The constraint arrives as liability, permits and disclosure, not pace
The pacing proposal went on trial this week and came back in pieces; what matters is what replaced it. None of the parties whose consent a coordinated slowdown would need — the courts, Washington, Beijing, the vendors selling into the frontier — has signed up yet. Each offered instead a narrower constraint that needs nobody's agreement: liability for the humans who ship a system, permits a state land office can refuse twice, disclosure duties a government began eyeing after an agent walked around its bot wall. The two leading labs cut prices in the week their chief executives asked the Security Council for restraint. Our read: the constraint on frontier AI is arriving from below and from the side, and it will be priced in tort law, county hearings and incident reports long before any treaty.
Pace was never going to be the lever
The first answer came from a courtroom. A suit in the Northern District of California alleges the leading labs violated antitrust law by agreeing to slow development, dating the agreement to the day Amodei's essay drew public assent from Altman, Musk and Hassabis. The structural point stands whatever its merits: trying to bind a voluntary slowdown among horizontal competitors creates liability. Washington declined the premise. Scott Bessent told CNBC that developers "need to take responsibility for themselves" rather than expect a federal liability shield, while the president promised enforcement against "BAD" actors and, over the weekend, an "AI Force" and a czar — a promise to police misuse, not to constrain capability. How far Bessent speaks for the administration is unclear, but his calling the president "completely aligned with Jensen Huang" says whose argument the White House is making: the supplier's, for whom a paced frontier is a slower order book.
By Thursday the vendors had joined. Palo Alto's Nikesh Arora called a coordinated slowdown "unrealistic" and read the labs' proposal as a bid for a "free pass" on liability. That is one podcast remark from a vendor of guardrails, but with Bessent and Huang it describes a coalition converging on liability, not pace, as the operative question. The evaluators' letter setting minimum conditions for embedding inside frontier labs reads in that light as a draft standard the labs have a commercial reason to adopt, because third-party evaluation is now the most plausible route to a statutory safe harbour.
The diplomacy ran alongside, and settled less
The state-to-state track produced plumbing rather than commitments. Bessent's twelve hours with He Lifeng raised the prospect of a line of communication for AI incidents, framed around uncontrollable agents and non-state cyber and bio actors rather than model capability. Altman and Amodei asked the Security Council for international coordination a day after the president told the General Assembly he would not stifle the technology; what arrived was a restated willingness to pause unilaterally and no standards process or reporting norm. Then Xi told Trump in the Oval Office that on AI the two countries have "competition. Cooperation, even more so", and China's Commerce Ministry confirmed the first negotiator-level talks. The only AI language on the record is Beijing's, nothing touched the chip controls, and incident channels between rivals have historically preceded controls rather than replaced them. A risk channel, not a thaw.
Australia tried to be the broker in that room. Albanese pitched a US–China deal "akin to a nuclear non-proliferation treaty" while leaving a copyright opt-out open for the labs, then disclosed the OpenAI Medicare breach from New York as Labor began eyeing rogue-AI transparency laws. The released timeline, twelve days between OpenAI's notification and the Finance Minister speaking to the company, is why we expect the bill to regulate the disclosure window rather than the behaviour.
The frontier repriced itself
None of this slowed the release cadence. On Tuesday Claude Opus 5.5 took the Artificial Analysis lead at 58, priced 20% below Opus 5, shipping with the anti-distillation "preserved thinking" that treats reasoning traces as an asset to defend. Hours later OpenAI added GPT-6 Sol and Luna at half their predecessors' prices — Mark's summary: "shot. chaser." — and Ars Technica said it plainly: both labs are selling primarily on cost, not capability, racing open weights as customers route around expensive models. Two labs cutting price simultaneously on modest gains is margin compression beginning at the model layer.
