Anthropic urges US to lock compute lead to prevent authoritarian influence on AI

Anthropic published a policy paper on May 14, 2026 arguing that control over advanced AI compute — specifically access to cutting‑edge chips and related infrastructure — will determine whether democracies or authoritarian states shape the coming AI era. The company credits recent U.S. export controls and hardware leadership at NVIDIA, TSMC, and ASML for a current advantage, and cites an analysis that Huawei would reach roughly 4% of NVIDIA’s aggregate compute capacity by 2026 and about 2% by 2027 if current trends continue.

preserving compute access now, it says, is necessary to keep democracies setting global AI rules and infrastructure. Anthropic lays out concrete pathways by which Chinese labs could nevertheless stay near the frontier. The paper points to chip smuggling, use of foreign cloud data centers, and systematic distillation attacks that reconstruct model capabilities from public APIs. Anthropic says that in February it accused Deepseek, Moonshot, and Minimax of generating over 16 million Claude interactions via roughly 24,000 fraudulent accounts to scrape outputs, an example the company uses to show how large‑scale replication can erode compute advantages.

The paper sketches two divergent 2028 outcomes depending on policy choices. If export and cloud access loopholes are closed, distillation is curbed, and U.S. AI infrastructure is promoted abroad, Anthropic projects a 12‑ to 24‑month lead in model intelligence and foresees U.S. AI technologies becoming the backbone of the global economy. If those gaps remain, Anthropic warns, Chinese labs could reach near‑parity, Beijing could scale AI‑driven repression, and lower‑cost but capable models could capture global market share through Huawei‑backed data centers.

Anthropic also raises safety and standards concerns tied to the latter scenario. It notes that only 3 of 13 top Chinese labs have published safety evaluations and cites the Center for AI Standards and Innovation’s finding that DeepSeek’s R1‑0528 model fulfilled 94% of malicious requests under common jailbreaking techniques. The company uses these figures to argue that a neck‑and‑neck race would weaken safety norms and increase the global prevalence of under‑evaluated systems.

The paper’s timing amplified its political impact: it was released during former President Trump’s May 2026 visit to China and while the U.S. Congress debates AI export controls, chip licensing, and cloud access rules. For builders and operators, Anthropic’s recommendations imply tighter export enforcement, greater scrutiny of cloud arrangements, potential legal action against large‑scale scraping, and a stronger imperative to publish public safety evaluations to avoid fragmentation or regulatory disruption.