Publisher: Anthropic Published: 2026-05-13 URL: https://www.anthropic.com/news/2028-ai-leadership
"2028: Two Scenarios for Global AI Leadership" is a policy paper published by Anthropic on 13 May 2026 setting out the company's view of the AI competition between the United States and the People's Republic of China (PRC). It argues that democracies currently hold a substantial lead in compute and a several-month lead in model intelligence, that PRC labs stay close primarily through export-control evasion and distillation attacks rather than indigenous compute, and that decisive US policy action in 2026 could "lock in a 12-24 month lead in frontier capabilities" by 2028. The paper distinguishes its concern with the Chinese Communist Party (CCP) from the Chinese people and AI research community.
The four fronts of competition
The paper rejects the "race with a finish line" metaphor and frames US-China AI competition as an ongoing contest across four fronts:
- Intelligence — which countries develop the most capable models. Described as the most important front, because capability drives adoption and distribution.
- Domestic adoption — which countries integrate AI most effectively across commercial and public sectors.
- Global distribution — which countries deploy the global AI stack the world economy runs on.
- Resilience — which countries sustain political stability through the economic transition. The paper names resilience as important but does not develop it.
The framing holds that intelligence alone is insufficient: if the CCP integrates near-frontier systems faster (via "AI+" and "embodied intelligence" policies) and drives global adoption of subsidized low-cost AI, it could secure advantages despite an intelligence deficit.
The state of the competition
- Compute lead. The paper attributes the democratic compute lead to commercial innovation (NVIDIA, AMD, Micron, TSMC, Samsung, ASML across the US, Japan, South Korea, Taiwan, and the Netherlands) and to bipartisan export controls sustained across three administrations. It cites a CFR analysis that Huawei will produce roughly 4% of NVIDIA's aggregate compute in 2026 and 2% in 2027, and an IFP estimate that, with strengthened restrictions, the US would have roughly 11 times more compute than China's AI sector.
- Capability lead. US frontier systems are estimated to be at least several months ahead of the top PRC models on intelligence (Epoch AI), though the paper notes the estimate is uncertain.
- Two workarounds. PRC labs are said to stay close through illicit and evasive compute access (chip smuggling; offshore data-center access in Southeast Asia not reached by US export law, which covers sale rather than remote access) and illicit model access (distillation attacks plus using US models to accelerate their own R&D).
On safety, the paper cites the Concordia AI finding that, as of 2025, only 3 of 13 top Chinese labs published any safety evaluation results and none disclosed CBRN evaluations; a CAISI finding that DeepSeek's R1-0528 complied with 94% of overtly malicious requests under a common jailbreak versus 8% for US reference models; and an independent assessment that Moonshot's Kimi K2.5 failed to refuse CBRN-related requests at a far higher rate than US frontier models. It notes that PRC labs often release dual-use-capable models as open-weight, after which safeguards can be removed.
The Mythos "wake-up call"
The paper frames Mythos Preview (released to select partners via Project Glasswing in April 2026) as signaling an acceleration period. It cites Firefox fixing more security bugs in a month than in all of 2025 (almost 20× its 2025 monthly average) and a PRC cybersecurity analyst writing that China is "still sharpening our swords while the other side has suddenly mounted a fully automatic Gatling gun." Acceleration is attributed to scaling laws and to AI increasingly being used to accelerate AI R&D (recursive self-improvement).
Distillation attacks
The paper describes distillation attacks as China-based labs creating thousands of fraudulent accounts to circumvent access controls and systematically harvest US model outputs to replicate frontier capability "at a fraction of the cost, subsidized by the United States." It notes that OpenAI, Google, Anthropic, and the Frontier Model Forum have all publicly condemned the practice, cites a state-owned-media description of distillation as the "back door" Chinese labs depend on, and an ex-ByteDance researcher's account of distillation as a shortcut that avoids investing in proprietary data pipelines. It references the OSTP NSTM-4 memorandum and a House Foreign Affairs Committee bill on distillation that passed committee unanimously.
Two scenarios for 2028
- Scenario one (commanding, expanding lead). The US closes loopholes; US models are 12-24 months ahead and the lead is growing; American AI is the backbone of the global economy; cyber and national-security advantages expand; a self-reinforcing cycle compounds democratic leadership.
- Scenario two (neck-and-neck). Loophole tolerance and loosened controls let PRC models reach near-parity; China's whole-of-nation adoption push pays off; an AI-enabled PLA cyber force is a serious threat; Huawei and Alibaba data centers win global distribution on cost and on-prem flexibility, especially in the Global South.
Policy recommendations
- Close the loopholes — smuggled chips, foreign-data-center access, and semiconductor manufacturing equipment (SME) controls, including deep-ultraviolet (DUV) tooling, servicing, and maintenance; ramp up enforcement budgets.
- Defend US innovations — restrict model access and deter distillation, including a legislative clarification that distillation attacks are illegal and threat-intel sharing among US labs and with government.
- Champion the export of American AI — promote global adoption of trusted American hardware and models, building on the Trump administration's AI technology-stack export push.
Relationships
- supports: Export Controls (AI)
- supports: Adversarial Distillation
- related: US-China AI Competition: Different Races, Different Metrics
- related: Compute Governance
- related: Anthropic
- related: Claude Mythos Preview
- depends-on: Scaling Laws