The recent call from industry leaders of leading AI enterprises to “pace the frontier” of AI development are driven by genuine, well-founded concerns about the trajectory of the technology.
Yet, a closer look at these proposals reveals fatal structural flaws. The most prominent one is the ‘Supremacy Paradox’ that gives the perception of ‘Hitting the Brakes to Win the Race’. The proposal simultaneously argues that democratic nations must maintain unassailable technological supremacy to enforce global safety norms, while also demanding an internationally coordinated slowdown in development. This is a geopolitical contradiction. It asks foreign adversaries and allies to halt progress while explicitly declaring our intent to dominate them. The proposal further ignores the fiduciary pressures driving the AI market. To maximize the societal value of AI safely, we must first recognize that the current blueprint for pacing the frontier breaks under its own mutually exclusive assumptions, where good intentions collide with broken mechanics.
The article We Must Pace the Frontier1 drew strong support from two other business leaders2,3 at the forefront of AI development. The article brings up serious risks in the current trajectory of progress in the development of AI including recursive self-improvement, biosecurity threats, or rogue agentic swarms require urgent, serious oversight. The article highlights immense benefits from AI, including difficult to achieve goals such as curing most diseases.
Supremacy Paradox, Global Cooperation and Enforcement
The essay warns that not building AI places the technology directly into the hands of authoritarian powers, yet his proposed solution relies heavily on achieving “global coordination” with authoritarian nations like China to establish safeguards. China has been explicitly called out in that article by Amodei as that country is openly the most advanced in AI besides the USA. Relying on an unenforceable international treaty with adversaries to pause AI development is a fragile strategy.
In the prevailing climate, making treaties requiring this level of cooperation is going to be far more difficult than it has ever been – leave alone enforcement. However, a treaty if a necessary step. Historical arms control agreements for nuclear or chemical worked largely because development cycles were slow and capabilities were physically verifiable – neither of which applies to training neural networks. We do not yet have an effective way of knowing if those controls have any effect at all for biological weapons. Even treaties for the heavy hardware based nuclear capabilities have not been fully effective. We may not know about certain developments in chemical or biological weaponization until a negative event, or may never know the certain that a harm occurred from violation of one such treaty.
Stated differently this is problem of diplomatic realism than technical safety. A successful international arms control regime functions because it offers mutual security, not unilateral hegemony. If the subtext of the American proposal is that the world pauses together but the US remains the supreme leader, the framework fundamentally fails the basic test of international relations. Allied nations will likely offer public support in diplomatic forums while quietly accelerating their own sovereign compute capabilities to avoid being permanently locked out of the frontier.
If democratic nations pace their progress while state-backed foreign adversaries do not, the pause only works to transfer geopolitical leverage. Moral frameworks transfer easily on paper, the hard leverage to enforce them disappears the moment you fall behind. This becomes the classical case of prisoner’s dilemma.
Asking foreign adversaries to halt progress while explicitly declaring our intent to dominate them is a near certain failure strategy. Even to nations allied to the US this would reasonably read this as a call to (permanently) cede their technological sovereignty to a handful of American corporations.
The geopolitical reality also likely lies on technological supremacy – whoever wins the AI race have far higher leverage in dictating the global norms for its deployment. If Western democracies hit the brakes, they forfeit the economic and military leverage required to force bad actors into adopting necessary safety guardrails.
The solution most likely lies in developing sophisticated guardrails appropriate for this need. Guardrails through regulation are not sufficient for this, we would need technological innovation as well and proactive defensive posture to protect against harms of those who do not follow objective harm reduction by AI.
Leaky Bucket of Enforcement by Export Controls
The article proposes a strategy that assumes the West can hit the brakes on its own capability progress while artificially holding China back through chip export controls and anti-smuggling crackdowns. This is guided by the premise that democratic supremacy is a prerequisite for global safety.
Authoritarian regimes operate with state-backed opacity, allowing them to publicly engage in coordination talks while secretly accelerating dual-use military AI applications. A self-imposed pause in the West, paired with the assumption that adversaries are successfully suppressed, is a high-risk gamble.
Hardware restrictions and trade embargoes are historically temporary friction, not permanent walls. While export controls effectively delay adversaries in the short term, they also force them to invest massively in indigenous supply chains and alternative architectures. If democratic nations voluntarily impose a “speed limit” on their own frontier models, they give authoritarian states the exact window of time needed to adapt and close the gap. Safety and supremacy are not sequential steps, but deeply intertwined. The power to enforce global safety standards and responsible deployment norms comes entirely from holding an insurmountable technological lead. The moment that gap narrows—even if due to a well-intentioned pause for alignment research—the West loses the leverage required to compel bad actors to respect those guardrails.
Nations like France, the UK, and Japan are currently investing heavily specifically to avoid becoming digital colonies reliant on a handful of American tech giants. If the United States proposes an international pause where the underlying architecture cements American corporate dominance under the guise of a shared “democratic coalition”, allies will inevitably view it as a regulatory moat masquerading as a safety pact. No matter how closely allied with the US – will allies negotiate away its future economic engine and national security infrastructure in good faith.
Given these, one could easily recognize that a genuinely global pause is impossible if it is designed, from the ground up, to protect American market dominance at the expense of European and Asian technological sovereignty – however well-intentioned the goal is. We need to ponder how do we expect the US to lead and develop such a multilateral agreement that gives a raw deal even to our allies?
Structural Market Incentives
Tech giants have poured tens of billions of dollars into frontier AI and the data centers required to run it. Executives have a legal duty to shareholders to generate returns on those massive capital expenditures. An altruistic pause simply cannot survive contact with quarterly earnings pressures and the need to amortize billion-dollar compute clusters. We cannot ignore this fiduciary trap.
When current market leaders call for strict regulations, costly audits, or a pause on next-generation models, it inherently looks like they are pulling up the ladder. One can argue that asking the world to hit the brakes exactly when a few US companies hold the technological high ground functions as a regulatory moat, protecting them from open-source disruptors and sovereign AI efforts.
If one company authentically decides to halt training for safety alignment, they know their domestic rivals will immediately absorb their market share. If Anthropic pauses, OpenAI scoops up their enterprise contracts; if OpenAI pauses, Google sweeps the board. This dynamic ensures that voluntary coordination is inherently unstable. This is now a corporate prisoner’s dilemma and market dynamics mathematically prohibit unilateral disarmament. Genuine safety requires external enforcement, not a gentleman’s agreement among rivals.
Despite the well-intentioned and well-informed proposal in ‘pacing the frontier’ of AI development, the preceding sections highlight the internal friction points in the mechanics of accomplishing the goal. The core incongruence manifests as ‘geopolitical paradox’, ‘allied sovereignty friction’, and ‘fiduciary reality gap’. While they do not rule out a global treaty, they highlight a strong need to come up with alternative ways to achieve the protection from potential harms of AI. The alternative approaches will likely involve guardrails in different formats (regulatory innovation), technological innovation and building proactive defense capabilities.
1. https://darioamodei.com/post/we-must-pace-the-frontier
2. https://x.com/sama/status/2098811563415150910
3. https://x.com/elonmusk/status/2098789109980332057
=========
Post-publication edits:
One word was removed when highlighting how the public market works in this context, without any change in meaning.
In pharmaceutical part, these few words were introduced: ‘including novel pricing strategies’