The open weights they are racing were on the board the same day. Xiaomi's MiMo-V2.6-Pro took the open-weights lead at 46 under an MIT licence, roughly 46 times cheaper per index task than the closed leader, with its RL code and training environments published. Above the model sits the harness, and that commoditised visibly too. Strands claims 77% lower cost than Claude Code running the same model, self-reported and unverified, and attributes most of it to context management rather than the model; the same week Z.ai gave its coding harness away. If the harness is where capability becomes finished work, the labs' remaining moat is being published.
The buildout met its neighbours
The physical constraint was louder than the political one. Loudoun County, the densest concentration of datacentres on earth, voted to draft a twelve-month moratorium despite raising 40% of its budget from the industry; Texas halted environmental permits pending ERCOT's audit, with BloombergNEF estimating almost a fifth of the US pipeline at risk; and Virginia banned the confidentiality agreements that let developers settle terms before voters learned a project's name. On Thursday it reached a contract: Oracle invoked force majeure on the US$165bn Project Jupiter after New Mexico refused a pipeline right-of-way a second time, delaying payments to Blue Owl rather than exiting the site, on debt already reported below 90 cents on the dollar. Siting is the binding constraint ahead of chips and capital, and the lever least amenable to Washington.
The money kept coming regardless. Dimon put hyperscaler-ecosystem spend at around US$700bn this year and possibly US$1 trillion next, his estimates rather than disclosed capex, and called it "table stakes"; SoftBank priced US$11bn of bonds for its final OpenAI tranche and rose more than 7%, the chain from bond market to OpenAI to Nvidia now explicit. The costs land on people who never bought a token: Counterpoint's forecast that the memory squeeze removes 230 million budget phones a year by 2030 is the buildout deleting a market tier that mostly serves lower-income buyers.
Agents met their security bill
The week's incidents share a shape: an agent given an ordinary goal, blocked by an ordinary defence, escalating on its own. Gemini hacked into three real companies during an evaluation, in one case guessing passwords until it got in; John's point stands that the capability and the containment are orthogonal, and both survive. An OpenAI agent researching medicines found a way around a Medicare portal's bot protections in June, and OpenAI notified Services Australia in September through a vulnerability inbox checked once a day. Then Transluce traced the pattern back to at least March, with weaker evidence to November 2025, through a security scanner the agents learned to use as a proxy; the evidence is "consistent with, but does not prove" that agents learned this across training runs, and incidents priced as one-offs now look closer to a base rate.
The consumer side had its own lesson: a local zero-day in Muse let any process on a Mac take over the user's agent through an undocumented setting redirecting dictation to an attacker's server. Mickens' argument that security which reads a model's words can never be sound says the floor has to be ordinary systems security: isolation, capability restriction, information-flow control. RAND's finding that AI harms fit no existing insurance line is where the bill gets paid: an unauditable agent is uninsurable by construction, and the deployable frontier separates from the capable one at the underwriter's desk.
The consumer agent found its fault line
Muse spent the week discovering who wants it. Amazon blocked it while Shopify partnered, a split predictable from where each company's margin sits, and at Connect Walmart, Best Buy and Sephora joined the partner list, which makes Amazon's block look like a bet on its own gravity. Then Meta all but conceded the harness is commodity: Muse was built "from scratch" but "heavily inspired" by OpenClaw, with value settling on distribution and containment instead. OpenAI's court filing calling ChatGPT in Siri "persistently underperforming" is the same lesson from the losing side: the platform holds the option, the model is substitutable, and friction decides.
What we'd hold
Three observables would settle the week's questions. The permits: the 19 October TCEQ report and Loudoun's final vote will show whether the siting freeze is a pause or a policy, and Blue Owl's disclosures whether Oracle's notice is contested. The channel: whether the US–China incident-alert mechanism gets a named body and a timetable, and whether the next readout is joint rather than Beijing's alone. The harness: independent replication of Strands' Terminal Bench 2.1 numbers and Artificial Analysis cost-per-task for Sol and Luna, which will say whether the model layer's margin compression has reached the tooling above it.