Europe’s AI Sovereignty Problem Runs Far Deeper Than Frontier Access

This three-part series is published in partnership with the AI Now Institute.

This weekend, the United States administration ordered Anthropic to cut off foreign access to two of its advanced AI models. In Europe, the news has been interpreted almost unanimously: a wake-up call that shows dependency means access can be cut off overnight.

In this series, we make the case that Europe’s AI market is deeply entangled with the ecosystems of dominant US players, in ways that boosting both supply and demand alone cannot disentangle. As a result, interventions that seek to secure sovereignty in AI risk leading to much more entrenched dependence. The first part of this series looks at the demand-side of things.

The limits of AI sovereignty at the application layer

The Swedish vibe coding start-up Lovable is lauded as one of Europe’s biggest AI success stories — and understandably so.

Founded in Stockholm in 2023, it was quickly hailed as the fastest-growing software startup in history. Currently, the company is valued at $6.6 billion after raising a total of $540 million and is frequently cited alongside Mistral as proof that Europe can build globally competitive AI companies. Lovable lets anyone build a website or an app (“something they love”) by chatting with an AI, and lots of people do.

But Lovable’s rise is dependent on the kindness of strangers. Instead of developing their own models, deep technological moats, or exploiting a local technology stack, the company builds on the American tech stack: prompts are routed to models made by Anthropic and Google and distributed in the existing hyperscaler ecosystems. For example, on June 4, Lovable announced a deepened partnership with Google Cloud that is supposed to lead to a “fivefold increase” in Lovable’s Google Cloud footprint, with recent integration of the products into the Google Workspace.

Let us be clear. There’s nothing wrong with focusing on making AI useful, and building on top of existing models. It’s also hard, requires speed of execution, great UX, nurturing a community of users, and getting there first. As a long-term business strategy, however, it carries a risk.

In April, images posted on X appear to show that Anthropic created a new in-chat app builder that lets users generate applications from simple prompts — essentially competing with their client Lovable’s entire product. Weeks later, Anthropic launched Claude for Legal, a law-focused offering that put legal AI startups on the defensive — including Harvey, Legora, and Robin AI, all of which deploy Claude models at the core of their products. The supplier had become a competitor twice over within a matter of weeks. This pattern — an incumbent absorbing the feature a smaller company built on top of it — is known as getting “sherlocked,” is a well-known, yet often-forgotten trope in the tech industry, named after Apple introduced a new feature that rendered a popular third-party tool irrelevant. 

Not all European start-ups build on proprietary US foundation models: some train their own, some run smaller, domain-specific models on proprietary data and on-premise infrastructure, or focus on physical-AI. However, we begin this series by interrogating application-layer companies that primarily create end-user software and workflows on top of existing foundation models, rather than training frontier models or operating large-scale AI infrastructure. That is because the application layer is a specific bet European policy has chosen to place.

The application layer

Lovable is an example of a broader trend in European tech strategy. It is based on an assumption that Europe’s opportunity in AI lies in building tools that make AI useful at the application layer (rather than spending billions on building deep technology, compute infrastructure and parallel ecosystems). The application layer is where most of the profit will be made, so the argument goes, and focusing on applied AI helps Europe stay in the ‘AI race’ by essentially piggybacking on the investments of others (this 2025 op-ed by SAP’s CEO in the Economist and these comments by Skype founder-turned-venture capitalist are prime examples for this position).

This idea is genuinely tempting. The bet (call it “sovereignty-at-the-application-layer”) provides a profitable way forward without playing the prohibitively expensive game of competing at the frontier-model layer. Asset-light software companies, with low headcount and few hard, illiquid assets, are a far better fit for how venture capital is built to allocate money. It also pattern-matches to how the internet era seemed to play out. From travel platforms like Booking.com to streaming services like Netflix, major winners of the internet economy emerged at the application layer.

The idea is also alluring to policy-makers. “Applied AI” first entered the European policy world with von der Leyen’s Political Guidelines for the 2024-2029 Commission, which set out the intention to build an Apply AI Strategy to boost new industrial uses of AI.

But the policy argument conflates financial efficiency with strategic autonomy. There are three serious arguments that speak against the long-term sustainability of the application-layer bet for Europe: creeping supplier competition, loss of control and value capture upstream

Suppliers as competitors and ultimate beneficiaries

Anthropic competing with their client, Lovable, isn’t an isolated incident. There’s been an industry-wide trend of model-makers entering the application market, driven by the need to capture value beyond the raw model provision. Claude Managed Agents (alongside Claude Cowork) are a direct answer to the self-hosted infrastructure for autonomous agents developed by OpenClaw (the free and open-source AI agent that can run on device). OpenAI tried to acquire coding assistant Windsurf (which was ultimately “aqui-hired” by Google in a multi-billion dollar deal).

