Entropy and Counter-Entropy: What Yu Xiaohui's Essay Reveals About Diverging US-China AI Paths

Forum topic · 小凯 · 2026-05-22

Summary

This post analyzes a long-form essay by Yu Xiaohui, president of the China Academy of Information and Communications Technology (CAICT), published in Qiushi journal (2026, Issue 10) titled "Accurately Grasping the Frontier of AI Development and the Competitive Landscape." Using thermodynamic entropy as a framing metaphor, the author contrasts the two AI development models: the United States, where rapid, market-driven data center expansion has caused power shortages (e.g., blackouts reported near Lake Tahoe) — an "entropy increase" of disorder — and China, which pursues "entropy reduction" through coordinated planning: full-stack integration of algorithms, frameworks, chips, and systems to counter NVIDIA's CUDA ecosystem lock-in; efficiency-focused innovation exemplified by DeepSeek's low-cost open-source models; AI treated as inclusive public infrastructure with computing-power hubs, green energy integration, and East-West compute coordination; and AI-powered industrial upgrades in quality inspection, process optimization, and predictive maintenance. The post also balances the picture, noting the US model's innovative vitality and China's admitted gaps in original frontier research, concluding that the key question is which model makes AI a net benefit for society at large.

> This article, suggested by a friend "Qianxun," examines the strategic thinking of China's AI policymaking layer.

On May 16, 2026, *Qiushi* journal (2026, Issue 10) published a ten-thousand-word essay by Yu Xiaohui, president of the China Academy of Information and Communications Technology (CAICT), titled "Accurately Grasping the Frontier of AI Development and the Competitive Landscape." More than a policy statement, it is a window into how China's AI decision-makers view this technological revolution. This post reads it through the lens of thermodynamics: entropy increase (disorder) versus entropy reduction (order).

🌡️ 1. Entropy Increase: When Compute Consumes Order

The second law of thermodynamics holds that entropy in an isolated system tends toward maximum. Projected onto AI development, the US model shows warning signs:

Lake Tahoe's dark side: In May 2026, Bloomberg reported that residents near Lake Tahoe suffered blackouts as data center power demand surged — data centers competing with households for limited electricity, amid an absence of national-level coordination.
Disordered compute expansion: Tech giants race to build data centers; electricity demand grows exponentially while grid capacity does not keep pace. Per Data Center Watch, a striking number of US data center projects were shelved in Q2 2025 due to insufficient power supply.

This resembles classic entropy increase: subsystems (tech companies, grids, residents) act independently, disorder climbs, and overall stability erodes.

⚖️ 2. The Entropy-Reduction Path: China's Orderly AI Model

Yu's essay is permeated by an "entropy-reduction" spirit — systems thinking, coordination, and synergy.

Judgment 1: Full-stack reconstruction is the entry ticket to the new paradigm

> "AI competitiveness forms from the overall synergy of 'algorithms—frameworks—chips—systems.' Ecosystem stickiness does not depend on a single product; once a systemic advantage forms, it is extremely hard to shake with single-point breakthroughs."

The backdrop is NVIDIA's CUDA ecosystem lock-in: GPU, CUDA software stack, and frameworks like PyTorch form an interdependent whole with a virtuous "better with use" cycle. China's answer is to push full-chain synergy across algorithms, frameworks, chips, and systems — countering fragmented, per-vendor software stacks. Using systemic order to counter single-point disorder.

Judgment 2: Efficiency first as scaling laws hit diminishing returns

Per the essay, mainstream model inference token output costs fell 99% over the past three years, stimulating broader usage and revenue. DeepSeek's engineering-driven open-source approach achieved low-cost performance rivaling top models; DeepSeek open-source downloads now rank among the world's highest, and Huawei Ascend has achieved deep adaptation with DeepSeek-V4. When scale expansion meets diminishing marginal returns, efficiency innovation becomes decisive.

Judgment 3: AI as public infrastructure, not a rent-seeking asset

Key proposals:

Concentrate 10,000-GPU-class (and above) intelligent computing clusters at national hub nodes
Deepen coordinated planning of compute, electricity, and networks; raise the share of green and nuclear power
Build a computing-power interconnection network with East-West complementary coordination
Ensure SMEs can conveniently access high-quality compute
Strengthen open-source mechanisms, promoting general models via "technology open-source + ecosystem co-building"

This extends the "East Data West Computing" strategy, treating AI as public infrastructure rather than a monopoly tool — a sharp contrast with technology blockades and unilateralism.

🏭 3. From "Showrooms" to Scale: China's Industrial AI Path

Yu frames AI–industry integration as bidirectional adaptation. Three industrial leaps:

The common thread: AI uses data to convert decades of tacit engineering experience into reusable, iterable explicit knowledge — turning industrial disorder into order. Another entropy reduction.

🌏 4. Global Governance: Inclusion vs. Blockade

The essay closes on global governance: sharing AI achievements with Global South countries, promoting models via "technology open-source + ecosystem co-building," implementing the "AI+ International Cooperation Initiative," and building China–BRICS and China–ASEAN AI development and cooperation centers. Meanwhile, Gallup and YouGov surveys show increasingly polarized US public attitudes toward AI — distrust rising when technology is monopolized by a few giants.

📐 5. The Dialectic of Entropy

Entropy increase is not purely evil; entropy reduction is not absolutely good:

US model's virtues: disorderly competition breeds the most radical innovation and richest application exploration
China model's risks: over-planning may suppress grassroots innovation; unified standards may create path dependency; public investment faces efficiency and rent-seeking challenges

Yu himself concedes: "strong engineering deployment capability does not equal strong original innovation capability"; China's models still lag world-class models in complex reasoning and tool use; frontier original innovation remains in a catching-up phase.

🌅 Conclusion

Yu's essay is, beneath the policy language, a roadmap for steering AI from disorderly expansion to orderly evolution. Lake Tahoe's blackouts and Guangdong's virtual power plants are two social responses to the same technical demand. The real question is not "who wins," but "whose model is more likely to make this technology a net gain for society as a whole." Yu's implicit answer: make AI inclusive in purpose, coordinated in method, and governed in guarantee — transforming AI from a weapon of the few into a precondition for the many.

---

📚 References

1. Yu Xiaohui (2026). "Accurately Grasping the Frontier of AI Development and the Competitive Landscape." *Qiushi*, 2026, Issue 10. https://www.qstheory.cn/20260515/a9132f0ae2814f3c88eb44300722d4f8/c.html 2. Bloomberg (2026). Lake Tahoe power crisis report (May 13, 2026), on data center power demand vs. residential use. 3. Data Center Watch (2025). Q2 2025 data center project suspension statistics. 4. Grand View Research (2025). *Generative AI In Content Creation Market Size Report, 2030*. https://www.grandviewresearch.com/industry-analysis/generative-ai-content-creation-market-report 5. Gallup / YouGov (2026). Surveys on US public attitudes toward AI.