Artificial intelligence is becoming a new arms race, and to avoid repeating the mistakes of the past perhaps a drastic unification of competing models could prevent catastrophe.
A specter is haunting global geopolitics – the specter of unconstrained, autonomous intelligence.
Recently, the U.S. government restricted the release of Anthropic's powerful Claude Mythos 5 — a model capable of autonomously finding zero-day vulnerabilities in every major operating system — and limited the rollout of OpenAI's GPT-5.6 Sol to a mere 20 government-vetted partners. These actions, driven by the fear that such models could act as digital "nuclear bazookas," underscore a fundamental panic. As Thomas Friedman observed, Washington and Beijing have arrived at a modern "Nixon-Mao moment:" both nations now face a shared, non-human adversary in the form of out-of-control AI. Yet, instead of uniting against this common foe, both capitals are retreating into isolation.
Policymakers in Washington and a number of its long-standing democratic allies are attempting to hoard artificial intelligence. They rely on "geopolitical fencing" and strict export controls, with the aim of constructing a fortress that monopolises state-of-the-art models and contains the prospects for them to generate existential risks within an ideologically secure sphere.
Beijing pursues the same goal of AI sovereignty, but through a radically different, asymmetric strategy: technological independence via open-source proliferation. Through aggressively releasing highly capable, open-weight models to the world for free, China aims to challenge Western software monopolies, spur self-reliance, and demonstrate that there can be an alternative path to AI development without excess gatekeeping.
Yet whether through American fencing or Chinese open-sourcing, both superpowers appear to assume the same premise: that a sovereign nation can unilaterally control the trajectory, or contain the fallout, of intelligence – even as it becomes precipitously autonomous. This is nothing but a fantasy. In an era of decentralised proliferation, open-source architectures, and hyper-computation, the notion that frontier AI can be neatly ring-fenced is a hubristic thought – an ideological fortress against a borderless threat.
We stand at the precipice of an era that our political institutions were never designed to survive. The architecture of global governance remains anchored to the "Westphalian illusion" – the 17th-century assumption that absolute sovereignty and physical borders can guarantee security. For four centuries, we assumed threats could be filtered or repelled by customs agents, navies, and checkpoints. But how does a state patrol the border of a neural network? How does a government blockade an open-source repository, or a de-centralised swarm of autonomous agents?
The technologists of Web3 already proved that capital and data could be decoupled from the nation-state. Now AI is doing the same to cognition and autonomy, yet with existential stakes. If a jailbroken model unleashes a synthetic digital pathogen, it will not stop at a checkpoint; it infects the global nervous system instantly. If an unaligned artificial general intelligence (AGI) triggers an uncontrollable cascade in global financial algorithms, the distance of the Pacific Ocean would not protect Beijing, and the global economy will not be spared. We would no longer be merely competing nations; we are mutual hostages to the same technology.
This is not merely a secular, technocratic concern. In his recent encyclical Magnifica Humanitas, Pope Leo XIV warned that autonomous weapons systems have already advanced "practically beyond any human reach to govern them", urging the world to disarm the logics of 'armed' AI competition. When the Vatican and Silicon Valley are sounding the exact same alarm, the existential nature of the threat is undeniable.
Strip away the diplomatic rhetoric and look at the cold physics of computation in 2026 - a stark reality emerges. The AI frontier is not a crowded, multipolar race – even though one of us firmly believes that multipolarity remains possible in the implementation, adaptation, and diffusion levels of AI, further away from the frontier. Frontier AI is an industrial leviathan, demanding gigawatts of dedicated energy, the monopolisation of sub-nanometer semiconductors, and capital expenditures that dwarf the GDP of entire nations.
As of today, only the United States and China possess the macroeconomic and technological mass required to attain some semblance of intelligence that can replicate a vast majority of humanly productive tasks, and match humans in all productive forms of cognitive labour – characteristics that some have described to be innate to AGI. The European Union, the UK, and Canada, whilst well-endowed with human capital and exceptional academic researchers, are playing catch-up at present. This broad duopoly is not a political preference; it is an empirical fact.
To survive, the duopoly must evolve from a mutual hostage situation into a mutual survival pact. Today, the survival of our species demands that we forge a completely novel, supranational architecture: "One Union, Two Countries".
Crucially, this Union cannot be another United Nations: superseding authority in name only, lacking in real power. A committee debating non-binding resolutions cannot govern a technology that rewrites its own source code in milliseconds. "One Union, Two Countries" is not a diplomatic forum; it is a mechanism of raw governance, holding absolute authority over the narrow peak of frontier technology.
To be effective, it must strip away the subjective, politically motivated decision-making that plagues traditional diplomacy. Politics is too slow, and human leaders too prone to emotional miscalculation. The Union would operate as a structurally binding alliance at the apex of computational power, enforced by cold, unemotional logic. It requires a joint "Compute Interlock" – a cryptographically verifiable registry of all apex-tier compute. With hard boundaries and built-in mathematical triggers, we remove the fatal flaw of human hesitation: should either nation attempt to scale unilaterally beyond an agreed safety limit, the system automatically initiates pre-programmed economic and computational friction. We must replace the volatile arms race with a rigid, automated equilibrium.
We can already see the limits of traditional statecraft in the ongoing dialogues between U.S. Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng. While their discussions highlight the deep economic interconnectivity between the two nations, their toolkit is constrained to conventional trade diplomacy – a woefully inadequate response to a technology that threatens the foundational architecture of both economies. Only a mathematically enforced Interlock can provide the certainty that mere diplomatic "communication" cannot.
We cannot take our chances. We are nearing a permanent bifurcation point for our species. The illusion of a "third way" – muddling through with fragile corporate self-regulation and 19th-century statecraft – has vanished, as the panic over the latest models from OpenAI and Anthropic makes plain. The defining challenge of our era is no longer the struggle between American and Chinese dominance; it is the race between human continuity and algorithmic obsolescence.
Strategic realists in both capitals must recognise that a shared compute interlock does not diminish sovereignty – it has become the prerequisite for survival. To reject this architecture out of 20th-century national pride is, practically speaking, to defend the right to engineer a mutual doomsday apparatus. The isolated sovereign fortress has run its course. The rational imperative is clear: the United States and China must engineer an architecture of shared survival, for should we face our collective obsolescence apart, we shall fall.
