Anthropic releases Fable 5. Throughout the day, the same question comes up, almost mechanically: “What about us? Where’s our equivalent?” It’s posed as an acknowledgment of being behind, sometimes as an accusation. It’s the wrong question to ask. Not because Europe has nothing to be ashamed of, but because it confuses two very different things: not doing something, and not yet having the means to do it. These are not the same problem, and they do not call for the same response.
Let’s define the terms. Without complacency, and without self-flagellation.
What a cutting-edge model is, and why it costs what it costs
A cutting-edge model is the very latest in research: the largest, the most capable, the one that pushes the boundaries of what’s possible at the moment it’s released. GPT-4, Claude, Gemini Ultra, and Fable 5 fall into this category. It’s not a matter of branding; it’s a matter of scale. And that scale comes at a cost that must be clearly stated.
Training such a model requires tens of thousands of state-of-the-art graphics processing units (GPUs)—Nvidia’s H100s or H200s—which are monopolized for months on a single training campaign. The computational cost of a state-of-the-art model runs into the hundreds of millions of dollars—sometimes more—excluding salaries and upstream research. Then there’s inference: running the model for millions of users costs even more, on an ongoing basis, and this cost cannot be covered by day-one revenue. Cutting-edge research is not an isolated intellectual feat. It is a resource-intensive, capital-heavy industrial activity that requires guaranteed access to scarce chips and a steady stream of funding that never dries up.
What’s Missing—and What Isn’t
That is the real issue. Four conditions make frontier research possible, and Europe currently meets none of them fully.
First, computing power: access to cutting-edge chips in sufficient quantities depends on a supply chain dominated by an American designer, Nvidia, and a Taiwanese foundry, TSMC. The capacity of data centers dedicated to cutting-edge training in Europe remains incomparable to that deployed across the Atlantic. Next, capital: cutting-edge research requires venture capital willing to lose billions over five to ten years before seeing any return—an approach that European funding, which is scarcer and more cautious, does not adopt on this scale. Finally, the domestic market: a product or service launched in the United States immediately reaches a homogeneous base of more than 300 million people, which helps offset costs. Europe remains fragmented by language, legal systems, and national markets. And the return cycle: an American or Chinese player can sustain years of losses because the promise of future dominance justifies the investment. The European ecosystem provides little funding for such a gamble.
The key takeaway can be summed up in one sentence: Europe is not lagging behind in cutting-edge research; it simply does not yet have the industrial foundation to carry it out. This distinction is not merely superficial. As for talent, there is no shortage of it. European laboratories train some of the best researchers in the field, and many leave precisely for places where that foundation exists. The problem isn’t in people’s minds; it’s in the ground beneath their feet. You can’t build the same thing with the same hands when the structural resources differ so greatly.
What Europe Can Do—and Is Already Doing
All too often, this assessment leads to a defeatist conclusion. Wrongly so, because the cutting edge is not the only playing field—nor even the most useful one for most organizations.
Mistral, in France, has shown this: it is possible to produce smaller, better-optimized, open models that cover the overwhelming majority of real-world use cases at a fraction of the cost of a state-of-the-art model. Research into efficiency, quantification, and local inference on modest hardware is an area where the disparity in resources matters far less. Next come vertical models: a model specialized in law, healthcare, industry, or a regional language does not need to be the largest in the world; it needs to be the best in its field and compliant with its regulations. This is an advantage of proximity, not a race for size. That leaves governance and standards, where Europe is ahead rather than behind: the regulatory framework, sovereignty criteria, and requirements for auditability and data protection shape the rules that models must follow to operate on the continent. Defining the playing field is a form of power that the race for metrics tends to overshadow.
Separating Use Cases to Ask the Right Question
Hence the only question that matters. Not “Where is our Fable 5?”, but “What do we really need, and for what risk?”.
For certain uses, dependence on a foreign cloud model is a serious risk. A critical application for defense, intelligence, or vital infrastructure cannot rely on a service that a foreign decision could shut down or restrict. The processing of sensitive data subject to sovereignty requirements calls for controlled inference, on domestic territory, under the appropriate legal framework. A heavily regulated sector needs a model whose entire chain it controls. Here, the question of autonomy truly arises, and the answer lies in effective, local, vertical models—not necessarily in a foreign competitor.
For many other uses, the issue carries less weight. Which of the organizations calling for a European GPT-5 actually has a use for it? Writing, summarizing, categorizing, and assisting with code: these tasks already run very smoothly on smaller models, and the risk of dependency is managed through portability and provider choice, not through the sovereignty of the model itself. Lumping all use cases together leads to the same flawed reasoning as pitting a low-cost hosting provider against a hyperscaler: we’re comparing entities that don’t follow the same risk logic.
AI sovereignty isn’t achieved by demanding a European clone of the latest American model. It’s built by separating use cases—to isolate risks—and by investing in the foundational infrastructure where it’s truly lacking. As long as we ask, “Where’s our Fable 5?”, we’re just talking about pride. The day we ask, “What use case? What risk? What level of dependence is acceptable?”—that’s when we’re finally addressing the issue.
Sources
- NIST, AI Risk Management Framework, nist.gov
- Stanford HAI, AI Index Report 2026, training costs and computing capacity
- Mistral AI, documentation on open models and local inference
- ANSSI, sovereignty criteria for AI services, March 2026
- European Artificial Intelligence Regulation (AI Act), implementation 2025–2026
- Epoch AI, estimates of computational costs for state-of-the-art models