AI risks are accelerating even as companies and governments have powerful reasons to keep competing: leading labs have warned about frontier development following reported AI-system intrusions and possible bioweapons-related use, while commercial investment and national-security rivalry make unilateral slowdowns difficult.

Past arms-control deals have generally worked when countries had strategic parity or a weaker state accepted limits because it could not win the race; agreements including SALT II, START, New START and the Anti-Ballistic Missile Treaty placed comparable constraints on both sides.

Against that backdrop, US policymakers are considering an AI safety agreement with China, but reportedly want to widen the US lead by as much as two years first and use tighter export controls as leverage; China is estimated to have about 4% of US AI compute, rely heavily on smuggled or remotely accessed US chips, and trail US labs by six to 12 months.

The White House Accord on Super Intelligence, signed on September 29, establishes internal safety controls, oversight and audits for AI models, but leaves out autonomous weapons, rogue cyber agents and AI-powered drones, and does not legally require signatories to act.

The conflict in Eastern Europe highlights the physical stakes: Russia is described as up to four years behind the United States and China, while Ukrainian drones reportedly damaged three Yandex data centres using Nvidia A100 chips; the drones were said to be AI-powered autonomous weapons.

Voluntary safeguards also face commercial pressure: Anthropic’s leaked filing showed an $8 billion operating loss in 2025, while Chinese AI companies were contending with falling share prices and attempted IPOs.

Even so, an agreement may be possible because China has expressed serious concern that AI could threaten political stability, and both powers could cooperate if they genuinely prioritize reducing existential risks and avoiding an uncontrolled race toward autonomous warfare.