The AI Act did not create Europe's lag. It freezes the cost of it.

The familiar line is that Europe regulates while America innovates.
It is too easy, and partly false: the capital gap predates the AI
Act and has many causes beyond the rules. The real problem is more
precise, and more troubling. The AI Act adds a fixed compliance
cost to an ecosystem already short of funding. Incumbents absorb it
without blinking. For a new entrant, it can decide whether the
product exists at all.

The AI Act did not create Europe's lag. It freezes the cost of it.

In the spring of 2021, in Brussels, thousands of pages of
consultations circulated among lawyers, lobbyists, and civil
servants. At the same moment, in San Francisco and Shenzhen,
engineers were training models. The two scenes did not move at the
same pace.

It is a fair image. But it does not prove what it is often made to
prove. The European regulation did not, by itself, open the gap
between Europe and the rest of the world. That gap already existed,
and it owes as much to capital, cloud, semiconductors, market
depth, and public procurement as to regulation.

The real problem lies elsewhere. It is more precise.

The size of the gap

The Stanford AI Index 2026 figures are stark. In 2025, the United
States attracted $285.9 billion in private AI investment. That is
23 times China's estimated $12.4 billion, and 48 times the United
Kingdom's $5.9 billion. California alone captured $218 billion,
more than three quarters of the US total.

Stanford adds an honest caveat: these figures measure private
investment only. They understate China's effort, which is partly
carried by government guidance funds estimated at $184 billion
between 2000 and 2023. The 23-to-1 ratio is a private-capital gap,
not a total-spend gap.

So the gap is not only regulatory. It is financial, industrial,
and infrastructural. Keeping that in view avoids the trap of the
single explanation.

What a fixed cost changes

The AI Act's logic is defensible. Not all uses carry the same
danger. A system that decides a hire, a loan, or access to an
essential service deserves more scrutiny than a photo filter.

The problem begins when protection takes the form of a fixed cost.

For a high-risk system provider starting with no existing
compliance setup, a 2021 CEPS study of the original proposal
estimated the full creation of a quality management system at
between €193,000 and €330,000, plus roughly €71,400 in annual
maintenance. These amounts apply neither to every company nor to
every system. Some requirements overlap with ordinary practice.
But they reveal an asymmetry.

For a global platform, the cost spreads across millions of users.
For a twelve-person startup, it can decide whether the product
exists.

A name for the asymmetry

The economist George Stigler gave a framework for this mechanism.
A regulation designed in the name of the public can be shaped or
exploited by incumbents, who have the means to absorb its cost,
and sometimes to turn it into a barrier to entry.

Careful with the shortcut, though. More than 1,000 stakeholders
contributed to the AI Act consultations, including startups,
trade associations, and SMEs. Access was not reserved for the
large. But formal access is not real influence.

A multinational mobilizes lawyers, lobbyists, and engineers for
years. A startup delegates a few hours of its founder's time, or
joins a federation that speaks on its behalf. The door is open to
everyone. The weight behind the door is not the same.

The uncomfortable question

Is a candidate better protected when their CV is screened by a
large provider's system, fully equipped to satisfy the AI Act,
than by a smaller European solution, perhaps more transparent, but
priced out of the market by the fixed compliance cost?

The honest answer is: sometimes yes. A small company is not
automatically more virtuous, safer, or less discriminatory. That
has to be conceded.

But a regulation that concentrates the market also reduces the
diversity of solutions and the number of competitors able to
challenge the incumbents. You may gain in compliance. You lose in
plurality. And plurality, when it comes to algorithms that decide
for us, is not a luxury.

The delay does not erase the problem

The AI Act entered into force in August 2024 and applies in
stages. The most immediate prohibitions have been active since
2025, as has part of the general-purpose-model regime. In May
2026, following a Commission proposal from November 2025, the EU
agreed to push back the main high-risk obligations: December 2027
for Annex III uses, August 2028 for certain product-embedded
systems.

This delay is an explicit adjustment to implementation
difficulties. It does not prove the whole architecture was wrong.
But it does not solve the cost asymmetry either. It moves it in
time.

Other architectures existed

The United States leaned more on sectoral law, ex post liability,
agency powers, and state-level rules. China layered specific
regimes onto recommendation algorithms, synthetic content, and
generative AI, with registration and controls. Neither model is
liberal by nature, nor free of cost.

But their diversity proves one simple thing: a horizontal
compliance regime, imposed ex ante on every classified system,
was not the only conceivable option.

The real question was never whether to regulate AI. It is who
bears the cost and who captures the benefit. When the cost falls
mainly on those who do not yet exist, and the benefit on those
already here, the price of citizen protection deserves a second
look.

Europe has the researchers. It has the universities. It has not
yet decided whether it wants the companies capable of competing
with them.

SOURCES ───────────────────────────────────────────────────

[1] Stanford HAI, AI Index Report 2026, Economy chapter (2025 private investment: US $285.9B, China $12.4B, UK$5.9B; caveat on Chinese public funds). [2] European Commission, AI Act — application timeline and risk categories (in force August 2024). [3] Council of the EU, political agreement on the AI Omnibus, May 2026 (Annex III to December 2027; certain products to August 2028). [4] CEPS, "Clarifying the costs for the EU's AI Act", 2021 (QMS cost €193,000-330,000, +€71,400/yr, on the original proposal; methodological caveats). [5] Epoch AI / Stanford AI Index 2026 (notable models 2024: US 40, China 15, Europe 3).