Anthropic outage forces European AI founders

Anthropic switched off its two newest AI models for every customer worldwide for 17 days in June. The US Department of Commerce mandated the outage, blocking foreign nationals' access to Claude Fable 5 and Mythos 5, and the company complied by cutting service entirely.

For a European AI startup founder, the immediate question is no longer just about product. It is what happens the morning the core model the entire business runs on becomes unreachable. A year ago, such a query might have seemed like theoretical risk management. Now, it is a pressing operational concern.

Stripped of political jargon, the concept of "AI sovereignty" presents a basic engineering question. A competent technical lead should be able to state what breaks if a key supplier vanishes and how long finding an alternative would take.

The landlord owns the building

Using leading models from companies like OpenAI and Anthropic is not inherently wrong. Their tools are remarkable, enabling small teams to achieve what once required large engineering departments. The problem emerges when an entire company's operations flow from a single account with one provider.

Governments impose restrictions. Providers retire or reprice models or silently alter terms of service. No malice is required; a startup can simply become a rounding error in another entity's compliance decision. For a 15-person company in Zürich, that is a likely scenario.

Required vigilance varies by sector. A financial services brokerage that cannot explain its contingency plans for a supplier outage may face serious regulatory scrutiny. This makes it essential to retain the ability to run models on internal infrastructure, continuously test available options, and plan for sudden obsolescence.

Open-weight models now make this contingency planning feasible. Options like France's Mistral, Google's Gemma, and China's DeepSeek and Qwen are downloadable and runnable on owned hardware. While a quality gap to the frontier models exists, it is narrowing. For many practical applications, the difference is irrelevant.

Very few European startups should train their own foundation models, as the economics are prohibitive. Instead, founders must clearly name their technical dependencies, specify what they would use as an alternative, and estimate the swap time. This is not grand strategy. It is basic operational housekeeping.

The three barriers to alternatives

Building independent European AI capacity consistently fails on three fronts: funding, infrastructure, and talent retention.

The funding challenge is acute. The author recounts co-founding a European company in 2020 and spending more time trying to raise capital locally than building the product. The team ultimately secured US investment after shorter meetings and a definitive yes. This funding gap is more critical for AI, where training a frontier model requires vast expenditure before any proof of success.

The infrastructure problem is more mundane and severe. An AI data centre is a facility full of hardware requiring massive, uninterrupted power. In much of Europe, securing a grid connection of that scale is a multi-year planning ordeal. Political declarations about AI's strategic importance do not accelerate the delivery of physical transformers.

Some progress is being made. The EU has established 19 AI Factories and 13 Antennas linked to EuroHPC supercomputers, explicitly prioritising startup and SME access to high-performance compute they would otherwise never reach.

The third barrier is people. Cities like London and Zürich host serious AI research clusters. Europe reliably produces excellent researchers, but a significant portion then accepts job offers in California or from the European offices of American firms. From a sovereignty perspective, the effect is similar.

What you actually own

Europe has many AI startups, most of which will never need to train a frontier model. The critical issue is determining what portion of their product they genuinely control.

If a company possesses only a polished interface, some prompt engineering, and a wrapper around a publicly available API, two uncomfortable questions arise. How quickly could a skilled team replicate it? What remains of the product if the underlying model disappears?

For numerous specific tasks, a smaller, open model is sufficient. Sometimes it is preferable, being cheaper and operable on owned hardware in a physical location.

Some companies must undertake the harder work of building models to ensure Europe is not merely a customer. This demands capital, power infrastructure, and retained talent. This list will remain unchanged in five years without systemic shifts. Europe also needs companies compelling enough to dissuade researchers from viewing a move to San Francisco as the default career progression.

The situation does not demand drama. It requires Europe to cultivate enough internal options so that a Saturday decision in Washington becomes an inconvenience, not a catastrophe. For founders, the process starts modestly. They must ask what they would run tomorrow if today's model stopped answering.

If formulating that answer takes more than a minute, there is no fallback plan. There is only a supplier and a hope.