2028: Two scenarios for global AI leadership

What it argues for

This is Anthropic's statement of what US policy on AI competition with China should be, and its argument is that compute access decides the contest. It opens: "It's essential that the US and its allies stay ahead of authoritarian governments like the Chinese Communist Party, or CCP," because AI "will soon become powerful enough to be used to repress citizens at unprecedented scale." It contrasts two 2028 worlds. In one, America has defended its compute lead and holds a 12–24 month edge. In the other, policymakers "have not tightened loopholes on the CCP's access to compute" and Chinese labs are neck-and-neck. The document credits bipartisan export controls with today's lead, and says Chinese labs have kept close only through two workarounds. The first is illicit or evasive compute access, meaning smuggled chips and offshore data centres, "a route current controls do not reach because US export law covers the sale of chips, not remote access to them". The second is distillation attacks, which it calls "systematic industrial espionage of a technology critical to long-term US national security interests." Its three asks follow from that. The first is to close the loopholes on smuggled chips, foreign data-centre access and semiconductor manufacturing equipment, with more enforcement money. The second is to restrict model access and deter distillation, including "a legislative clarification that distillation attacks are illegal" and threat-intelligence sharing between labs and government. The third is to champion the export of American AI. It argues that a close race is itself a safety risk, because firms would "feel more pressure to release new models and products faster, without taking prudent pre-deployment safety measures". It also says that "Anthropic supports international AI safety dialogue with AI experts in China, when possible," while holding that such engagement goes best from a position of strength.

Stated positions (12)

• Democracies must lead: "Democracies, not authoritarian regimes, must lead in AI development and deployment," because the political systems that build the most advanced AI "will shape the rules and norms for how the technology is developed and deployed."
• The concern is the regime, not the people: "we do not seek to undermine the interests or ingenuity of the Chinese people"; the CCP is singled out as "the regime that is most able to use frontier AI to cement authoritarianism."
• Compute is the decisive input — "The race for global AI leadership is in large part a race for compute" — and export controls have worked: Huawei will produce "just 4% of NVIDIA's aggregate compute in 2026."
• If the US and its allies act now on export-control evasion and distillation, "it may be possible to lock in a 12-24 month lead in frontier capabilities"; 2026 is likely to be seen as "the breakaway opportunity for American AI."
• Close the loopholes on "Smuggled chips, foreign data center access, and SME": "Tightening controls and ramping up enforcement budgets can help close these loopholes that prop up the CCP's AI ecosystem," including deep ultraviolet lithography and the servicing and maintenance of chipmaking tools.
• Remote access is the named gap: Chinese labs train on export-controlled chips in Southeast Asian data centres, "a route current controls do not reach because US export law covers the sale of chips, not remote access to them"; a footnote records that a House bill to close it passed but "has not passed the Senate."
• Distillation attacks — fraudulent accounts used to harvest US model outputs — are "systematic industrial espionage"; policy should punish and disincentivise them, including "a legislative clarification that distillation attacks are illegal," and support threat-intelligence and technical sharing between American labs and with the US Government.
• Champion the export of American AI: the administration "should continue its efforts to promote the global adoption of trusted AI hardware and models developed and shaped by democratic principles," to deny the CCP global footholds.
• A close race undermines safety: with Chinese labs at or near parity, US and Chinese firms would release faster "without taking prudent pre-deployment safety measures," and governments "could become reluctant to enact policies to encourage responsible AI development and deployment, for fear of falling behind."
• Chinese safety practice lags: "only 3 out of 13 top Chinese AI labs published any safety evaluation results," and open-weight release of dual-use models means "safeguards that do exist can be removed."
• Supports safety engagement with China: "Anthropic supports international AI safety dialogue with AI experts in China, when possible," but "The prospects for productive engagement are best when the US maintains a large capabilities advantage."
• Resilience — sustaining political stability through the economic transition — is named as a fourth front of competition, but the essay explicitly sets it aside.

About this document

A long-form post on anthropic.com under the "Policy" label, dated May 14, 2026, with no byline, written in the corporate first person. It describes itself as "a new paper that explains our views on the competition on AI between the US and China", but it exists only as this web page; no PDF edition is linked. A short unheaded introduction sets out the two scenarios. A "Summary" follows, then headed sections: "The imperatives of staying ahead" (with subsections on authoritarian AI, dual use and the safety cost of a neck-and-neck race), "Our policy objective: creating and maintaining a lead for democracies", "Opportunities for engagement on AI safety", "The Mythos Preview wake-up call", "Four fronts of the competition", "The state of the competition" (how the compute lead was built and the loopholes that erode it), "Two scenarios for 2028" (a commanding lead versus a neck-and-neck CCP), "Ensuring democracies lead" (three bolded policy asks) and a "Conclusion". It links heavily to outside sources — CAISI's DeepSeek evaluation, CSET translations, Financial Times reporting, prosecutions for chip diversion, the White House AI Action Plan, OSTP's memo on distillation — and carries one footnote on the House bill targeting remote access to export-controlled chips. It is addressed to US and allied policymakers, and names the Trump administration and Congress as the actors whose decisions this year will settle the outcome.

How this sits against AI law

Each stance compared with what EU and US instruments actually require. Where no instrument addresses a theme, that gap is shown rather than hidden.

This is a US industrial-security agenda, and the EU barely features in it. EU law has no counterpart to what Anthropic wants: the dual-use Regulation licenses listed items on a common control list, but it has no China-specific AI compute controls, no provision for remote access to compute and no treatment of distillation. The EU also does not promote exports of a European AI stack as a strategic lever. In Washington the document is mostly pushing on an open door. America's AI Action Plan already calls for tougher chip export-control enforcement, controls on semiconductor manufacturing sub-systems, full-stack AI export packages and countering Chinese influence in standards bodies. Anthropic goes further on two points. It wants remote access to chips in offshore data centres brought within export controls, which existing controls on chip sales do not reach. And it wants legislation declaring distillation unlawful, which Executive Order 14409 does not provide, since its IP-protection language is a statement of policy backed only by existing computer-crime enforcement.

Source

Date on the page:

May 14, 2026

Source checked:

opened and confirmed on 2026-09-29