Superpowers and AGI

Center for a New American Security | Jacob Stokes

The United States and China are accelerating their strategic competition to achieve artificial general intelligence (AGI), a development that could fundamentally reshape the global balance of power. While Silicon Valley treats AGI as a primary objective, Beijing is integrating these advanced capabilities into its broader national AI policy to enhance state control and economic productivity.

This technological race is driven by the premise that possessing human-level AI performance will not automatically guarantee geopolitical dominance. Power depends on successful translation. Instead, the state that successfully deploys AGI must effectively convert raw computational power into tangible economic growth, information control, and novel military capabilities. However, systemic risks like large-scale system misalignment or loss of control present severe escalation dangers. To mitigate these hazards, Washington is advised to balance its technology bets while establishing robust bilateral tracking mechanisms with Beijing to manage the transition during the upcoming phases of emergence.

Comment

The technological bottleneck in the Sino-American race for artificial general intelligence lies in the physical infrastructure required to train frontier models. United States export controls, enforced by the Bureau of Industry and Security, directly target this vulnerability by restricting China's access to advanced silicon like Nvidia's H100 graphics processing units. This hardware asymmetry limits Beijing's capacity to train dense, multi-trillion parameter models at scale. Consequently, Chinese developers are forced to focus on algorithmic efficiency rather than brute-force compute scaling.

This forced pivot to efficiency manifests in the optimisation of open-source architectures like Meta's Llama series, which Chinese research institutes adapt to bypass hardware constraints. By utilising techniques such as low-rank adaptation and quantization, groups like the Beijing Academy of Artificial Intelligence attempt to close the capability gap without relying on massive clusters of restricted H100 chips. Ultimately, these software workarounds face diminishing returns, leaving the Beijing Academy of Artificial Intelligence structurally disadvantaged against American labs utilising unrestricted clusters of Nvidia's Blackwell architectures.

Strategic Question for Discussion

How would the strategic efficacy of the Bureau of Industry and Security's restrictions on Nvidia's Blackwell architectures change if the Beijing Academy of Artificial Intelligence successfully offsets hardware limitations through algorithmic optimisation?

The pattern suggests that while algorithmic breakthroughs can temporarily narrow the performance gap, they cannot entirely compensate for the massive compute deficits in training foundational models. My assessment is that physical hardware restrictions will remain the primary lever of American technological containment, as next-generation architectures like Nvidia's Blackwell still dictate the absolute ceiling of frontier AI capabilities. Consequently, software-level workarounds by Chinese labs will likely yield diminishing returns over successive generational shifts in hardware.

Share your assessment in the comments below.

To read in details click here