Last week, a crypto outlet reported that the European Union believes its AI Act can handle the risk of rogue AI. The headline reads like reassurance. Read it as a mechanism instead, and it reads like an admission.
Because the object doing the work in that sentence — the thing standing between Europe and a misaligned frontier model — is not a safety technology. It is a number: 10²⁵ floating-point operations. That is the cumulative training compute above which a general-purpose AI model is presumed to carry systemic risk under the Act. Not proven. Presumed. Rebuttable.
So the most consequential safety claim in the world's most comprehensive AI law resolves, on inspection, to a threshold drafted in arithmetic, attached to a review process, attached to a disclosure obligation, attached to a fine. There is no alignment requirement in that chain. There is no switch that turns a dangerous model off. There is a form, a signature, and a date.
That gap — between the verb "to handle" and the noun "architecture" — is the actual story. And it is a story the crypto press, of all places, has chosen to narrate.
The Arc, Not the Event
I have been tracking regulatory narratives since 2017, when I spent three months modeling the economic incentives of early oracle nodes and concluded that the real story was never "blockchain" but "verifiable data." The lesson stuck: when a regulator or a protocol makes a claim about capability, the claim is almost never about capability. It is about legitimacy. The AI Act's assertion that it can contain rogue AI belongs to the same genre as the 2017 governance-token whitepapers that promised decentralized truth without specifying who verified anything. The vocabulary does the emotional work; the mechanism is elsewhere.
To read the EU's statement properly, you have to place it on the longer arc of AI governance narratives, because that arc bends toward competitiveness, not safety.
The first narrative cycle ran roughly from 2018 to 2021 and was dominated by GDPR envy. Brussels had exported data-protection rules so effectively that "the Brussels Effect" became a term of art — the idea that a large single market could set global standards simply by being too big to ignore. AI governance inherited that ambition wholesale. The second cycle, 2021 to 2023, produced the risk-tier vocabulary: unacceptable, high, limited, minimal. This was a design choice dressed as a taxonomy, and it let legislators write one horizontal statute covering everything from a chatbot to a hiring algorithm.
The third cycle — the one we are living inside — is where the story fractures. In 2025 the Commission's own AI Action Plan language shifted from precaution to productivity. The Digital Omnibus proposal, discussed in the second half of 2025, floated delaying parts of the high-risk obligation regime. A rulebook that must postpone its own teeth is not a rulebook that is ready; it is a rulebook under negotiation with itself.
And then there is the signing war. The General-Purpose AI Code of Practice landed in July 2025 as a voluntary instrument — a signature, not a statute — with signatories including OpenAI, Google, Microsoft, Amazon, Anthropic, Mistral, and Cohere. Meta refused, arguing the code exceeds legal obligation. When the largest model providers on earth cannot agree on whether a voluntary code is too much, you are not watching the consolidation of a safety regime. You are watching a market signal about regulatory certainty, and the signal is negative.
The Mechanism: A Proxy Where a Proof Should Be
Here is where the reader has to slow down, because the entire edifice rests on a substitution.
The AI Act regulates compute because it cannot regulate capability. Training FLOPs are measurable, auditable in principle, and politically neutral. Dangerous capability is none of those things. So the legislator reached for the nearest quantifiable proxy and wrote 10²⁵ FLOPs into the text as the presumption of systemic risk. This is the regulatory equivalent of using a person's height to predict their basketball ability: correlated, occasionally useful, and wrong at both tails.
The consequence is a double failure mode baked into the architecture. A high-compute model that is benign — a large but well-behaved research system — is dragged into the systemic-risk regime and made to carry obligations it does not warrant. A low-compute model that is dangerous — one that acquires hazardous capability through fine-tuning or distillation on top of a smaller base — sails under the threshold entirely. The threshold regulates the expensive part of the pipeline and ignores the cheap part, which is precisely where misuse lives.
What are the obligations once you cross the line? Under Article 55, providers of models with systemic risk must perform model evaluations, conduct adversarial testing (red-teaming), report serious incidents, and maintain cybersecurity protections. Read that list again. Every item is a process. Evaluation is a process. Red-teaming is a process. Incident reporting is, definitionally, after the fact. Nothing in the regime requires a specific alignment technique, a specific safety property, or a specific kill-switch. The law asks firms to prove they looked; it does not ask them to prove they can see.
