The first piece in this series made one structural claim: the EU holds roughly 5% of global AI compute against the United States' 74%, and a jurisdiction that regulates a technology it does not produce at the frontier cannot set the terms on which that technology is made available. Everything published since has been an argument for or against that claim, conducted not in words but in five days of events. The argument is over. The claim won.
What Actually Triggered the Directive
The public story, an Amazon-flagged jailbreak in Fable 5, was real but not the proximate cause. Reporting from the Washington Post, corroborated independently by Wired, Korea JoongAng Daily, The Elec, and UPI, establishes the actual sequence: Anthropic submitted a list of roughly 111 partner organisations for early access to Mythos, then an expanded list of about 50 more. Reviewing that expanded list, White House officials identified a South Korean telecommunications company they suspected of undisclosed ties to China. Anthropic revoked that company's access the same day the concern was raised. The episode is what one administration official described to the Washington Post as having "badly damaged" the government's confidence that Anthropic could vet its own access list. The Amazon jailbreak, days later, became the public trigger for a directive whose private trigger was a trust failure that had already occurred.
The company has since been named, independently, by Wired and Korea JoongAng Daily: SK Telecom, South Korea's largest wireless carrier, an Anthropic investor since 2023, and a Project Glasswing partner. SK Telecom's on-record response, June 16: "It is absolutely untrue that there is any linkage with China." KT and LG Uplus, the other two Korean carriers, separately confirmed they had never held Mythos access at all.
Sources: Washington Post (June 15), Wired, Korea JoongAng Daily, The Elec (June 17)
This matters for sequencing, not for blame. It means the directive was not primarily Pentagon-feud spillover landing on an unrelated ally, as the previous piece in this series framed it. It was a separate, specific trust failure, compounded days later by a jailbreak dispute, inside a relationship the Pentagon dispute had already left with no slack. Two different mechanisms, same outcome: the government acted unilaterally, consulted no one, and the rest of the world found out from the company's own statement.
The Ledger, Five Days Later
Every actor with a stake in this outcome has now been heard from, or conspicuously not heard from. The pattern across all of them is identical.
Who Said What, and What It Actually Means
Read the ledger in order and a single shape repeats. Every actor capable of speaking has spoken, mostly to manage the optics of dependence rather than to change it. Every actor invoked as a threat has stayed silent, because confirming the framing does not serve them either. Nobody has produced an alternative.
Ruthless, Not Erratic
It is tempting to read the week's broader pattern, export controls on a close ally's AI access, a tariff threat against France over a digital tax, a reported 14-point framework agreement with Iran, a public statement that Russia "should make a deal" on Ukraine, all inside the same 72 hours, as incoherent. It is not. Dean Ball, a former Trump administration AI adviser, told Axios the export action sits in real tension with the administration's own prior policy of encouraging US AI export abroad. That tension is genuine. But a single actor pursuing one consistent objective through several uncoordinated agencies will produce exactly this kind of friction: Commerce wants export dominance, national-security-adjacent concerns want lockdown, and nobody is required to reconcile the two in public. The apparent contradiction is not evidence the strategy lacks a centre. It is evidence the centre does not bother to coordinate its own departments, because it has not needed to.
The allies are not surprised because the rule changed. They are surprised because they had assumed, without ever testing it, that the rule did not apply to them.
The Finding, Restated
There is no country to use as a basis, because there is no country, currently, with both the will and the productive capacity to act as an independent check on US frontier AI policy. The EU has the rhetorical will and a fraction of the compute. Canada has the clearest diagnosis on record and no frontier lab of its own. South Korea has real infrastructure partnerships and just learned they are revocable by a phone call it was not party to. China and Russia have the stated ambition and, for now, the same dependency problem in different form. What every non-US actor produced this week was theater: declarations of intent to diversify, summits, trusted-partner frameworks, sovereignty packages, each one a real political artefact and none of them a unit of compute. Theater is not free. It costs money, time, and in the EU's case, a regulatory architecture that increasingly governs a market it does not supply.
This is not a story that resolves with the next news cycle. SK Telecom's access dispute, the G7's trusted-partners scheme, the next jailbreak, the next directive, will all generate fresh headlines that look like new developments. They are not. They are the same structural fact, observed from a different angle, again.
Sources & Primary References
- Washington Post, 15 June 2026: White House officials identify South Korean telecom suspected of China ties on Anthropic's expanded access list (cited via Korea JoongAng Daily, The Elec, UPI summaries; original behind paywall)
- Wired, ~17 June 2026: SK Telecom identified by name as the company in question
- Korea JoongAng Daily, 17 June 2026: "White House officials pin Anthropic AI export block on Korean telecom"
- The Elec, 17 June 2026: US export controls raise questions for Korean telecom sector
- UPI, 18 June 2026: Anthropic opens Seoul office, signs MOU with Ministry of Science and ICT
- Khan (Kyunghyang Shinmun), 17 June 2026: SK Telecom on-record denial of China ties
- Semafor, 13 June 2026 (Reed Albergotti): original "China-linked group" report, single-sourced, disputed by Anthropic
- CBC, AP, Washington Post coverage, 14 June 2026: Mark Carney's full remarks en route to G7, Westport/Ireland
- Reuters, The National, Arab News, US News, iTnews, 17 June 2026: Ursula von der Leyen's G7 AI lunch remarks
- POLITICO, ~16-17 June 2026: Thomas Regnier briefing, "wary of stoking Anthropic row at G7"
- European Commission, 17 June 2026: 2026 State of the Digital Decade report (9% global chips market share, 5% ICT specialist labour share)
- Axios, 16 June 2026 (via Khan): Dean Ball on the tension with prior US AI export policy
- CNBC, Euronews, Wikipedia (52nd G7 summit): general summit context, Iran framework agreement, AI industry lunch attendance
- Companion pieces: The AI Governance Wars Begin and Three Months Ago, Forever
Read the Series From the Start
This piece assumes the structural argument made in the first two. Start there for the full picture.
The AI Governance Wars Begin → Three Months Ago, Forever →