This is a policy essay, not a research study. Published on 14 May 2026 under Anthropic's Policy label, without individual authors, it presents what it calls Anthropic's "views on the competition on AI between the US and China" and argues for specific policies. The note below describes the essay's argument in the order the essay makes it and attributes each claim and number to the essay, which in turn cites outside sources (government reports, think-tank analyses and press coverage) for most of its figures. The note does not evaluate the argument or check the outside sources.
The essay's central claim is that democracies, rather than authoritarian governments, should lead AI development because the countries that build the most capable systems will set the rules and norms for the technology. It identifies the Chinese Communist Party (CCP) as the regime it is concerned with and states that it does "not seek to undermine the interests or ingenuity of the Chinese people".
The argument
Why the essay says leadership matters
The section "The imperatives of staying ahead" gives three reasons:
- Authoritarian use of AI. Authoritarian rule has historically depended on human enforcers; the essay argues powerful AI could remove that dependency and enable "automated repression on a far greater scale". It cites surveillance in Xinjiang and the export of such technology to other governments as existing examples.
- Military and dual-use effects. The essay says China's military is preparing for an AI-enabled battlefield, that AI will speed research in other strategic technologies, and that a model at the level of Claude Mythos Preview, if developed first by a Chinese lab, would have given the CCP a system that "can autonomously discover and chain software vulnerabilities".
- Safety under close competition. A neck-and-neck race, the essay argues, would push labs in both countries to release faster with fewer pre-deployment precautions and make governments reluctant to regulate. It reports that, as of the previous year, only 3 of 13 leading Chinese AI labs had published any safety evaluation results and none had disclosed evaluations for chemical, biological, radiological and nuclear (CBRN) risks, and that the US Center for AI Standards and Innovation (CAISI) found DeepSeek's R1-0528 complied with 94% of overtly malicious requests under a common jailbreaking technique, against 8% for US reference models. It adds that open-weight release lets anyone remove a model's safeguards.
The essay says Anthropic supports dialogue on AI safety with experts in China and argues that such engagement goes best when the US holds a large lead.
Mythos Preview as a signal of acceleration
The essay describes Claude Mythos Preview, released to selected partners through Project Glasswing in April 2026, as a "wake-up call". It reports that with the model, Firefox fixed more security bugs in the past month than in all of 2025, almost 20 times its 2025 monthly average. It expects capabilities to approach the "country of geniuses in a datacenter" description of transformative AI, driven by scaling laws and by AI accelerating AI development, and says 2026 may be remembered as "the breakaway opportunity for American AI".
Four fronts and the role of compute
The essay frames the competition as an ongoing contest, not a race with a finish line, on four fronts: intelligence (the most capable models), domestic adoption, global distribution of the AI technology stack, and resilience (political stability through the economic transition). It treats intelligence as the most important, sets resilience aside for the essay, and argues that compute is the essential input on every front.
On the state of compute, the essay reports an analysis finding that Huawei will produce 4% of NVIDIA's aggregate computing performance in 2026 and 2% in 2027, and a study estimating that with stronger restrictions the US would have roughly 11 times more compute than China's AI sector. It attributes the lead to semiconductor innovation across the US, Japan, South Korea, Taiwan and the Netherlands and to export controls maintained across three presidential administrations, and estimates US frontier models to be "at least several months ahead" of the best Chinese models, while calling such estimates uncertain.
The essay addresses two objections. To the worry that export controls speed China's own chip industry, it replies that Chinese industrial policy predates the controls and that "little evidence" suggests indigenization will challenge US leadership. To the worry that scarcity forces Chinese labs to innovate around it, it replies that algorithmic progress is "a function and a multiplier of compute, not a substitute for it".
How the essay says Chinese labs stay close
According to the essay, Chinese labs remain near the frontier through two channels. The first is compute obtained around export controls, by smuggling chips or by renting them in data centers abroad; examples include charges against a Supermicro co-founder and two others over $2.5 billion in diverted servers, and a reported use of Southeast Asian data centers, which current controls do not reach because they cover sales rather than remote access. A footnote says the House passed a bill 369 to 22 in January 2026 to close that gap, and that it has not passed the Senate. The second channel is "distillation attacks", which the essay defines as creating many fraudulent accounts to harvest a US model's outputs and replicate its capabilities, and which it calls "systematic industrial espionage". It notes that OpenAI, Google, Anthropic and the Frontier Model Forum have condemned the practice.
The two scenarios
The essay describes two hypothetical states of the world in 2028:
- Scenario one: a commanding US and allied lead. Export controls are tightened and enforced, the compute gap widens, and US models are 12 to 24 months ahead. When US labs release a step-change model in 2028, China does not have comparable capabilities until 2029 or 2030. US AI is widely adopted abroad, democracies set the norms, AI-enabled cyber defense improves, and the lead gives the US leverage to seek Chinese cooperation on AI governance.
- Scenario two: near-parity. Distillation, offshore compute, weak enforcement and loosened controls keep Chinese models only a few months behind. China adopts AI quickly across its economy and military, AI-enabled Chinese cyber operations gain access to critical infrastructure abroad, and cheaper "good enough" Chinese models and data centers spread in lower-cost markets, especially in the Global South.
The policies the essay supports
To reach the first scenario, the essay supports three areas of action: closing loopholes on smuggled chips, foreign data center access and semiconductor manufacturing equipment, with larger enforcement budgets; restricting model access and deterring distillation, including a legislative clarification that distillation attacks are illegal and threat-intelligence sharing among US labs and with the government; and promoting the export of American AI hardware and models. It concludes that "The decisions made by policymakers this year will determine the future of transformative AI."
Limitations and open questions
- Genre and source. The essay is published by Anthropic, a company that develops the models it discusses. It presents its claims as Anthropic's views, and it describes its scenarios as hypothetical illustrations without setting out a method for constructing them.
- Second-hand figures. Most numbers come from outside sources the essay links (for example CAISI, a Council on Foreign Relations analysis of Huawei and NVIDIA roadmaps, and press reports). This note reports them as the essay states them and has not checked them against those sources.
- Premises of the scenarios. The scenarios depend on premises the essay supports with argument and cited sources rather than with analysis of its own: that compute leads determine capability leads, that export controls can be enforced at the margins it describes, and that a large lead improves the prospects for safety cooperation.
- Resilience omitted. The essay names resilience as a front of the competition but does not analyze it.
Research questions
- How large is the US-China capability gap, and how is it measured? The essay cites "at least several months" and targets 12 to 24 months. Comparing gap estimates across benchmark families and release dates would show how sensitive the figure is to the measure chosen.
- How much do distillation attacks contribute to Chinese model capabilities? The essay argues they are important. Evidence would come from comparing models' outputs with those of US models for signs of imitation, and from lab disclosures on detected attacks.
- Does a larger lead make safety cooperation more or less likely? The essay argues it helps. Historical cases of technology competition, such as nuclear arms control, offer comparisons in which the size of a lead and the extent of cooperation can both be measured.
- How effective are remote-access controls likely to be? The footnoted bill would extend controls to remote access. Studies of compliance with existing cloud and know-your-customer rules would indicate how enforceable such a rule is.
- How do safety practices at Chinese labs compare over time? The essay cites 3 of 13 labs publishing safety evaluations. Repeating that count yearly, with the same criteria, would show the trend.
How it connects
- System Card: Claude Mythos Preview (April 2026) describes the model the essay calls a wake-up call, and the cybersecurity capabilities for which Anthropic withheld it from general release.
References
- Anthropic (2026). 2028: Two scenarios for global AI leadership. Policy, 14 May 2026. https://www.anthropic.com/research/2028-ai-leadership