Why Unfinished Rules Are Costing Europe Its AI Investment

Deck: Siemens, Aveva and other multinationals say Europe's AI rules are pushing investment toward the US, but the sharper problem is implementation delay, not the rules themselves.

Roland Busch didn't mince words. Standing at Hannover Messe on April 20, 2026, the Siemens CEO said most of the company's planned €1 billion industrial-AI investment would go to the United States. Not Europe. His reason: the AI Act and the Data Act "miss the mark." Five months later, the Financial Times gathered similar signals. Aveva, Prysmian and Novo Nordisk voiced related worries. One major US bank now rates countries red or green for AI deployment readiness. The FT framed the concern around "AI openness." Aveva's CEO spoke of countries' AI "maturity, readiness, openness and capabilities."

The claim worth testing is narrower than the usual complaint. Europe's 2026 AI investment problem traces less to what the AI Act says than to how incompletely it has been built out. Delayed harmonized standards. Undesignated national authorities. Deadlines renegotiated this year. That's a different thesis than "Europe over-regulates," and the evidence can confirm or break it.

What Openness Means to Builders

For anyone designing AI systems, "openness" breaks into a few distinct ideas. Model openness can mean open weights, permissive licensing, reproducible evaluations and, more rarely, training-data transparency. Infra openness means you can self-host on your own hardware. Run inference on a Pi cluster or a NAS box, rather than a hyperscaler's region. Regulatory openness is about the path from "we built this" to "we can legally sell this." These concerns aren't identical. They share one trait, though. All of them get harder to reason about when the rules keep shifting underneath you.

A prototype that runs cleanly in a lab can become a substantial compliance project. That happens the moment it touches a regulated use case. The gap between technical readiness and legal clarity is where hesitation starts.

Rules on Paper, Reality on the Ground

The AI Act's high-risk obligations were supposed to roll out through 2026 and 2027. That timeline slipped. CEN and CENELEC missed their August 2025 deadline for harmonized standards. Regulation 2026/1744, effective July 27, 2026, pushed stand-alone high-risk obligations to December 2, 2027. Product-embedded high-risk systems moved to August 2, 2028.

As of September 7, 2026, six member states had no single point of contact listed for AI Act enforcement. Austria was one of them, with twelve more designations pending final adoption. Austria's RTR AI Service Desk exists, but it's advisory. It isn't a formally designated authority. A founder hunting for the relevant supervisory contact or notified-body route still finds an unclear picture across much of the EU.

That's a different problem than strictness. Unfinished rules hurt investment more than strict ones do. Strict rules can at least be budgeted for.

How Capital Moves

Companies run a rough calculation. Time to approval. Ongoing documentation and audit cost. Whether the model stack is swappable. Where the GPU clusters and engineers already sit. And whether the rules will still look the same in eighteen months. Fines backstop this logic. Relevant violations can trigger €15 million or 3% of global turnover, whichever is higher. Prohibited practices reach €35 million or 7%.

Two honest caveats belong here. Most ordinary B2B software, analytics tools and local inference systems aren't automatically high-risk under the Act. America's capital lead also predates this regulation by years. Energy prices, deeper venture markets and compute access may explain as much as any rulebook.

Stanford's AI Index puts 2025 private AI investment at $20.92 billion in Europe. The US figure was $285.88 billion. That metric excludes public money and is skewed by mega-rounds like OpenAI's $40 billion raise. A Commission-hosted StepUp StartUps report found AI took 18% of Europe's VC pool from 2020-2025, against 34% in the US. EU investor participation in deals above €25 million has fallen to 26%. US and UK capital fills the growth-stage gap. None of this proves the AI Act caused the shortfall. It fits what executives describe: a scale-up stage that gets harder after the pilot.

Where Europe Still Has an Opening

A Commission-hosted report on open-source AI from December 2025 makes the sovereignty case directly. Adoption data back it up. Eurostat recorded 13.5% of EU enterprises using AI in 2024, rising to 20.0% in 2025. Enterprise appetite isn't the limiting factor.

InvestAI launched at the Paris AI Action Summit in February 2025. It targets €200 billion in public and private AI investment, with €20 billion earmarked for up to five AI gigafactories. The tender opened July 30, 2026, with construction planned for 2027. Whether that becomes real capacity depends partly on conformity and sandbox processes moving fast enough. Financing, chips and energy access matter too. A sandbox only helps if it produces a binding yes or no. Article 57 guarantees guidance and supervision, not market authorization. The national sandbox deadline just moved to August 2, 2027. Fixing that end-point is the cheapest move Brussels has left.

If You're Building in Europe

Design for portability from day one. An architecture built around one model provider and one region has already made a jurisdiction decision. Nobody meant to make it. Map high-risk triggers before a customer asks about them. Treat technical documentation as a living file, updated continuously rather than written once for an audit. Favor vendors with clear licensing and on-prem options. Treat self-hosting as insurance against regulatory whiplash. And if a sandbox pilot or standards consultation opens, join it. Specific technical feedback helps regulators more than generic cost complaints from people who haven't shipped anything.

Europe's real choice has little to do with safety versus growth. The question is whether to finish implementation properly or leave builders guessing for another two years. Siemens will keep writing checks in the meantime.

The Severity-Uncertainty Matrix

A simple way to read any regulatory environment is to score it on two axes. How strict the rules are, and how certain their application is.

• Strict and certain: slower, but budgetable. Builders proceed with eyes open.
• Lenient and certain: fast, good for pilots, thin ground once regulators decide something is high-risk.
• Strict and uncertain: today's Europe. The worst quadrant for capital, because cost and timeline are both unknowns.
• Lenient and uncertain: rare, and the most dangerous long-term, since nobody has priced in the eventual correction.

Chasing lenient regimes misses the point. Price uncertainty as its own line item, separate from severity, before deciding where to build.