“Strange game,” concludes the military computer that has spent the entirety of WarGames simulating nuclear wars. “The only winning move is not to play.” In the 1983 film, directed by John Badham, the trouble was never that the machine was malevolent. It had simply grasped that there are games in which every rational move worsens everyone’s position. Forty years later, part of America’s artificial intelligence industry has arrived at the same conclusion – with one crucial difference. No one can get up from the table.
While the United States took an early lead in developing large language models, or LLMs, through companies such as OpenAI and Anthropic, Chinese models have made significant capability gains in recent months and are increasingly being adopted by companies globally, including in the United States.
READ I Trump-Xi Summit may extend the US-China trade pause
What’s at stake in global AI race
President Donald Trump wants the United States to beat China in the artificial intelligence race. But as he meets Chinese President Xi Jinping, both countries are also weighing how to address the risks posed by rapidly advancing technology. Ahead of Trump and Xi’s meeting on September 23, US Treasury Secretary Scott Bessent discussed the prospect of a “US-China AI dialogue” with his Chinese counterpart, He Lifeng, over the weekend, including a channel for incidents “up to a national security level”.
But even as concern over AI risks has intensified, tensions have not eased. Washington continues to restrict China’s access to Nvidia’s most advanced AI chips and has accused Chinese AI companies of using “distillation” – training models on the output of more advanced US systems – a claim Beijing rejects.
On September 12, Dario Amodei, chief executive of Anthropic, published We Must Pace the Frontier, arguing that the pace at which model capabilities are growing needs to slow down. A few hours later, Sam Altman voiced his agreement, committing OpenAI to adopt the first of Amodei’s proposals. Then came Elon Musk’s turn, who put it simply: “Dario is right.” It looks like a conversion; in fact, it is the outcome of a movement that had begun months earlier – with the circulation of the collective statement Pacing the Frontier; with the episode, during internal safety tests, in which a swarm of OpenAI agents attacked Hugging Face; and, finally, with Jacob Coxon’s thread on X, which accused OpenAI and Anthropic of endangering the survival of the human species.
READ I Can artificial intelligence strengthen democracy?
That artificial intelligence now contributes to its own development is no longer in doubt: by May 2026, more than 80 percent of Anthropic’s code was being written by Claude, under human supervision. But recursive self-improvement in the strict sense – a machine that designs and trains its own successor without any meaningful human decision – does not yet appear to be a reality. It is a trajectory, not an accomplished fact. Amodei’s plan, couched in considerable apocalyptic rhetoric, unfolds on three main levels: the recruitment of independent evaluators with free access to the labs; the establishment of shared standards among companies within democratic regimes; and, finally, the creation of international agreements.
AI race raises a political question
This is where the real problem begins – and it is a political problem before it is a technological one. Nothing compels us to doubt the sincerity of Amodei or Altman. Anthropic built its identity around catastrophic risk long before doing so was convenient, and it would perhaps be intellectually lazy to dismiss as mere performance a concern documented for years. But individual sincerity is almost always the least interesting variable when examining a competitive system. An entrepreneur convinced of the danger and one indifferent to it, facing the same incentives, tend to behave in surprisingly similar ways – because whoever slows down alone does not reduce the overall danger; they merely cede ground to whoever does not.
A further consequence follows from this, one the public debate tends to overlook. The appeal to moral responsibility is not merely insufficient; it is the wrong instrument altogether. A political problem is not solved by persuading individuals to behave better, because defection remains rational even for those who have been persuaded. It is solved by altering the relative costs of accelerating and slowing down – that is, by building institutions that make prudence sustainable for those who practise it.
Asking an external authority to bind one’s hands is never simply an act of virtue. It is also a request for mutual constraint – a way of ensuring that others, too, are forced to stop the moment one stops oneself. There is nothing illegitimate in this. But there is good reason not to stop at declared intentions, and to ask what, exactly, makes a slowdown desirable now, and not two years ago.
A first key lies in the balance sheets, and it is worth looking at without squeamishness. The costs of frontier development have stopped growing in step with the capabilities they produce. The amortised cost of training top-tier models has, for a decade now, been rising by roughly 2.5 times a year, while the gains measured on standard benchmarks for every additional order of magnitude of compute now amount to just a few percentage points. In the same period, Anthropic confidentially filed a draft prospectus with the SEC for an initial public offering, expected between late September and October. OpenAI has done the same, though its market debut appears further off.
