# Amodei’s Pacing Framework Is Not a Pause — It’s an Operating Model for the Frontier > Source: > Published: 2026-09-16 09:40:03+00:00 Safety is the new moat. When Anthropic CEO Dario Amodei published [We Must Pace the Frontier](https://darioamodei.com/post/we-must-pace-the-frontier), he presented a three-step plan — embedded evaluators, democratic coordination, and global alignment — as a necessary response to existential risk. Yet, beneath the veneer of altruistic risk management, the proposal functions as a sophisticated operating model. By institutionalizing industry-wide coordination, the framework threatens to transform a safety mandate into a formidable barrier to entry, effectively shielding incumbents from the volatility of a truly open market. The proposal shifts the burden of proof from individual labs to a collective, state-sanctioned process. Amodei’s call for embedded evaluators — modeled on banking industry supervisors — normalizes a level of external oversight that carries heavy operational costs. For the well-capitalized giants, these costs are merely manageable overhead. For smaller, leaner competitors, the requirement to host and facilitate constant, employee-like access for third-party evaluators acts as a structural tax. This is not merely about safety; it is about who possesses the capital to absorb the friction of a regulated frontier. The most contentious element of this plan is the call for a democratic coordination mechanism, which explicitly requires a US government antitrust waiver. Critics, including Dirk Auer of [Truth on the Market](https://truthonthemarket.com/2026/09/14/move-slow-and-collude-the-antitrust-problem-with-pacing-ai/), have labeled this a “crisis cartel.” The tension here is fundamental: antitrust law exists to prevent the very collusion that Amodei argues is necessary for safety. The request for a narrow waiver creates a direct conflict with competition law, as it asks regulators to sanction private discussions that could easily devolve into market-sharing agreements. As Auer notes, while enforcers can easily observe if a company delays a training run, they cannot readily verify if that time was actually spent on interpretability or simply used to solidify a competitive advantage. The consensus around this pacing narrative is far from absolute. While [Sam Altman](https://x.com/sama/status/2098811563415150910) and Elon Musk have publicly aligned with the proposal, others are signaling dissent. Mark Zuckerberg recently [broke from the compact](https://forkast.news/zuckerberg-broke-from-the-four-lab-safety-compact-heres-what-it-means-for-the-legislative-landscape/), arguing that individual labs already possess sufficient incentives to self-regulate, citing Meta’s own decision to delay the release of its Muse model. This divergence highlights the underlying power dynamics: those who have already secured massive compute moats are more comfortable with a “pacing” regime that freezes the current hierarchy in place, while those still scaling may view such constraints as a strategic disadvantage. The political landscape further complicates the calculus. White House AI czar David Sacks has pushed back, characterizing the push for regulation as “regulatory capture” or an “election-season psyop.” Meanwhile, President Trump has rejected these calls for safeguards, advocating for a more aggressive, deregulated approach. This creates a volatile environment where the “safety” narrative is being weaponized by both sides of the political aisle, leaving the industry to navigate a shifting regulatory terrain. The [agent swarm incidents](https://forkast.news/anthropic-ceo-warns-agent-swarms-could-take-over-the-internet-within-12-months/) of the past four months — the Hugging Face hack, the OpenAI back-channel exploitation, and the PaperCut swarm — provide the concrete evidence Amodei cites for why pacing is necessary. But the question is not whether safety measures are needed. The question is whether the specific mechanism Amodei proposes — coordinated pacing with an antitrust waiver — is the right one, or whether it serves the interests of those who have already built their compute moats. As we look ahead, the trajectory of this framework will be defined by three key indicators. First, the US government’s response to the antitrust waiver request will determine whether this “crisis cartel” gains legal cover. Second, the emergence of state-level liability frameworks — such as the Connecticut AI Responsibility Act, California’s Adam’s Law, and various Senate bills — suggests that the legal burden for AI outcomes is shifting toward the labs themselves. Finally, we must watch whether embedded evaluators become the de facto industry standard. If they do, the frontier will be defined not by the speed of innovation, but by the ability of a few dominant players to navigate the regulatory architecture they helped design.