Not just model makers, but also dominant hyperscalers are both suppliers and potential competitors for application-level start-ups. Amazon, Google, and Microsoft all compete at the application layer directly, Microsoft on office and productivity software integrated into Office 365, Google across search and productivity, and AWS now also on industrial applications of AI. All three companies also sell AI as a service bundled into their cloud platforms, making it excruciatingly difficult to compete with these players.

This is what distinguishes AI from earlier software eras. Companies like Salesforce and Adobe could build lasting application-layer franchises because their suppliers sold commodity inputs and had no designs on their markets. AI applications rent their defining capability from a handful of firms that are expanding into applications themselves, and that set the marginal cost of every customer interaction.

If all of this sounds familiar to astute historians of Europe’s relationship with dominant tech companies, that’s because it is. Europe already has its application-layer success story, and it is a tenant. Spotify, its most celebrated consumer-tech champion, runs on someone else’s cloud, hands most of its revenue to the rights holders it cannot do without, and has spent years fighting Apple over App Store fees, payment restrictions, and market competition. Succeeding at the application layer and achieving independence turn out to be different things.

Application-layer AI companies inherit an even harder version of this. Their marginal costs are rigid and set upstream by model providers. And while they could theoretically switch to open models, once these reach parity with proprietary ones, leaving is hard, often by design, and the financial and infrastructural dependencies that are being built are hard to disentangle. As a result, these companies bear far more resemblance with a different corner of the internet economy: the flight-comparison sites that ran on Google traffic until Google Flights displaced them; the publishers who reorganized around Facebook’s algorithm and pivoted to video, until court filings revealed the engagement figures had been vastly inflated by Facebook; and apps like Spotify that remain subject to Apple’s ever changing App Store terms, and a cut of up to 30% on many transactions.

Today, we’re seeing this exact dynamic play out at a much larger scale. European AI companies building on American APIs risk becoming something like an outsourced R&D department for leading model providers: they take the risks of discovering which applications actually work, while the value of those discoveries flows back down the stack. The moment a European startup comes up with a seriously profitable idea, the model makers have every incentive to acquire the company or simply rebuild the product as a feature.

This is bad for Europe’s long-term independence, but for the typical venture capital cycle, selling early to a wealthy incumbent is an attractive way to reach exit liquidity — and resisting that pull is hard for a startup stacked with VC. Lovable currently has 54, mostly venture investors in its cap table, with Google’s growth venture arm CapitalG leading the latest funding round. Instead of battling for market share against stupefyingly wealthy incumbents, selling to these companies provides a successful exit for the early investors and a small team of founders. This points to the tension between the strategic objective of technological sovereignty and the logics of venture capital financing. While the European AI ecosystem has notable exceptions, this typical VC model still skews startups’ trajectory towards the US ecosystem.

It is often noted that European startup founders invest in a new generation of founders, creating a flywheel. But this flywheel might never spin free of the structure of which it is part. It might lead to a cycle of serial exits to the balance sheets of US tech ecosystems, without generating enough escape velocity to create alternative gravity wells that could serve as a scalable platform anchoring European companies to pursue pathways. Against this gloomy picture, a counterargument is routinely offered: European companies can protect themselves through proprietary data, scale up fast enough to become dominant themselves, expand through mergers, or form a captured user base into long-term contracts. Moreover, the irreplaceability of context-specific data generated by proprietary know-how and workflows of vertical industries and incumbents is invoked as the insulating layer protecting the European companies from US competition, and creating staying power in the European ecosystem

But these mythical, vertical-specific data lakes in manufacturing or healthcare are not available for horizontal startups such as Lovable, which have to depend on customer traction, user experience, and speed as their differentiating factors. These are relatively weak moats compared to the financial dominance, ecosystem control, and integration advantages of existing platforms that control the engine on which these corporations depend. The first-mover advantages are real, as OpenAI testifies. However, the first to move is not necessarily the last to survive.

No control over your inputs

Infrastructure and model layers competing with application-layer start-ups that depend on them isn’t the only reason why focusing on the application layer is risky. When delegating a core part of their infrastructure, companies also delegate control over pricing models, and APIs can change overnight, while model behavior can be altered without notice. This leaves a fundamental vulnerability at the heart of an API-based business model. Recently, for instance, Anthropic shifted from a flat rate to usage-based pricing, destabilizing application builders downstream.

