Key Takeaways

  • Mario Draghi and corporate leaders like Jamie Dimon claim Europe is falling behind, but Dominik Leusder argues that Europe captures consumer and enterprise welfare without funding the underlying capital expenditure.
  • Chasing American hyperscaler capex on frontier models is a financial trap; European policy should prioritize sovereign tech stacks and enterprise adoption over prestige foundation models.
  • Regulators and operators confirm that near-frontier open models like Mistral handle the vast majority of corporate use cases without requiring multi-billion-dollar training runs.
  • Severe domestic compute constraints have pushed European firms like Mistral to build on top of Chinese model chassis and pre-trained weights, creating a distinct dependency risk.

The Frontier Capex Trap

European policy discussions are haunted by panic over falling behind the United States. Corporate leaders like Jamie Dimon and former European Central Bank president Mario Draghi regularly cite lagging gross domestic product growth as proof of European decline. When real growth and purchasing power parity data contradict that gloom, critics pivot to the next perceived failure: foundation models.

Leusder sees this pivot as a moving target: “People in the Europe-US discussion have sort of shifted the goalposts a few times by saying, 'Well, you can forget about PPPs. Forget about real growth then. Whatever, even if there hasn't been any decline relatively speaking so far, Europe is sort of renting the future, and it will miss out on the capital expenditure-generating technology of the day which is generative AI, and these frontier models.'”

Spending tens of billions of dollars on data centers to train frontier models makes little economic sense for Europe. The United States and its hyperscalers are bearing the risk of massive capital depreciation. Europe has historically absorbed technological gains without paying the initial development bills. As Leusder notes, “consumers have pocketed the welfare nonetheless by lower prices for tech goods and services, and that's despite the fact that US tech companies have pretty decently entrenched moats in Europe, legally entrenched ones. And I think with AI, it's going to be much the same.”

Near-Frontier Models Fit Enterprise Reality

Most enterprise workflows do not require bleeding-edge frontier intelligence. They require deterministic outputs, data privacy, predictable latency, and low serving costs. A company processing internal invoices or classifying customer support tickets does not need the largest American frontier model running at maximum compute cost.

Leusder points out that European policy and business leaders are coming to the same conclusion: “My impression talking with people who are in this field and also who regulate AI at the European level, in their opinion, near-frontier models like Mistral are totally fine for the vast majority of business needs of most European firms.”

Trying to match Google, Meta, or Microsoft dollar for dollar in cluster size is a fool's errand. Instead, Europe benefits by letting foreign capital subsidize raw model training while local companies apply open weights to actual industrial processes.

The Unseen Compute Chokepoint

Skipping the frontier race does not mean Europe is entirely free of risk. The continent faces an acute shortage of local compute infrastructure. Because European developers lack the hardware clusters required to train large base weights from scratch, they must adapt architectures built elsewhere.

Leusder highlights a striking example in France's flagship AI champion: “The interesting fact, of course, that Mistral is launching models with a Chinese chassis, so with basically with the weights of a Chinese model because Europe doesn't have the compute to actually train their own weights, raises all sorts of other dependency issues, I suppose.”

True technological independence will not come from building a European rival to OpenAI. It will come from securing independent compute infrastructure and controlling model weights, ensuring domestic businesses are not left vulnerable to foreign export bans or closed application programming interfaces.

What to Do With This

Audit your product's token expenditures across every active route tomorrow morning. Migrate routine classification, extraction, and drafting tasks off frontier American APIs and onto self-hosted open-weight alternatives like Mistral. Confine your highest-tier frontier model calls exclusively to the small fraction of workflows that actually demand top-percentile reasoning.