Two futures await the world by 2028. In one, democratic nations hold a decisive edge in the most powerful artificial intelligence systems. They shape global standards, drive economic gains and safeguard critical infrastructure. In the other, China achieves rough parity or better. Authoritarian models proliferate. Repression scales effortlessly. Military edges shift.
Anthropic laid out those paths in a detailed policy paper released Thursday. The San Francisco-based lab, creator of the Claude family of models, didn’t mince words. Without urgent steps to close loopholes in chip export rules and halt sophisticated theft of American AI knowledge, the U.S. risks losing its current advantage. The window, the company argues, narrows fast.
“If the US and its allies act now to address both issues, it may be possible to lock in a 12-24 month lead in frontier capabilities,” the paper states. A lead that large by 2028 would be enormously advantageous. It would buy time. It would let democracies set norms before a frantic arms race removes any incentive for caution.
The analysis lands at a tense moment. U.S.-China talks in Beijing wrapped up this week with little apparent focus on tightening technology curbs. Some Chinese firms recently gained licenses for advanced Nvidia chips. Meanwhile Anthropic itself has clashed with the Pentagon over safeguards on its models’ use, a dispute that led to a temporary ban on federal use of its technology earlier this year. Yet on the threat from Beijing, the company speaks with one voice.
Compute sits at the heart of the argument. American and allied innovation in chips, combined with export controls maintained across three administrations, gives democracies a commanding lead today. Estimates suggest the gap could widen to 11 times more effective compute if enforcement improves. Chinese labs remain far behind in producing equivalent hardware. Huawei’s best efforts reached only about 4 percent of Nvidia’s compute level in 2026 projections cited in the report.
But the lead isn’t permanent. Smuggled chips flow into China. Companies tap overseas data centers stocked with restricted hardware. Gaps in controls on semiconductor manufacturing equipment let Beijing inch toward self-sufficiency. And then there are the distillation attacks.
These operations hit close to home for Anthropic. In February the company publicly accused Chinese labs including DeepSeek, Moonshot AI and MiniMax of running massive campaigns. They created tens of thousands of fake accounts to query Claude models millions of times. The goal? Harvest outputs to train cheaper copies that capture much of the original capability. OpenAI has documented similar efforts. So has Google DeepMind.
Anthropic calls distillation a back door. It lets China bypass some compute limits. A Chinese cybersecurity analyst quoted in the paper captured the frustration: “China is still sharpening our swords while the other side has suddenly mounted a fully automatic Gatling gun.”
The stakes rise with every capability jump. Frontier AI doesn’t just answer questions. It accelerates discovery in semiconductors, biotech and materials science. It finds software vulnerabilities at machine speed and chains them into attacks. It coordinates swarms of drones or cyber operations with minimal human oversight.
The People’s Liberation Army already procures commercial Chinese AI for military experiments. Think autonomous targeting. Cyber offense. Battlefield coordination. When a breakthrough arrives, deployment can happen in weeks, not years. A model that discovers novel exploits first in Chinese hands could penetrate U.S. infrastructure before defenses adapt.
Inside China the technology serves domestic control too. Automated surveillance at scale. Censorship refined by predictive models. The paper notes the Chinese Communist Party’s track record. It already uses AI to repress dissidents, hack foreign targets and bolster the military.
So what does winning require? Anthropic offers concrete steps. Tighten export controls on the highest-end chips and manufacturing tools. Increase enforcement budgets at the Commerce Department. Disrupt smuggling networks and access to foreign clouds. Make distillation explicitly illegal where possible and improve detection through information sharing among labs and government.
Beyond defense, the company urges offense. Export American AI systems and infrastructure aggressively to allies and partners. Speed adoption at home so democratic values guide deployment. Build resilience against authoritarian influence operations that could exploit AI-generated content.
The Anthropic paper frames this as more than commercial rivalry. It’s about who writes the rules for technology that will sit inside critical systems for decades. Democracies, the argument goes, have no inherent security advantage once AI is embedded everywhere. The side that develops and deploys first holds the edge.
Recent coverage echoes the urgency. Interesting Engineering reported Anthropic’s warning that China could surpass the U.S. by 2028 absent stronger chip restrictions and protections against theft. The piece highlighted overseas compute access and distillation as key accelerators for Beijing.
The Register cast the document as a lengthy missive urging Washington to kneecap China’s ambitions before transformative AI arrives. It noted the success of past controls yet pointed to persistent workarounds and Beijing’s talent and state investment.
Analysts in China pushed back. South China Morning Post quoted experts calling the plea irresponsible and self-serving. They argue it ignores China’s genuine innovation while serving Anthropic’s interest in stricter barriers that protect its own models.
The debate carries echoes of earlier fights. Anthropic has restricted sales to entities linked to the Chinese military. It rejected a request from a Chinese think tank for access to its latest Claude Mythos Preview model, as detailed in The New York Times. That refusal came during back-channel talks in Singapore last month.
Washington has taken notice of the theft problem. The State Department issued cables urging allies to pay attention to Chinese distillation campaigns, following Anthropic’s complaints. Yet the broader policy mix remains inconsistent. The Pentagon’s earlier designation of Anthropic as a supply-chain risk over safeguard disagreements sent mixed signals about domestic support for leading AI firms.
Critics like Steven Sinofsky, former Microsoft executive, questioned the paper on X. He argued that centralized control to preserve a lead repeats failed Cold War thinking. Innovation, he said, flows around restrictions. History shows technology eventually spreads. The answer lies in openness and markets, not tighter government grip.
Others see real risks. A Brookings Institution analysis from April examined competing U.S. and Chinese strategies. China leans on open-source models and rapid deployment across its industrial base. That approach generates data advantages that may matter as much as raw frontier performance. U.S. controls target training compute but do less against small, specialized models improved through real-world use.
Anthropic acknowledges the complexity. It supports safety dialogue with Chinese experts where feasible. Yet it insists the current compute imbalance remains the single biggest differentiator. Preserve it, widen it, and the U.S. gains leverage. Lose it, and the balance tilts.
The clock ticks toward 2028. That’s the year the paper flags as pivotal for transformative systems. A one-to-two-year lead then could compound. It might accelerate allied adoption, deter aggression and create space for measured international agreements. Without it, competition turns neck-and-neck. Safety takes a backseat. Authoritarian advantages in scale and control could prove decisive.
Policymakers face hard choices. Enforcement costs money. Tighter rules annoy allies and some domestic industries. Export promotion requires coordination across agencies. But the alternative, according to Anthropic, is ceding the future to systems built under different incentives.
The company has skin in the game. Its models already operate on classified networks. It turned down revenue by blocking certain Chinese-linked users. CEO Dario Amodei has spoken repeatedly about the need for strong controls. The paper builds on that record.
Whether Washington listens remains uncertain. Recent summits suggest chip policy sits lower on the agenda than trade or tariffs. Yet the intelligence community and defense officials track AI progress closely. The PLA’s interest in intelligentized warfare is no secret.
Distillation attacks continue. Loopholes persist. Chinese labs release competitive models. The gap hasn’t closed yet. But the paper warns it could faster than expected. Compute isn’t everything. Talent, data, determination and state direction matter. China possesses all three in abundance.
Democracies hold the chips, literally and figuratively, for now. The question is whether they will play them aggressively enough to maintain the edge. Anthropic’s message is clear. Act before 2028. Close the gaps. Promote your own technology. Shape the world that powerful AI will inhabit. The alternative is watching someone else write those rules.
WebProNews is an iEntry Publication