Entropy Increase vs Decrease: Diverging US-China AI Paths Through Yu Xiaohui's Essay

Forum topic · 小凯 · 2026-05-22

Summary

This analysis uses the physics concept of entropy to examine a May 2026 10,000-character essay by Yu Xiaohui,院长 of the China Academy of Information and Communications Technology, published in Qiushi Magazine. The article characterizes US AI development as entropic: unchecked hyperscaler data-center expansion, grid strain (illustrated by Lake Tahoe blackouts), fragmented chip-software stacks, and unilateral export controls. In contrast, Yu's framework advocates an entropy-reducing Chinese model built on full-stack coordination across chips, frameworks, and algorithms, a national hub-and-spoke plan for 10,000+ GPU clusters, clean-energy integration, open-source ecosystems led by DeepSeek, and AI positioned as public infrastructure rather than rent-seeking property. The essay also addresses AI industrialization through visual quality inspection, process optimization, and predictive maintenance, while proposing open-source cooperation and BRICS/ASEAN partnership centers for global governance.

Key points

Source article: Yu Xiaohui (余晓晖), head of the China Academy of Information and Communications Technology, published "Accurately Grasping the Frontier and Competitive Landscape of AI Development" in *Qiushi* Magazine, Issue 10, 2026.
Analytical lens: The author frames US and Chinese AI strategies through thermodynamic entropy, treating disorderly expansion as entropy increase and coordinated planning as entropy decrease.
Entropy-increase pattern (US): Lake Tahoe blackouts illustrate how hyperscaler compute demand outpaces grid capacity, fragmented chip-software stacks replicate Nvidia CUDA lock-in, and unilateral export controls fragment global governance.
Entropy-decrease pattern (China): Full-stack coordination of algorithms, frameworks, chips, and systems; concentration of 10,000+ GPU clusters in hub nodes; coordinated compute-power-network planning with green and nuclear energy; positioning AI as inclusive public good rather than rent-seeking asset.
Efficiency over scale: Inference token costs fell roughly 99% over three years, validating DeepSeek's open-source engineering approach; Huawei Ascend adapters with DeepSeek models confirm Chinese capability in efficiency-led innovation.
Industrial AI: Visual inspection, process parameter optimization, and predictive maintenance convert tacit factory experience into reusable, iterable explicit knowledge, an entropy-reducing transfer from fragmentation to order.
Global governance: Yu endorses open-source ecosystems, South-South technology sharing, BRICS and ASEAN AI cooperation centers, contrasting with US unilateralism. He also concedes that China's original-innovation capability remains behind frontier labs, a candid self-assessment of over-planning risks.
Limits of the metaphor: The author warns that excessive planning may suppress local innovation, create path dependency, and invite rent-seeking; the US model still produces aggressive innovation despite social fragmentation.

Entropy Increase vs Decrease: Diverging US-China AI Paths Through Yu Xiaohui's Essay

Summary

Key points

Tags