A strategy story. And it is the one Europe must read correctly.

Part II — Europe: learning to fight judo

Why a frontal response fails

Begin with the arithmetic, because it ends the debate. An MIT scenario report finds that Europe controls just 5% of global AI compute against the United States' 80%, that the EU's €200 billion InvestAI pledge is dwarfed by a single year of American spending, and that summer 2026 is the last actionable window before the gap becomes self-reinforcing. Mistral, the continent's most serious lab, operates at a scale that barely registers against frontier US or Chinese labs. European industry pays roughly double the electricity rate of its US counterparts. And the flagship response — four AI gigafactories under that €200 billion mobilisation target — is not due to be operational until 2027–2028, on public-procurement timelines, against a window the same analysts say closes this summer.

Put plainly: if Europe enters the sumo ring, it loses before the bout begins. It has less compute, less capital, costlier power and slower institutions. Every euro spent matching American scale head-on is a euro spent losing more slowly. The honest conclusion is not despair but redirection. The frontal game is unwinnable, so Europe must find the dimensions where the giants' mass becomes a liability.

Where Europe's leverage lies

Europe has more of those dimensions than the declinist commentary admits. They fall into two groups: assets it already holds, and games it can still choose to play.

The chokepoint nobody mentions. The entire US–China compute race runs through one European company. ASML, of the Netherlands, holds a 100% monopoly on EUV lithography and about 94% of the overall lithography market — the machines that print every advanced chip on Earth; no Nvidia GPU, no Apple silicon, no AI revolution without them. That position took three decades, tens of billions of euros and a supply chain of some 5,000 specialist partners to build, with machines now priced around €200 million each, and it has made ASML Europe's most valuable technology company, worth over $553 billion. The implication inverts the declinist story: Europe already owns the most defensible position in the entire AI stack — its base. When Washington wanted to slow China, the lever it reached for was chip-tool access — which is to say, European technology.

The cleanest cheap power on the continent. AI compute is, in the end, electricity wearing a silicon costume — and European industry already pays roughly twice the US rate. Except where it does not. France runs on about 67% nuclear, with carbon-free sources at 95% of its mix and a carbon intensity of 21.3 grams of CO2 per kilowatt-hour — among the lowest in the world, against a global average near 470. Electricity there costs roughly one-sixth of Germany's, a rare pairing of low operating cost and low emissions for always-on AI compute, and data-centre load is highly stable baseload, varying only about 5% across the day — the exact profile nuclear serves best and intermittent renewables serve worst. France is already a net electricity exporter at a record 89 TWh. Europe cannot out-build US compute, but France and the Nordics can offer what the strained American grid increasingly cannot: abundant, cheap, low-carbon baseload for training and inference. Cheap clean power is a moat measured in decades, not quarters.

The deepest talent bench. Europe trains more AI talent than either rival: by one mapping it has roughly 30% more AI professionals per capita than the United States and nearly three times as many as China. Its weakness is not the base but retention — it exports its best people to the US, the UK and the Gulf. Yet the current is turning: tightening US immigration is pushing researchers back across the Atlantic, and the EU's €500 million “Choose Europe” initiative, with long-term grants and relocation top-ups, is a first attempt to catch them. The move is not to out-produce talent; it is to stop giving it away.

The data the giants do not have. The frontier is migrating from text to physical AI — robotics, agents, industrial systems. The US labs were built on consumer-internet data; they do not own the factory-floor, machine-tool, grid, aerospace and precision-manufacturing data the next paradigm runs on. Europe does — Siemens, Schneider, Bosch, Airbus, ASML, and the automotive and pharmaceutical bases. Whoever holds the proprietary data for the coming paradigm has a moat no amount of web-scraping replicates. This is uncontested water: compete where the incumbents have no foothold rather than where they are entrenched.

Turn the compute gap into an efficiency mandate. The 5%-versus-80% deficit is, on sumo logic, a death sentence; on judo logic it is precisely the H800 condition that produced DeepSeek's edge. Constraint is the most reliable forcing function for algorithmic efficiency — and efficiency, not raw scale, is the dimension on which the incumbents are least defended. Europe should stop apologising for its compute deficit and start treating it as the discipline that yields a smaller, cheaper, faster class of model.

Weaponise openness and regulation. Mistral should not try to out-scale frontier labs; its real assets are an open-weight heritage and a home market where trust is currency. Open weights commoditise the layer the US labs enclose. And the AI Act and GDPR, routinely cast as handicaps, are a leverage point: for banks, hospitals, defence ministries and public administrations, data residency and provable compliance are not friction — they are the product. This is the Brussels Effect turned into a go-to-market: the jurisdiction that writes the rules can sell the only stack that natively satisfies them.

Own the agent layer. Value in AI is migrating from the model to the agent — systems that plan and execute multi-step work rather than answer single prompts. That layer rewards small, fast teams over capital-heavy incumbents; it is the one place a European startup can ship weekly while a frontier lab deliberates. Mistral has shown the template, raising $830 million for a sovereign data centre with the French public bank Bpifrance as lead lender and committing €1.2 billion to infrastructure in Sweden. The opening is less to clone OpenAI than to own the orchestration and application tier that sits above any model — European or otherwise.

Be the trusted third pole. Much of the world wants AI beholden to neither Washington nor Beijing. Europe's rights-based, rules-first posture — usually derided as a handicap — is exactly the brand a neutral buyer would pay for: governments, regulated industries and the many states unwilling to hardwire dependence on a superpower. Sovereignty as a service, layered onto a single market of roughly 450 million people across two dozen languages, lets Europe offer multilingual, jurisdiction-respecting AI that neither superpower will build.

What to take from the playbook — and what not

Europe should copy China's strategic logic and refuse its specific tactics. The logic is legitimate and sound: pick the dimension where the incumbent is weakest, turn its mass against it, move faster than it can react, and commoditise the layer it monetises. The tactic to reject is the contested one — free-riding on a rival's frontier outputs. Beyond the legal exposure, it is strategically shallow: a one-time exploit that invites exactly the retaliation it drew, from export controls to access countermeasures. Durable asymmetry is built on assets a rival cannot revoke — cheap clean power, a lithography monopoly, proprietary industrial data, institutional trust — not on borrowed capability.

Two older ideas should anchor the discipline. Gary Hamel and C.K. Prahalad's “strategic intent” describes how resource-poor challengers — postwar Japan, then Korea — won by setting ambition wildly disproportionate to their means and leveraging scarce resources rather than lamenting them. And Alexander Gerschenkron's “advantages of backwardness” explains how late industrialisers leap precisely by adopting the frontier and skipping the incumbent's sunk stages. Europe's deficits, read correctly, are the preconditions for a leap. One caveat the data forces: Europe does not actually lack capital — it lacks channels, with vast household savings sitting idle. Mobilising patient European capital toward its own asymmetric bets is the financial half of the same strategy.

The choice

The image to end on is the one we began with: a giant on the floor, undone not by a bigger giant but by a smaller, faster opponent who refused to grapple on the giant's terms. Europe has spent years staring at a sumo ring it cannot win, mistaking its absence from that contest for irrelevance. It is not irrelevant. It owns the chokepoint the entire stack depends on, the cleanest cheap power on the continent, the deepest talent bench in the world and a trust neither superpower can manufacture. What it has lacked is not assets but a theory of the game — the recognition that the winner of an asymmetric contest is rarely the strongest party, but the one who best chooses where to fight. The window is open through this summer. Europe's choice is stark and simple: design its own game, or remain a tenant in someone else's compute empire.