2028: Two Scenarios for Global AI Leadership — Anthropic

TL;DR

Anthropic's policy essay argues that the US-China frontier AI gap will be decided by 2028, primarily through compute access (export controls). Two scenarios: US maintains 12–24 month lead vs. China catches up via distillation attacks and loophole exploitation.

Authors & Org

Anthropic (policy/research essay, May 14 2026). No named individual authors — institutional position paper.

Method

Scenario analysis: two futures projected from current trends in compute access, export controls, model distillation capabilities, and democratic AI adoption.

Results

Scenario A (US leads): Export controls tightened, distillation attacks disrupted, democracies accelerate AI adoption → 12–24 month frontier lead sustained → US/democratic norms govern global AI.

Scenario B (US falls behind): Loopholes unaddressed, China uses distillation to close the frontier gap → China potentially overtakes → non-democratic norms propagate.

Anthropic's recommendation: defend compute advantage, close export loopholes, restrict model distillation, promote global deployment of American AI stack.

Significance

First major public Anthropic position paper directly engaging with US-China AI competition. Context: published alongside Gates Foundation partnership and PwC expansion — Anthropic positioning as a trusted policy/enterprise partner. Operationalizes the Anthropic Institute's "Threats & resilience" research pillar (see Anthropic).

Model distillation attacks is a key new framing: smaller models trained on frontier model outputs can close capability gaps. Restricting this is a proposed policy lever — significant if enacted, since much of open-source LLM progress relies on distillation.

Open Questions

• Is a 12–24 month frontier lead a stable equilibrium or a temporary gap?
• Can distillation restrictions be technically enforced at scale?
• Does "democratic AI stack" deployment actually constrain downstream norm-setting?

Cite

"2028: Two scenarios for global AI leadership" — Anthropic research, 2026-05-14. (link)