AI Race to 2028: Geopolitics, Compute Limits, Open‑Source Debate

Anthropic marks 2028 as the year when self‑improving AI could give the first mover an unassailable lead. If the United States secures that lead, global AI norms may follow a democratic template; if China catches up, authoritarian norms could dominate. The speaker argues that self‑improving AI would let an authoritarian regime deploy automated repression at a scale never before possible.

The Compute Advantage

Compute dominates AI progress, outweighing data volume and talent. Current U.S. export controls restrict China’s access to frontier chips, slowing its frontier‑model development. China pushes an “indigenization” strategy but confronts steep hurdles in high‑bandwidth memory production and EUV/DUV lithography. Because algorithmic advances multiply with available compute, any compute shortfall caps both model size and algorithmic innovation.

The Four Fronts of Competition

1. Intelligence – building the most capable models, a priority for Anthropic.
2. Domestic Adoption – embedding AI across commercial and public sectors.
3. Global Distribution – supplying the AI stack that powers the world economy.
4. Resilience – preserving political stability during rapid economic transition.

The speaker disputes Anthropic’s claim that intelligence is the decisive front, insisting that cheap, efficient open‑source models could accelerate worldwide adoption of Chinese‑built AI and shift the balance of power.

Proposed Solutions & Disagreements

Anthropic calls for tighter export‑control loopholes and limits on model access to block distillation attacks. The speaker agrees on blocking distillation but rejects the anti‑open‑source stance. Open‑source models, the speaker contends, are essential for global diffusion of AI technology; restricting them could undermine U.S. strategic interests.

Mechanisms & Explanations

A distillation attack trains a smaller model on the outputs of a larger, more capable model, achieving high performance with far less compute. Self‑improving AI describes systems that conduct their own R&D, creating recursive performance gains that outpace human‑led cycles. The “finish line” argument holds that once a nation attains self‑improving AI, exponential gains make it impossible for rivals to close the gap, effectively ending the race.