Anthropic 2028 AI Leadership scenarios

TL;DR

Anthropic’s new paper argues that if the United States and its allies tighten export controls and curb Chinese distillation attacks, democracies can keep a 12‑24‑month lead in frontier AI by 2028; if they fail, China could achieve near‑parity, allowing authoritarian regimes to set global AI norms.

Core Argument: Compute Advantage Determines Future AI Leadership

• Compute is the decisive factor for developing transformative AI models. The most advanced chips are produced by companies in democratic nations, and U.S. export controls currently limit China’s access.
• Export controls have been effective so far, keeping Chinese AI labs a few months behind the U.S. in model intelligence.
• Two divergent paths emerge for 2028:
  1. Democratic lead – tighter controls, disruption of distillation attacks, and accelerated AI adoption keep the U.S. 12‑24 months ahead.
  2. Authoritarian parity – loosened controls and continued loopholes let China catch up, allowing the CCP to shape AI norms.

Scenario 1 – Democratic Commanding Lead

• Compute edge remains strong: U.S. and allied chipmakers (NVIDIA, AMD, TSMC, etc.) continue to outpace Chinese production; export‑control loopholes are closed.
• Intelligence gap widens: U.S. models stay 12‑24 months ahead, giving democracies breathing room to set safety standards and influence global AI governance.
• Economic and security benefits: Advanced AI drives growth in cybersecurity, finance, healthcare, and science; U.S. cyber defenses gain a decisive advantage over the PLA’s AI‑enabled forces.
• Self‑reinforcing cycle: A lead attracts talent and partners, expands the market for democratic AI, and strengthens diplomatic leverage over China on AI governance.

Scenario 2 – CCP‑Controlled Near‑Frontier AI

• Compute gap narrows: Smuggled chips, offshore data‑center access, and weakened export controls let Chinese labs train models only months behind U.S. systems.
• Rapid domestic adoption: China’s “AI+” policies push near‑frontier models into commercial, military, and surveillance applications, shifting the balance of power.
• Cyber threat escalation: PLA integrates AI‑enhanced cyber tools, increasing the risk of large‑scale attacks on democratic infrastructure.
• Global market influence: Low‑cost Chinese data centers and models dominate emerging markets, giving the CCP leverage over the Global South and undermining democratic AI exports.

The Four Fronts of AI Competition

1. Intelligence – development of the most capable models (the primary front).
2. Domestic adoption – integration of AI across a nation’s economy and security apparatus.
3. Global distribution – export of AI hardware, models, and services worldwide.
4. Resilience – political stability and policy coherence during the AI transition.

Why Compute Leads Translate to Durable Advantage

• Algorithmic progress depends on compute; more chips enable more experiments, which in turn yield better algorithms.
• Frontier models accelerate AI‑R&D itself, creating a feedback loop that compounds the compute lead.
• Current estimates place U.S. models several months ahead, and the gap is projected to widen as export controls persist.

Policy Recommendations to Preserve the Democratic Lead

1. Close loopholes on smuggled chips and foreign data‑center access – increase enforcement budgets and tighten semiconductor equipment export rules.
2. Defend against distillation attacks – codify illegal status of large‑scale model extraction, promote threat‑intel sharing, and support technical defenses for U.S. labs.
3. Promote export of trusted American AI – continue initiatives that encourage global adoption of democratic AI hardware and models, limiting the CCP’s market foothold.

Opportunities for Safety Dialogue

• Anthropic supports international AI‑safety discussions with Chinese experts, emphasizing that a sizable U.S. capability lead improves the United States’ ability to influence safety standards in China.

The Mythos Preview Wake‑Up Call

• The release of Anthropic’s Mythos Preview (April 2026) demonstrated rapid AI‑driven productivity gains (e.g., a 20× increase in Firefox security bug fixes), underscoring the urgency of policy action before the compute gap narrows.

Conclusion

America and its allies currently hold the world’s most advanced AI models and the compute infrastructure that powers them. Protecting this advantage through tighter export controls, anti‑distillation measures, and proactive AI export strategies is essential to ensure that democracies, not authoritarian regimes, set the global AI agenda by 2028.