Two Roads to 2028: Anthropic Maps the US-China AI Race

Anthropic published a policy paper this month arguing that the outcome of the US-China competition for AI leadership will probably be settled before 2028, and that the critical decisions are being made right now. The paper, titled "2028: Two Scenarios for Global AI Leadership," does not predict a winner. It describes two credible futures -- one in which allied democracies hold a commanding advantage, and one in which China closes the gap to competitive range -- and argues that the difference between them depends almost entirely on what policymakers do with advanced semiconductor export controls over the next 18 months.

Why 2028 Is the Frame

The report's central claim is that transformative AI systems could arrive as soon as the end of 2026. If that timing holds, the infrastructure and adoption patterns being built today will determine which countries and companies are best positioned to deploy those systems at scale. A lead of 12 to 24 months in frontier model development, Anthropic argues, could translate into durable advantages in economic productivity, scientific research, and national security that would be extremely difficult to close afterward. The word "transformative" is doing real work here. Anthropic is not talking about incremental capability improvements. It is describing a threshold beyond which AI systems can contribute meaningfully to research, software engineering, and economic planning in ways that compound over time.

The choice of 2028 as the frame reflects Anthropic's development timeline expectations rather than a fixed geopolitical deadline. But the language of the paper is notably urgent. Dario Amodei's recent remarks at the Council on Foreign Relations covered similar ground, describing a world in which AI writes most new code within a year. The 2028 paper extends that framing into explicit policy territory.

What China Has Already Done

The paper pays particular attention to a technique Anthropic calls "distillation attacks." These are not network intrusions. Chinese AI labs create accounts on US AI services, generate large volumes of outputs from frontier models, and use those outputs to train their own models. The result is that labs in China have built models far closer to US frontier capabilities than their own independent hardware and research would allow. Anthropic estimates China's current position as "competitive at the near-frontier," a phrase that would not have applied two years ago.

The report also addresses chip smuggling. Despite tightened export controls on advanced semiconductors introduced in 2022 and 2023, Chinese firms have found routes around them, including re-export through third countries and procurement through front companies. Anthropic describes this as an ongoing operational problem that current controls have not solved. The company's point is not that controls have failed entirely, but that the gap they have preserved is narrower than intended and is actively being closed.

"AI labs in China have built models close in intelligence to America's because of their talent, their knack for exploiting loopholes around export controls, and their large-scale distillation attacks that illicitly extract the innovations of American companies." Anthropic, "2028: Two Scenarios for Global AI Leadership," May 2026

Three Things Anthropic Wants Done

The paper makes three specific recommendations. The first is tightening export controls on advanced chips: closing loopholes that allow re-export through third countries, cutting off Chinese labs' access to cloud compute hosted outside China, and strengthening enforcement of existing restrictions. The second is actively disrupting distillation attacks. Anthropic calls for treating them as economic espionage rather than a terms-of-service problem, with coordination between AI companies and government agencies to detect and block them systematically. The third is accelerating global adoption of American AI systems, on the logic that a US platform embedded in the world's technical infrastructure is harder to displace than one that has won a benchmarks competition but has not scaled into production.

Each recommendation carries trade-offs the paper does not fully resolve. Tighter export controls risk pushing allied countries toward Chinese alternatives. Aggressive enforcement of distillation rules requires international legal cooperation that does not currently exist. And global AI adoption pushes US companies to ship in markets where regulatory frameworks are still being built. Anthropic acknowledges these tensions but argues that the risks of inaction are greater.

An Unusual Step for an AI Lab

Policy papers of this specificity are rare for AI companies. Most prefer to engage with government through lobbying and testimony rather than published scenario documents. Anthropic has been more willing than most to make its policy positions explicit in writing. Part of the explanation is structural: the company's ongoing dispute with the Department of Defense over supply chain designations has made its relationship with the US government an active, contested space rather than a quiet background condition.

The paper arrives at a moment when Anthropic's business case and its stated security concerns have converged. A world in which American AI companies maintain a meaningful advantage over Chinese counterparts is commercially favorable for Anthropic. The company is honest about this in the text, but argues the alignment of interests does not invalidate the analysis. Whether readers find that convincing will depend on how much weight they give to the underlying technical claims about distillation, compute, and the trajectory of Chinese AI development. What the paper establishes clearly is that Anthropic intends to be a participant in that policy debate, not just a subject of it.