The 2028 Clock: Why Anthropic Says the US Must Lock In Its AI Lead Now

Anthropic dropped a policy paper Thursday that frames the next four years as a decisive window for democratic control of advanced AI. The Claude maker argues the US and its allies have until roughly 2028 to cement a meaningful lead in frontier systems—or risk ceding the trajectory to authoritarian models built under Beijing’s direction.

The core argument is about compute. American and allied chip innovation, backed by export controls maintained across three administrations, currently gives democracies an estimated 11-to-1 advantage in effective computing power for training large models. Chinese hardware efforts, including Huawei’s latest projections, reach only about 4 percent of Nvidia’s compute level. That gap is the single biggest strategic asset the US holds.

But it’s leaking. Smuggled chips reach Chinese labs. Companies access overseas data centers stocked with restricted hardware. And distillation attacks—where adversaries harvest outputs from models like Claude to train cheaper copies—are a persistent back door. Anthropic publicly accused Chinese labs DeepSeek, Moonshot AI, and MiniMax of running massive campaigns earlier this year, creating tens of thousands of fake accounts to query Claude millions of times.

The paper calls for tightening chip export rules, increasing enforcement budgets at Commerce, disrupting smuggling networks, and making distillation explicitly illegal where possible. It also urges aggressive export of American AI systems to allies, faster domestic adoption to embed democratic values in deployment, and resilience against authoritarian influence operations powered by AI-generated content.

Critics see self-interest. Steven Sinofsky, the former Microsoft executive, argued on X that centralized control repeats failed Cold War thinking—innovation flows around restrictions, and openness beats tighter government grip. Chinese analysts quoted in the South China Morning Post called the paper irresponsible and self-serving, noting it ignores genuine Chinese innovation while serving Anthropic’s interest in barriers that protect its own models.

Still, the stakes are hard to dismiss. Frontier AI accelerates discovery in semiconductors, biotech, and materials science. It finds software vulnerabilities at machine speed and chains them into attacks. The People’s Liberation Army already procures commercial Chinese AI for military experiments, including autonomous targeting and cyber offense. A model that discovers novel exploits first in Chinese hands could penetrate US infrastructure before defenses adapt.

Anthropic’s message to policymakers is direct: act before 2028, close the gaps, and promote your own technology. A one-to-two-year lead then could compound, accelerating allied adoption and creating space for measured international agreements. Without it, competition turns neck-and-neck, safety takes a backseat, and authoritarian advantages in scale and control could prove decisive.