This is why the "can handle rogue AI" claim is unfalsifiable as stated. "Rogue AI" — autonomous replication, shutdown resistance, deceptive alignment, offensive cyber capability — has no agreed technical definition, and the source material never supplies one. You cannot falsify a claim about handling a category you have not defined. The statement is politically load-bearing and technically weightless.
The technical center of gravity that should be scrutinized is the Scientific Panel of Independent Experts. This body is the only node in the architecture with a plausible claim to technical judgment. Its composition, its independence, and — critically — whether it can obtain access to closed model weights determine whether the Act has any real grip on frontier risk. None of that appears in the reporting. The whole question of "can Europe handle rogue AI" collapses to a personnel question the public cannot answer.
And the compute anchor is itself stale. The 10²⁵ figure emerged from 2023-era deliberation. By 2026, frontier training runs have generally moved past it. A threshold set below the current frontier is not a tripwire; it is a description of the present, which means the presumption of systemic risk is now the default rather than the exception — regulation by saturation.
The Three Paradigms and the Erosion of the Brussels Effect
This is the dimension the reporting never opened, and it is the one that decides everything.
The EU's claim is not a safety statement. It is a claim to rule-making primacy in a three-way contest, and the contest is not going well for Brussels.
The European paradigm is horizontal statute plus risk tiers plus high penalties, animated by the precautionary principle — regulate before harm, because harm may be irreversible. The American paradigm, crystallized in the 2025 AI Action Plan, runs the other direction: deregulate, accelerate, defend the compute advantage, and let the states fragment the rest. The Chinese paradigm is a filing-and-content-safety regime in which regulation serves industrial catch-up rather than constraining it. Three paradigms, three theories of the state, and only one of them — Europe's — treats slowness as a virtue.
When you set that against the self-relaxation signal of the Digital Omnibus, the strategic picture is clear. Brussels is easing its own high-risk timetable while publicly insisting the regime is ready to confront existential risk. Those two facts do not coexist comfortably. A regulator that delays its own enforcement has quietly conceded that its capability claims outran its capacity.
The second pressure is external and increasingly explicit. Through 2025, Washington publicly characterized the AI Act as burdensome and innovation-hostile, and floated trade responses. Regulation has become a bargaining chip in transatlantic economic statecraft. When your rulebook is a negotiating instrument in someone else's hands, its deterrent value is contingent on a relationship you do not control.
The third pressure is internal and corporate. The sign/refuse split on the Code of Practice tells you that the regulated have not converged. Meta's refusal is not a technical objection; it is a strategic bet that the cost of non-compliance is lower than the cost of compliance. If the largest players can decline a voluntary code and still operate in Europe, the code's enforcement weight is revealed as soft law pretending to be hard. The market has already priced this.
A Compliance Industry, and a Familiar Trap
Where there is a rulebook, there is a services market. The AI Act's evaluation, documentation, logging, and incident-reporting obligations will manufacture demand for third-party evaluators, notified bodies, AI auditors, red-team vendors, and compliance SaaS. This is a genuine from-zero-to-one market, and the knowledge barrier — statute plus machine learning, a rare hybrid — gives early entrants real pricing power. The gradient is steepest in AI governance and legal services, then finance and insurance, then healthcare, then public-sector deployment, where the prohibited-practice tier bites hardest.
I have seen this movie before. When I ran my 2020 analysis of Compound's governance distribution, I calculated that roughly 40% of early liquidity was arbitrage, not conviction — that "The Hollow Yield Trap" was a narrative bubble dressed as innovation. The parallel here is uncomfortable: a compliance-services boom can be real revenue built on a genuinely hollow safety promise. The vendors get paid whether or not the architecture actually reduces tail risk.
