ISSUE 015 · The Aroko Letter

Full Speed Ahead

When frontier labs keep asking to slow AI development and nothing with the authority to do so has taken effect anywhere, can international cooperation step in?

SEP 21, 202614 min readPrimary-source verified

REGULATORY & POLICYGEOPOLITICSAI INFRASTRUCTURE

Eleven times since March 2023, nine of them from the people running the frontier labs, someone has asked for frontier AI to be slowed down. More than 300 model releases later, nothing anywhere confers the authority to delay, cap, or halt a training run. This issue reads the instrument record directly — the EU AI Act, China’s generative-AI measures and amended Cybersecurity Law, the US reporting default that was revoked in 2025 — then tests it against four restraint regimes that did hold (Montreal, the Limited Test Ban Treaty, the Biological Weapons Convention, the IAEA) to ask what those had that this does not. The answer is a substitute the incumbents could trade, which leaves power, chips, and grid connections the only brake that binds.

Anchor figure

Eleven public asks to slow frontier AI since March 2023, nine of them from frontier-lab leadership, and 300+ model releases in between: instruments with authority over a training run, zero.

Key takeaways

No instrument anywhere — US, China, or EU — currently confers authority to delay, cap, or halt a frontier training run. Reporting duties, transparency requirements, and voluntary commitments are not that authority, and the one US reporting default that set a compute threshold (EO 14110) was revoked in January 2025.
Every historical restraint regime that held had a substitute the incumbents could trade for it: DuPont had an alternative to CFCs before Montreal turned. Frontier labs have no substitute for frontier models, which is why eleven asks have produced no rule.
The EU AI Act obliges the Commission to publish a list of general-purpose models with systemic risk under Article 52(6) — the cheapest available route from a private count to a shared one, and still unpublished.
The multilateral machinery is thinner than the asks assume: the WTO Appellate Body has been without quorum since December 2019, and a January 2026 presidential memorandum directed US withdrawal from or defunding of sixty-six multilateral bodies.
The forecast is not that AI goes ungoverned — unilateral and bilateral instruments keep accumulating, and export controls remain the longest-running active tool. It is that through 2030 nothing acquires authority over the training run, and nothing binds the US and China together.

Where it sits in the series

Closes the loop back to Issue 13’s physical ceiling: if no rule reaches the training run itself, the constraint actually pacing capability is the Grid-Silicon Order — power, chips, and grid connections.

READ THE FULL ISSUE Published in full, free, on Beehiiv.