China's AI strategy is diffusion, not frontier — and the bill is arriving
Every province produced an implementation plan within seven months. Data centre utilisation sits below 30%, compute rental prices halved in six months, and provinces are auditing subsidy fraud. Both facts describe the same policy.
The AI+ Initiative published by the State Council in August 2025 targets 70% penetration of AI applications in key sectors by 2027 and 90% by 2030. That framing is worth noticing: the objective is not building the best model. It is getting AI into industries that already exist.
Judged on diffusion, it is working. Judged on capital efficiency, it is producing exactly the waste that a subsidy-driven mobilisation produces.
The diffusion is real
Within seven months of launch, all 31 provincial-level regions had rolled out implementation plans, with over 120 sector-specific policies and more than ¥80 billion in dedicated subsidy funds. MIIT released a first batch of 100 benchmark AI+ application scenarios across 12 sectors to standardise adoption.
The results in industry are substantial:
- AI-powered quality inspection exceeds 75% penetration in automotive and 3C electronics, against a global average of 45%
- East Data West Computing has drawn over ¥1 trillion into 8 hubs and 10 clusters, reaching 1,590 EFLOPS
- SME adoption reached 21% in 2025, up 13 points from 2021 — outpacing the EU's 7-point rise over the same period, though behind the US at 34%
- China holds 38.6% of global AI patents and filed roughly six times as many generative AI patents as the US between 2014 and 2023
- It hosts 24.4% of global AI talent, second to the US at 32.6%
Set against the US, the pattern is consistent: China leads on application penetration in traditional industries and on policy-driven diffusion speed, and trails on enterprise-wide adoption (26% in manufacturing against 35–40%), top-tier talent, and fundamental research investment — $125 billion a year against over $300 billion.
Set against the EU, China leads on enterprise usage (25.9% against 19.95%) and invests roughly 2.8 times as much. The EU leads on regulatory framework and privacy protection, and on essentially nothing else in this comparison.
What the subsidies actually bought
This is where the source material is unusually candid, and it deserves to be read alongside the achievements rather than separately.
Idle infrastructure. Over 500 data centres with investment exceeding ¥1 trillion, at average utilisation below 30%. Smart computing centre utilisation sits below 50% nationally, and under 30% in parts of the central and western regions. Local governments built compute without matching demand.
Collapsing prices. Rental for A800/A100-equivalent compute fell from ¥3.8 to ¥1.9 per card-hour between August 2025 and February 2026 — over 50% in six months. General LLM API prices fell more than 70% in the same period, with some providers offering unlimited free calls. Industry average gross margin went from 45% in 2024 to below 20% by end-2025.
Overcapacity. AI server shipments exceeded demand by 30–40% in 2025. Domestic AI chip capacity utilisation runs at 20% or lower. Over 5,000 companies crowd the AI application market, with prices down 30–50% and margins compressed to 10–15%.
Duplication. Over 200 large language models above 10B parameters have been released, more than 70% in the general-purpose track, with a homogenisation rate above 60%.
Outright fraud. One Anhui enterprise produced 200 identical inspection robots purely to claim subsidies, with 82% of core components sitting idle. A firm claimed subsidies from both Shanghai and Jiangsu and had to return 30% for tax base duplication. A central province verified three cases of subsidy fraud in Q4 2025 involving over ¥20 million in falsified AI application scenarios. Guangdong and Zhejiang have launched audits, rectifying over 50 non-compliant projects.
And a behavioural problem that no audit fixes: 90% of enterprises treat subsidies as free technology upgrade funds, buying AI tools rather than re-engineering business processes. That is the difference between adoption statistics and productivity gains, and it is why penetration figures should be read carefully.
The security record
Fast adoption has produced documented failures at scale. The 2025 Harbin Asian Winter Games saw over 50 million cyberattacks, including the first significant AI agent attacks. A DeepSeek database leak exposed 1.3 billion records. A medical AI breach compromised three million patient records. Training data poisoning accounted for 87,800 data leakage cases, 43,000 of which directly exposed personal identities. One AI training data company was fined ¥210 million for bioinformation protection violations, and open databases produced over 159 data theft incidents.
