Anthropic's '2028' Essay Draws Battle Lines on US-China AI and Open Source
Anthropic has published its most concrete public policy statement yet, framing 2028 as the year AI crosses into recursive self-improvement — the point at which, it argues, the leading compute power becomes functionally uncatchable. Matthew Berman's 6,400-word section-by-section walkthrough of the essay, corroborated by Anthropic's own announcement, surfaces a policy document that has split the AI industry along a fundamental strategic fault line.
What the Source Actually Says
Anthropic's central thesis rests on three pillars: close export-control loopholes enabling CCP chip acquisition, defend US innovations against distillation attacks, and champion American AI exports — all to lock in a decisive compute lead before 2028. The essay presents binary outcomes: a US-led world where democracies set AI norms, or a CCP-led world where automated repression scales without the historical constraint of human enforcers. Anthropic cites its Mythos preview (Project Glasswing) — a 10-trillion-parameter model with autonomous cyber-offense capability, released only to select partners in April — as concrete evidence of why the current acceleration period demands immediate policy action.
Berman endorses the threat identification near-completely: PLA deployment of DeepSeek models for swarm UAV coordination and cyber offense, Huawei producing just 4% of Nvidia's aggregate compute in 2026 (falling to 2% in 2027), and China's inability to manufacture EUV/DUV lithography at scale. Where he diverges sharply is on open source. Enterprise leaders reviewing Claude Opus at $30/M output tokens against Chinese open-weight equivalents at $3/M will choose the cheaper option for the 99% of use-cases current models already saturate — handing China the norms-setting power that compute restriction is designed to prevent. He cites Jensen Huang's counter-position directly: "The day DeepSeek runs on Huawei first is a horrible outcome for our nation."
The essay also contains a self-contradiction Berman highlights: it declares intelligence "the most important of four competitive fronts," then concedes that near-frontier AI paired with subsidized global distribution could give China structural advantages over democracies even with an intelligence deficit — undermining the primacy of the compute restriction it prescribes.
Strategic Take
This essay sharpens the doctrinal fork for AI builders: Nvidia's open-source export thesis versus Anthropic's compute-restriction doctrine. Policy pressure on open-weight model licensing and export controls is intensifying — enterprise roadmaps built on open-weight Chinese models should assess regulatory exposure now, before the 2028 policy window narrows further.