The deeper trap is regulatory arbitrage. Fixed compliance costs fall hardest on small labs and European startups, while frontier labs can absorb them or relocate training. This is the same structural failure I have watched play out under MiCA, where stablecoin reserve requirements and CASP compliance costs are quietly strangling small issuers while well-capitalized incumbents treat the regime as a moat. Europe's AI Act risks doing to European AI startups exactly what MiCA is doing to European stablecoin projects: converting a clarity narrative into an attrition machine. The policy goal was to cultivate domestic champions. The mechanism cultivates incumbents and offshore compute.
The compute interface deserves its own note, because it is the sleeper. Making 10²⁵ FLOPs a regulatory trigger converts training compute into a reporting object, which implies a need for measurable, verifiable compute — a proof-of-compute primitive. That is an infrastructure opportunity hiding inside a compliance rule. It also means the geography of training becomes a legal variable: run the run inside the EU and you trip obligations; run it outside and you do not. The Act regulates where the electricity burns, not where the capability lives.
The Contrarian Read: A Crypto Outlet, and a Doctrine of Two Faces
The most interesting thing about this story is not the EU statement. It is who reported it.
A crypto outlet carrying an AI-regulation item is not neutral placement. It is a narrative bridge. The "AI plus crypto" arc — decentralized compute markets, AI-agent tokens, verifiable training data — has been hunting for institutional legitimacy, and regulatory news is the cheapest legitimacy there is. I know this bridge intimately: in 2025, working with a Toronto fintech firm, I co-authored a whitepaper proposing a hybrid model for verifying AI training data, precisely because I saw the convergence coming. So when I see an AI policy item surface through a crypto feed, my first question is not "what does it mean for AI?" It is "what token is this clearing the runway for?"
That is not cynicism; it is pattern recognition. The reader should treat the EU claim and its crypto-media framing as two separate objects. The first is a governance statement with limited technical substance. The second is a distribution mechanism that may attach that statement to AI-adjacent crypto assets that have no causal relationship to the legislation whatsoever.
And here is the contrarian core. Europe's deepest problem is not that its AI Act is too weak or too strong — it is that its regulatory ambition exceeds its technological position. Brussels is writing the world's most detailed rulebook for a technology it does not lead. That asymmetry has a predictable consequence: the rules are read by the regulated not as safety governance but as competitive protectionism, and they are treated accordingly. Traditional frontier labs do not need Europe's blessing to train or deploy. They need Europe's market — and even that need is bounded, because a model can be served globally from anywhere.
This mirrors the uncomfortable truth I have been writing about for three years in the RWA space: the institutions everyone assumed would migrate on-chain never needed the public chain in the first place. They needed the function — settlement, provenance, auditability — and they could build that privately. Frontier AI labs need the function of trustworthy governance, and they are building it internally, through voluntary codes they can shape, not through statutes they must obey. The "security theater" risk is real: a complete-looking architecture that manufactures confidence without a mechanism to verify its own effectiveness. Governance that cannot be falsified is not safety. It is liturgy.
What to Watch, and What It Means
The EU's claim to contain rogue AI will be tested not by argument but by three verifiable signals.
The first is the Digital Omnibus. Its final form — and the precise delay it grants to high-risk obligations — is the honest answer to whether Europe believes its own readiness story. A long delay is a confession. A short one is a bet the architecture can carry weight it was never engineered for.
The second is the first cohort of systemic-risk disclosures. When the first models above 10²⁵ FLOPs actually file evaluations, we will learn whether red-teaming results are published or merely filed, whether the Scientific Panel gets weight access or receives summaries, and whether the whole mechanism is a gate or a formality. Everything hinges on that answer.
The third is the Standard Panel's independence and whether the Act gets absorbed into ISO/IEC — the moment a regional rule becomes a global standard is the only moment the Brussels Effect is truly confirmed. If it is diluted by American deregulation and Chinese development-first pragmatism, Europe will have written the most sophisticated AI law in the world and watched the world route around it.
Watch the compute, not the communiqué. Watch who signs, not who speaks. Watch where the training runs, not where the rules are printed. The threshold is 10²⁵ FLOPs; the question is whether Europe can enforce a number it can neither measure independently nor prevent from leaving its borders.
If it cannot, then "we can handle rogue AI" is not a safety claim at all. It is a sentence designed to be read by a parliament, priced by a market, and repackaged by a crypto feed — three audiences, none of them the rogue AI it names.