Four AI Chiefs Agree to Slow Development - and Spark Cartel Fears
Sam Altman, Dario Amodei, Demis Hassabis and Elon Musk's weekend proposal to "pace the frontier" divides observers between those seeing safety progress and those warning of monopoly tactics.
A Weekend Pact That Split the Industry
Over the weekend of 14 September, four of the world's most influential figures in artificial intelligence reached a loose consensus that sent ripples across Silicon Valley and beyond. Sam Altman of OpenAI, Dario Amodei of Anthropic, Demis Hassabis of Google DeepMind, and Elon Musk of SpaceX collectively endorsed the idea of slowing the pace of frontier AI development - what they termed "pacing the frontier."
The proposal they aligned with includes embedding third-party auditors within leading laboratories, regulating domestic AI research facilities, and pursuing a coordinated global slowdown in the race to build ever more capable systems. For proponents of AI safety, the announcement represented a rare moment of unity among executives whose companies have been locked in a fierce competition to ship the next breakthrough model. For critics, however, the timing and composition of the agreement raised immediate red flags about anti-competitive intent.
At Opentechwire, we have tracked the consolidation of AI capability among a handful of well-funded laboratories in the United States and the United Kingdom. The involvement of these four leaders - whose organisations command the largest compute budgets, the deepest talent pools, and the closest relationships with regulators - means that any informal arrangement they strike carries weight far beyond a simple policy statement.
The Proposal's Core Elements
The framework the executives endorsed revolves around three pillars. First, mandatory third-party audits of frontier labs, with auditors granted access to training infrastructure, model weights, and internal safety documentation. Second, domestic regulatory regimes that would require government approval before deploying models above a specified capability threshold, likely tied to compute or benchmark performance. Third, a multilateral agreement to synchronise these slowdowns across major AI-producing nations, preventing any single jurisdiction from racing ahead while others pause.
Advocates argue that such measures would create breathing room for safety research, allow society to adapt to the disruptions already underway, and reduce the risk of catastrophic misuse or loss of control. The idea of pacing development is not new; it has circulated in academic circles and among longtermist groups for several years. What is new is the public backing of executives whose companies stand to lose competitive advantage if they unilaterally slow down.
That apparent willingness to sacrifice speed for safety is precisely what sceptics question. Critics note that the four signatories already occupy dominant positions in the AI landscape. OpenAI and Google DeepMind lead in large language model deployment; Anthropic has carved out a niche in constitutional AI and enterprise contracts; Musk controls significant compute resources and has launched his own model efforts. A coordinated slowdown, the argument goes, would freeze the current hierarchy in place, making it vastly harder for new entrants - particularly those relying on open-source foundations or operating outside the United States - to challenge incumbents.
Cartel Accusations and the Open-Source Flashpoint
Within hours of the announcement, the term "cartel" began circulating on social media and in policy forums. The comparison is not merely rhetorical. In economic terms, a cartel is an agreement among competitors to restrict output, fix prices, or divide markets in ways that harm consumers and stifle innovation. The AI slowdown proposal, critics contend, functions as an output restriction: by limiting the release of new capabilities, the leading labs protect their existing products from obsolescence and reduce pressure to compete on features or pricing.
The open-source community has been especially vocal. Developers working on models such as LLaMA derivatives, Mistral variants, and other freely available architectures see the proposal as a direct threat. If regulators adopt the framework the executives support, open-source projects could face audit requirements and deployment thresholds that are trivial for a company with OpenAI's legal and compliance budget but prohibitive for a distributed team of contributors. The result, they warn, would be to cement proprietary models as the only viable path to frontier AI, locking out researchers in Bangalore, Shenzhen, and São Paulo who cannot afford the overhead of a compliance apparatus.
There is also the question of what "frontier" means in practice. The executives have not published a technical definition. If the threshold is set high enough, the slowdown might exempt most current systems, allowing the signatories to continue iterating on their existing product lines while blocking genuinely novel architectures. If the threshold is set low, it could capture a much wider range of research, including academic work and specialised domain models. The ambiguity leaves room for regulatory capture, in which the firms best positioned to influence standard-setting processes define the rules in their favour.
The Safety Case and Its Limits
Supporters of the pact argue that these concerns, while not unfounded, miss the scale of the risk. They point to the rapid capability gains of the past two years - improvements in reasoning, multimodal understanding, and agentic behaviour - and note that no robust framework for evaluating catastrophic risk exists. Third-party audits, they contend, are a minimal baseline for ensuring that labs are not racing blindly toward systems they cannot control. A global slowdown, meanwhile, reduces the chance that competitive pressure forces a lab to deploy a model before safety work is complete.
There is merit to this view. The incentive structure in frontier AI is brutal: the first lab to ship a new capability captures user attention, investor capital, and talent. That dynamic creates pressure to cut corners on red-teaming, adversarial testing, and alignment research. If all major labs agree to slow down together, the theory goes, no single player suffers a competitive penalty for investing in safety.
Yet the argument assumes good faith and transparency on the part of the labs themselves. The proposal does not specify who would select the third-party auditors, what powers they would have, or how findings would be disclosed. It does not address the possibility that labs might comply with the letter of an audit regime while continuing to push boundaries in less visible ways - through internal research, partnerships with hardware manufacturers, or deployment in jurisdictions with lighter oversight. And it does not resolve the fundamental tension between the labs' fiduciary duties to shareholders and their stated commitment to safety.
Regional and Geopolitical Angles
From an Asia-forward perspective, the proposal raises questions about who gets to set the global pace. The four executives are all based in the United States; their companies are subject to US export controls, sanctions regimes, and national security directives. A slowdown coordinated through Washington and echoed in Brussels or London would do little to constrain AI development in Beijing, where domestic labs operate under different incentives and regulatory frameworks.
If the proposal were implemented as a multilateral treaty, it would require buy-in from China, India, South Korea, Japan, and other major AI research centres. The history of arms control and environmental agreements suggests that securing such consensus is difficult and that verification is harder still. In the absence of a truly global pact, a US-led slowdown might simply cede leadership in certain domains - such as inference optimisation, edge deployment, or specialised vertical models - to labs in Shanghai or Seoul that face no equivalent restrictions.
There is also the risk of bifurcation. If open-source development continues in jurisdictions that do not adopt the slowdown framework, the global AI ecosystem could split into a regulated, audited tier dominated by Western incumbents and an unregulated, rapidly iterating tier centred in Asia and other regions. That outcome would undermine both the safety goals and the competitive concerns that motivated the proposal in the first place.
What Comes Next
The weekend agreement is not binding. It is a statement of intent, not a treaty or a regulatory filing. The executives involved have not committed to specific timelines, capability thresholds, or enforcement mechanisms. What they have done is signal a willingness to explore collective action - and in doing so, they have forced a debate about whether such action serves the public interest or entrenches private power.
Regulators in the United States, the European Union, and the United Kingdom are already drafting AI governance frameworks. The question is whether those frameworks will incorporate the kind of slowdown and audit provisions the executives endorsed, and if so, whether they will apply symmetrically to all labs or carve out exemptions that favour incumbents. The open-source community, meanwhile, is mobilising to ensure that any regulatory regime does not foreclose distributed, collaborative development.
At Opentechwire, we see this moment as a test case for how the tech industry negotiates the tension between competition and coordination. The AI slowdown proposal is neither pure altruism nor pure cartel behaviour; it is both, and the balance will depend on the details of implementation. The challenge for policymakers is to design rules that genuinely enhance safety without locking in monopoly power - a task made harder by the fact that the firms with the most expertise to inform those rules are also the ones with the most to gain from shaping them.
