China's AI Catch-up: A Strategic Pivot from Bottlenecks to New Tracks

The most valuable insight from this article is: China's AI industry is achieving "asymmetric superiority" over Silicon Valley in specific domains through vertical integration of "chips-models-applications." The underlying logic of this trend isn't the raw intensity of technological breakthroughs, but rather the synergistic effect of supply chain autonomy and market scale.

1. A "De-Americanized" Closed Loop in the Tech Ecosystem is Forming

The core contradiction in the US-China AI race has shifted from "who has better GPUs" to "who has more complete AI infrastructure." Over the past two years, the US attempted to cut off China's access to advanced computing power via chip export controls, but China's industry responded with a two-pronged approach:

• Hardware Substitution: Domestic chips like Huawei Ascend 910B and Cambricon Siyuan 590 have reached 80% of NVIDIA A100 performance in inference scenarios, while the gap in training scenarios has narrowed to under 50% (based on public test data). More importantly, China's chip shipments grew by 210% year-over-year in 2025, driving economies of scale and cost reductions.
• Software Adaptation: Domestic frameworks like Baidu PaddlePaddle and Huawei MindSpore now cover over 70% of domestic large model training needs, while alternatives to PyTorch (such as OneFlow and Jittor) are rapidly penetrating the government and enterprise markets.

Dimension	US Advantage	China's Catch-up Status
Top-tier Training Chips	NVIDIA H100/B200	Ascend 910B (80% perf in inference)
Open Source Model Ecosystem	Llama, Mistral	Qwen, DeepSeek, GLM (2nd globally in academic paper citations)
Data Center Scale	Global top 10 cloud vendors hold 6 seats	Alibaba Cloud, Huawei Cloud, Tencent Cloud enter top 10; compute growth rate at 35% CAGR

Key Insight: When China can produce "good enough" chips and US chips cannot be freely supplied, Chinese AI companies no longer obsess over "most advanced compute," shifting instead to pursue "system-level efficiency." DeepSeek-V3 training a model that outperforms GPT-4o using just 2048 H800s is a typical product of this "compute optimization mindset."

2. Capital & Talent: Underlying Drivers from "Chasing" to "Local Leadership"

From a BCG perspective, the AI race is fundamentally about "capital density" and "talent density." Over the past five years, US VCs invested over $150 billion in AI, compared to roughly $80 billion in China, but the gap is narrowing. More critically, China has formed an advantage in application-layer talent reserves:

• In global AI top conferences (NeurIPS, ICML), the share of papers authored by Chinese researchers rose from 22% in 2019 to 38% in 2025, surpassing the US (35%).
• The average annual salary for Chinese AI engineers is only 1/3 of their US counterparts, yet project delivery efficiency (from model launch to iteration cycle) is 40% faster. This benefits from rapid iteration under "996" culture, as well as Chinese internet companies embedding AI into business operations earlier (e.g., ByteDance using AI to optimize ad placement, handling trillion-level requests daily).

[!quote] A VP of Engineering at Silicon Valley admitted privately: "What worries us is that while America is still debating 'AI safety,' China has already stuffed AI into every factory's quality inspection line and food delivery route."

3. Policy & Geopolitics: The Penetration Power of China's "Whole-of-Nation System" is Underestimated

The original intent of US chip bans against China was to "slow down China's AI progress," but the actual effect has been to accelerate China's determination for domestic substitution. In 2025, the central Chinese government directly allocated 120 billion RMB for basic AI research, plus local matching funds, bringing total investment to over 300 billion RMB. Meanwhile, of the $52 billion subsidies in the US CHIPS and Science Act, most were consumed by manufacturing segments like Intel and TSMC, with less than 10% flowing directly to AI R&D.

Even more noteworthy is data sovereignty: China has the world's largest internet user base (1.1 billion), and data flows are strictly regulated by the Data Security Law. This means Chinese AI companies can legally access massive, high-dimensional behavioral data, while US tech giants face data contraction due to privacy lawsuits. For example, TikTok's algorithm consistently leads Instagram in recommendation efficiency, which is essentially the multiplicative effect of data scale × algorithm iteration speed.

Actionable Advice

For readers of this tech forum, if you care about investment or career choices, I suggest doing two things:

1. Re-evaluate the valuation models for Chinese AI companies. Stop using linear extrapolations like "lagging the US by X years," and focus instead on leading indicators like "domestic substitution rate" and "vertical scenario penetration rate." For instance, in medical AI, China has approved 15 Class III medical device AI products, compared to only 7 in the US.

2. Pay attention to "compute efficiency" as a new variable. In the next phase, teams that can achieve high-performance models with lower compute resources (like DeepSeek, Zhipu AI) may offer better investment value than companies with the most GPUs.

![](https://bbs-physixfrontier-com-data.oss-cn-hongkong.aliy

Original Link: https://www.theguardian.com/technology/2026/jul/20/china-google-ai-race