The Wrong Race: The US, China, and AI Competition

Quincy Institute for Responsible Statecraft | Alvin W. Graylin

The United States and China are currently locked in a high-stakes competition over artificial intelligence that underwrites nearly one trillion dollars in annual capital expenditures and sweeping export control regimes. American capability leads over Chinese models have shrunk from 12–14 months to just 2–3 months despite expanding export restrictions on high-bandwidth memory and lithography equipment.

Chinese open-weight architectures like DeepSeek V4 Flash now achieve performance parity with American frontier models at one-tenth the application programming interface cost. Intelligence is cheap. This technological convergence demonstrates that high-end hardware denial fails to secure a permanent military or strategic advantage between nuclear-armed rivals. Furthermore, small specialized models deployed by non-state actors present a far more immediate existential threat than state-sponsored artificial general intelligence. Achieving long-term security requires Washington and Beijing to abandon winner-take-all frameworks during upcoming bilateral summits and establish shared international safety protocols alongside domestic labor protections for automated workforce displacement.

Comment

The compression of frontier model lead times reveals how export restrictions on ASML lithography scanners and High Bandwidth Memory fail to bottleneck algorithmic optimisation. When hardware access is constrained, software architectures adapt by prioritising efficiency over brute compute scaling, altering the economics of technological parity. DeepSeek V4 Flash illustrates how open-weight distillation circumvents hardware blockades, allowing resource-constrained entities to achieve near-frontier performance.

This efficiency shift relies heavily on Mixture-of-Experts architecture and lower-precision quantisation to reduce memory bandwidth requirements during inference. By reducing GPU interconnect dependencies, these software modifications allow clusters using legacy SMIC-fabricated silicon to achieve throughput comparable to unconstrained Nvidia H100 arrays. Consequently, mathematical optimisation across open-weight frameworks routinely bypasses the physical hardware limits enforced by the Bureau of Industry and Security.

Strategic Question for Discussion

If algorithmic techniques like Mixture-of-Experts continue allowing legacy SMIC-fabricated chips to match Nvidia H100 output, which factor will determine state-level AI dominance — physical semiconductor access or open-weight deployment speed?

The available evidence points toward deployment speed and systemic economic absorption overriding raw hardware advantages as the primary driver of strategic utility. While Bureau of Industry and Security controls constrain physical compute density, open-weight architectures allow rapid integration into broader commercial and defense workflows. Consequently, state capability will increasingly depend on organisational adoption capacity rather than isolated frontier model benchmarks.

Share your assessment in the comments below.

To read in details click here