China vs US: Who is winning the AI race, in four charts

Al Jazeera | Hanna Duggal

United States President Donald Trump and China’s President Xi Jinping are scheduled to meet in Washington, DC, to address critical bilateral issues, including a proposed artificial intelligence notification hotline to mitigate national security risks. This diplomatic push occurs as both superpowers aggressively accelerate their respective AI capabilities to secure global technological dominance.

Currently, the US maintains a significant lead in infrastructure, holding nearly three-quarters of global AI computing power and operating 5,400 data centres compared to China's 14 percent capacity share. Furthermore, US hyperscalers like Amazon and Microsoft are projected to spend $764bn on AI infrastructure in 2026, dwarfing the $102bn projected for Chinese counterparts like Alibaba and Tencent. The spending gap remains vast. However, Chinese open-weight models like DeepSeek V4 Pro are rapidly closing the performance gap, trailing US frontier systems by only eight months. China also dominates academic output, producing over 27 percent of global AI scholarly publications.

Comment

The rapid adoption of Chinese open-weight models like DeepSeek V4 Pro on platforms such as OpenRouter exposes a critical shift in the global software ecosystem. While Washington relies on capital-intensive hardware hoarding to maintain its lead, Beijing is bypassing this bottleneck through algorithmic efficiency and open-source distribution. This democratisation strategy allows global developers to integrate Chinese architecture directly into their workflows, creating path dependency. Consequently, OpenRouter's high volume of Chinese tokens indicates that global software development is actively decoupling from American proprietary APIs like OpenAI's GPT-4.

This architectural shift carries profound implications for future standards-setting and cyber security. As international developers standardise on open-weight frameworks originating from Chinese firms, Western regulatory frameworks like the EU AI Act lose their enforcement leverage. Furthermore, the widespread deployment of these models creates a distributed diagnostic network, allowing Chinese developers to refine their systems using global telemetry data. Ultimately, this telemetry feedback loop accelerates the optimisation of Chinese dual-use models, rendering unilateral US export controls on Nvidia H100 chips increasingly obsolete.

Strategic Question for Discussion

If global software developers continue to favour open-weight models like DeepSeek V4 Pro on platforms like OpenRouter, how does this shift the efficacy of US hardware-centric containment strategies?

The trajectory indicates that widespread adoption of open-weight architectures reduces the strategic leverage of US hardware embargoes by lowering the compute threshold required for state-of-the-art performance. My assessment is that this software-level optimisation allows Chinese developers to achieve comparable operational capabilities using older, unrestricted silicon. Consequently, containment strategies focused solely on physical supply chains will face diminishing returns as algorithmic efficiency outpaces hardware limitations.

Share your assessment in the comments below.

To read in details click here