Michael Darius Eastwood · Independent AI alignment researcher
Published
On 3 June 2026, the EU's tech commissioner Henna Virkkunen said, to CNBC: "We want to be sure nobody has a kill switch." Because the phrase "kill switch" is memorable and the phrase "we want to be sure nobody has" is not, the quote has been reported by headline editors as if it announced an EU kill-switch law. It did not. It announced the opposite. CADA's framing is anti-foreign-kill-switch sovereignty: the Union does not want any external actor (a foreign government, a foreign vendor) able to disable AI or cloud infrastructure that European institutions rely on.
The public draft describes three moving parts. First, a certification regime for cloud and AI service providers, layered by "sovereignty tier". Higher tiers require proof that the provider's compute cannot be unilaterally disabled from outside the block, and that its data and model weights are legally and technically insulated from foreign shutdown orders. Second, procurement rules that route sensitive workloads (defence, health, critical infrastructure, sovereign identity) to higher-tier providers. Third, disclosure obligations on providers about their upstream dependencies, so a customer can see which sovereign jurisdictions could, in principle, pull the plug on them.
The Fable 5 week (12 to 26 June 2026) demonstrated why the CADA framing is not a rhetorical hedge. When the US BIS ordered Fable 5 offline, European enterprise customers discovered in real time that they had no legal or technical recourse: a foreign government could remove a critical service they depended on, in hours, with no consultation. CADA is the pre-legislative response to exactly that failure mode. The "kill switch" being refused is the foreign one.
CADA still requires an enforcement layer. If a "sovereignty tier" claim is to mean anything, someone must be able to verify it, ideally at the hardware attestation level. That is why the same policy conversation keeps looping back to FlexHEG-class instruments (arXiv:2506.15093). Any mechanism strong enough to prove that no foreign actor can shut a cloud down is also, in principle, strong enough for an EU actor to shut it down. CADA does not resolve that tension; it names it and asks who holds the keys. That question is a democratic one, and it is a very old one.
Two litmus tests. First, does the piece quote the full Virkkunen sentence, or the "kill switch" fragment? Second, does the piece describe CADA as regulating providers, or as regulating end-users? A piece that gets both wrong is describing a different bill.
In the ARC/Eden convergence register, CADA is row 12, classified CONVERGENT, roughly five months after the January 2026 book that named "The Chokepoint" and argued that infrastructure-as-leverage would become an explicit policy vocabulary. CADA is the first legislative instrument that speaks that vocabulary in a democracy-facing document. The convergence claim is structural, independent institutions arriving at the same architectural conclusion.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.