Enforcement is running — 300+ data violation cases and 12 algorithm cases in 2025 — but the framework is still emerging, with high-level legislation pending.
How the three blocs fail differently
The comparison in the source is the most useful thing in it, because each bloc's risks are characteristic rather than incidental.
| China | United States | EU | |
|---|---|---|---|
| Dominant risk | Subsidy misallocation, overcapacity, price wars, data violations | Market concentration, funding winter, regulatory fragmentation | Compliance cost, slow adoption, SME exclusion |
| Subsidy duplication | Severe, driven by provincial competition | Minimal | Low, tightly targeted |
| Overcapacity | Severe across hardware and services | Minimal, supply concentrated | None — supply is tight, prices 30–50% above China |
| Regulatory risk | Moderate, legislation pending | High fragmentation, no federal law | Severe, strict AI Act enforcement |
| Monopoly risk | Moderate, antitrust active | Severe — 75% of compute, 80% of R&D in seven firms | Moderate, dominated by US hyperscalers |
| SME impact | Price pressure lowers access; subsidies favour large firms | High barriers, startup funding down 32% | Severe — Germany SME penetration only 14% |
China overbuilds and wastes capital. The US concentrates and excludes. Europe regulates itself out of adoption — German SME AI penetration at 14%, with compliance costs prohibitive for smaller firms and compute priced 30–50% above China.
Each is a real failure. They are not equivalent in consequence: overbuilt compute becomes cheap compute, and cheap compute eventually gets used.
Where the applications have actually landed
Manufacturing. A Shenzhen industrial park runs AI across 2,000+ devices, optimising production parameters 30 times an hour. A Zhuzhou pharmaceutical plant cut monthly scheduling from three days to 30 minutes with 33% higher output. Steel converter flame monitoring achieves 99.5% accuracy in slag detection.
Fintech. A ¥285 billion market in 2025, projected at ¥620 billion by 2030. Ant's Agentar-Fin-R1 financial reasoning model, Ping An's digital risk control, UnionPay's national AI pilot base with Huawei.
Healthcare. Tsinghua's DrugCLIP accelerates drug screening enormously. Insilico's ISM001-055 showed 98.4ml lung capacity improvement against −20.3ml for placebo in Phase IIa. Ant's AQ health app passed 100 million users. A traditional Chinese medicine AI system reports 92.3% prescription accuracy.
Transport. L3 autonomous driving has regulatory approval, with BYD, XPeng and Li Auto shipping L3-capable models. As of February 2026, 17 Chinese cities run full-scenario commercial L4 operation, with Apollo Go, Pony.ai and WeRide accumulating over 10 million robotaxi orders — the largest operational scale globally. 64-line LiDAR has fallen to ¥800, down 90% from 2021.
Smart city. City brain systems in Shanghai, Hangzhou and Shenzhen have cut peak congestion 20–30% through AI-optimised signals.
What this means if you are on the European side
The compute price collapse is an opportunity, not just a Chinese problem. Chinese compute at half its August 2025 price, against European compute priced 30–50% above Chinese levels to begin with, is a widening cost gap for anything trainable across borders. European firms with China operations should be looking at where that capacity sits idle.
Adoption statistics overstate transformation on both sides. If 90% of Chinese enterprises are buying tools rather than redesigning processes, headline penetration numbers are measuring procurement. The same caution applies to European adoption targets — and it means the real competitive gap is smaller than 25.9% versus 19.95% suggests.
Europe's regulatory lead is genuine and is costing it adoption. German SME AI penetration at 14% against prohibitive compliance costs is the AI Act's price, visible in the data. That is a legitimate policy choice, but it should be made knowingly rather than described as a competitive advantage.
L4 autonomy is commercially operating in 17 Chinese cities. For European automotive and mobility firms, the operational data being generated at that scale is the asset — and it is being accumulated somewhere else.
The industry has moved, in the source's own words, from a battle of a hundred models to an endurance race about real-world penetration and industrial integration. That is a more serious contest than the model leaderboard, and Europe is not currently entered in it.