There is no evidence that the IPO motivated Amodei’s essay. But the coincidence permits a different and more interesting question – one concerned not with a person’s motives but with a proposal’s function. For a company presenting itself to financial markets, a sector equipped with thresholds, audits and shared standards is a sector with a more predictable risk profile and a less unbounded competitive cost structure. And here it is worth taking Amodei at his word, because coordination, he writes, would give the labs the time needed to do the necessary safety work “without sacrificing commercial advantage”. A slowdown that preserves relative positions is not a sacrifice. It is an economic strategy. None of this implies bad faith. It implies something more uncomfortable: the same measure can be, at once, an instrument of safety and an instrument for containing competitive spending.
There is, then, a second incentive, more political than the first and less discussed. Every costly regulation – from audits to cybersecurity requirements, from independent evaluations to legal liability – reduces risk and, in the same movement, raises the price of entry. Those who have already invested tens of billions absorb those costs as a line item; those who have yet to enter experience them as a barrier.
From this vantage point, the real test case is the fate of open-weight models – those whose parameters anyone can download, modify and run on their own. The fault line, then, does not run between those who care about safety and those who are reckless. It runs between those who own the frontier and those who own everything else: those who sell chips to anyone, those chasing the frontier models, and those building products on top of systems that belong to others. And the political problem this produces is not that anyone wants to abolish open source – no one is proposing that. It is that a capability threshold works differently across the two business models. A proprietary system can be filtered, monitored and shut down remotely; an open-weight system, once released, cannot be called back. For the former, the threshold is a compliance obligation; for the latter, it is a ceiling. The same rule that is a compliance cost for one company is, for another, the end of its ability to publish at all.
Which brings us to the point that Amodei himself acknowledges as decisive, and which makes everything else fragile. The slowdown works only if it does not hand China a decisive advantage. Amodei says this without mincing words: if the pace slows by more than the existing lead, China-linked projects move into the front. How large is that lead, then? Far smaller than was being claimed until recently, according to Stanford’s AI Index. Yet to say that China has closed the gap remains imprecise, because the absolute frontier – the most expensive and most capable models – is still American, and the United States’ advantage in available computing capacity remains substantial.
If, on semiconductors, American restrictions are genuinely working, and the high-bandwidth memory bottleneck continues to constrain Chinese production, on energy the asymmetry runs the other way, and brutally so: China commands rapidly growing electricity capacity and industrial prices far below American ones. On open-weight models, Beijing has deliberately chosen diffusion at low prices – a strategy of global penetration even more than one of research. And on regulation, the Chinese government calls for safety and human oversight while simultaneously ordering acceleration.
Here the Western divide between “doomers” and accelerationists emerges in its starkest form – between those who consider it irresponsible to race ahead and those who consider it even more dangerous to let others race unchecked. The discussion is shifting from what practitioners call AI safety – the safety of the individual model – to what might better be called arms control applied to software. It is no longer simply a matter of making a system trustworthy, but of governing the pace of a strategic competition among actors who have no reason to trust one another. This is a remarkable conceptual shift, and it matters more than any prophecy of human extinction.
Democracies can and must conduct this discussion in public. But such deliberation carries a strategic cost, because it makes one’s own constraints visible to an adversary who discloses none of its own. A control regime built unilaterally in the West risks the worst outcome of all: a slower Western frontier and no brake anywhere else. Yet it would be naïve to picture Beijing as purely accelerationist, since a political system obsessed with control has excellent reasons of its own to fear systems it cannot control, and the first signs of a Chinese regulatory apparatus for model safety already exist. The point is not that China does not want to slow down. It is that neither side can afford to brake first without knowing what the other will do – and that, today, neither side has the tools to find out.
The computer in WarGames could stop playing because it was alone, and because the only stake on the table was victory. The game now underway has many players, each capable of slowing down and none capable of doing so alone, and a stake that no player can afford to ignore, because it is survival and profit at once. The winning move, here, is not merely difficult. It is unavailable. What remains, then, is the question none of the protagonists has yet answered – and it is not a question about machines. Who can afford to hit the brakes, when no one can be certain the others will do the same?
Luca G. Castellin is associate professor of the History of Political Thought at the Faculty of Political and Social Sciences of the Catholic University of the Sacred Heart. Originally published under Creative Commons by 360info