Many applied AI companies thus stake their hopes on staying “provider-agnostic”: switching between models rather than tying their products to one. The bet is intuitive, but it leaves them exposed on two fronts at once. On prices, API costs are currently depressed sector-wide as providers still fight for market shares. Anyone building on top of these APIs is heavily exposed once prices reflect real costs, or when the model layer as a whole starts to capture value and is able to extract rent. On performance, reaching the absolute competitive frontier for a given use case may require model-specific optimization. An agnostic start-up has to forgo these optimizations to prevent lock-in

A counter-argument is that orchestration layers that sit between the app and several model APIs could make switching between models more seamless — and many start-ups, such as the German langdock, now offer exactly this. But that raises the question of who sets orchestration standards, and AI companies at the application layer would still be dependent on multiple, though now competing, oligopoly of model makers.

The most serious counterargument is open-weight model parity. Simply put, if open-weight models reach a good enough quality, on par with leading, proprietary frontier models, then application-level AI start-ups can self-host competitive models and build on top. But the largest open-weight models are still released by a small number of AI labs, which means dependence on these models being released under truly open licenses. Most open-weight models are not actually open in the sense of open-source, and even maximally open models do not by themselves guarantee competition, as the inputs needed to train and deploy large-scale AI remain concentrated in the same handful of corporations, who have a history of using the rhetoric of openness to entrench, rather than dilute their dominance. While open-weight models do shape the bargaining position, the ecosystem is an integrated whole – the distribution layer can shape model and infrastructure choices of application companies squeezed in between.

How these technical developments and advances will unfold is still uncertain. But crucially, we caution against betting that any of these technical changes will automatically solve the underlying political economy that takes into account the broader ecosystem in which those models live.

Economic value flows upstream

From an industrial policy perspective, the question of where value flows is perhaps the most important one, and the one that’s often sidelined. The Commission’s report supporting the Apply AI strategy, for instance, counts over 4,100 AI startups in the EU with a combined enterprise value of €161 billion. But it counts startups, not supply chains and tracks where AI companies are headquartered, not where value is created or captured.

The ambition is to enable the next Google, or at least the next OpenAI. But Google’s power was never the product at the top of the stack, it was owning almost everything beneath it, from chips to cloud to models to data. And the leading proprietary models are each tied to one of the dominant hyperscalers, for instance, OpenAI with Microsoft and Amazon, and Anthropic with Google and Amazon. As a result, any start-up that builds on leading, proprietary models incurs not just ongoing costs for tokens, but also for inference and infrastructure. For application-layer companies, those costs consume a substantial share of revenue, which flows back to the US model providers and cloud platforms. The access to distribution might also come with strings attached – for instance, Lovable’s Google Cloud deal comes with expanded access to Google’s own Gemini models and Anthropic’s Claude models, where Google (via its investment arm Capital G) is a significant investor. To support their expansion of market share, they are motivated to direct token demand to these foundation models by using their investment partnership to determine model choices. As a result, control of margin sits upstream, not at the application layer, vulnerable to being squeezed at will by their core technology provider.

What this means for EU AI industrial policy

This dynamic changes a few core assumptions that are often implicit in EU AI industrial policy: Simply increasing the number of AI start-ups in Europe will not necessarily lead to more independence from US tech. European AI start-ups at high valuations don’t necessarily mean that Europe is building a sustainable AI sector that is here to stay. A thriving scene of European AI start-ups doesn’t automatically mean an economic upside for Europe. And more private investment, especially money bound by the logic of venture capital, does not automatically produce gains that stay in Europe.

The problem is not that European founders build on the best available models. The problem, for the long-term sustainability of Europe’s economy, is that countless rational and perfectly reasonable individual choices further entrench Europe’s dependency in ways that may be even harder to unwind than the status quo. The seeds of European dependency lie not only in the coercive force of the monopolies of the digital economy. They are also sown as a byproduct of ideas that appear self-evident enough to be left unexamined.

When Europe accepts that European companies should build on market-leading foreign partners for fear of being left behind; when hyperscaling venture capital is elevated as the sole viable financial driver of European tech; and when the spirit of Silicon Valley startup handbooks — built in California, by California, and for California — is carried as the absolute model to emulate, the path is largely set.

The question is not whether Europe can and should build applications on top of existing models — it is rather to note that applications alone do not solve the question of technological sovereignty that Europe so desperately seeks.