⚠️ Early Draft — This research is a work in progress. Data and analysis are subject to revision.

Research Briefing · February 2026

China's AI Strategy:
The Race to 2035

China is executing a whole-of-nation AI mobilization through a three-pillar incentive architecture — state investment that funds supply, subsidies that lower costs, and regulatory mandates that guarantee demand — producing deployment speed and scale no other country can match.

But the quality gaps, the compute reality, and the commercialization paradox reveal both the power and the limits of this model.

230+ findings 3 regions compared Confidence-scored

Section 1

The Strategy

No other country has a 20-year AI plan. China has had one since 2017 — and it's ahead of schedule.

China's AI ambition isn't a recent pivot. It's a phased, whole-of-nation program that began with the 2017 AI Development Plan (AIDP) and extends to 2035, with a military-civilian fusion timeline reaching 2049. The strategy has evolved through three policy generations, each building on the last.

The 2017 AIDP set the foundation: make China the world's primary AI innovation center by 2030, with a core AI industry target of 150 billion yuan by 2025. By 2023, the actual figure hit 578.7 billion yuan — four times the target, two years early. By late 2025, it exceeded 1 trillion yuan.

The 14th Five-Year Plan (2021–2025) added safety alongside development. The 15th FYP (2026–2030), previewed in October 2025, shifts emphasis from breakthroughs to application, conversion, and ecosystem building. Its core themes: scientific self-reliance, "New Quality Productive Forces," and high-quality development.

The August 2025 Pivot: AI+ Action Plan

State Council Guo Fa [2025] No. 11 moved China from building AI vertically — labs, models, chips — to diffusing it horizontally across science, industry, consumption, governance, and global cooperation. The targets are concrete and mandatory:

70%
AI penetration by 2027
mandatory for SOEs & local govts
90%
AI penetration by 2030
"as widespread as electricity"
2035
"Full entry into intelligent
economy and society"

These aren't aspirational — they're compliance mandates backed by the incentive architecture detailed in §3.

The military dimension runs in parallel. The PLA targets mechanization-informatization-intelligentization integration by 2027, basic modernization with intelligent combat systems by 2035, and world-class military status by 2049.

The Whole-of-Nation Model (举国体制)

What makes this unique isn't any single target — it's the institutional machinery. The model aligns the State Council, three co-equal lead agencies (NDRC, MOST, MIIT), the PLA, universities, SOEs, and the private sector around shared objectives. In January 2026, all four agencies jointly issued the first national framework for government investment fund governance, with rules mandating consolidation of overlapping provincial funds and establishing national-level coordination that distinguishes between strategic chokepoints (national funds) and regional strengths (provincial funds).

Sources: State Council, NPC, US DoD Annual Report, NDRC official notices. Confidence: 0.90–1.0

Section 2

The Machine

China has built the largest AI ecosystem outside the United States — and in some dimensions, the largest in the world. But "largest" doesn't always mean "best."

The Companies

China's national champions span the full AI stack. DeepSeek proved that algorithmic efficiency can partially offset hardware constraints — its V3.2 model trained for just $5.57 million while scoring gold-medal level (96%) on AIME. Alibaba's Qwen pushes multilingual reach (119 languages) and open-source distribution. Moonshot's Kimi specializes in long-context processing (2M Chinese characters), raising $500M at a $4.3B valuation. Zhipu's GLM-Image is strategically significant as the first major open-source model trained entirely on Huawei Ascend chips — no Nvidia, no CUDA — demonstrating that the domestic hardware stack is viable for at least some workloads.

The open-source strategy is deliberate. Chinese models command 17.1% of global downloads (surpassing the US), up from 1.2% at end-2024, with peaks reaching 30%. This is a competitive weapon: openness builds global developer dependency on Chinese models, contrasting with the US API-only, cloud-locked approach.

⚠ Confidence note: The "95% LLM developer mindshare" figure sometimes cited is from an industry observer with 0.60 confidence — treat with skepticism. The 17.1% download share is better sourced.

The Talent Pipeline

53,400
S&E doctorates/year (China)
NSF NCSES 2022 actuals
44,700
S&E doctorates/year (US)
1.3M
Engineers graduated/year (China)
vs ~130K in the US

China produces roughly 53,400 S&E doctorates annually (2022 actuals, NSF NCSES), 37% in engineering — 1.5x the US figure. CSET Georgetown projects this reaching ~77,000 by 2025, though actuals aren't yet available.

The talent machine extends beyond organic production. The Qiming Plan recruits overseas PhDs (under 40, 36+ months at QS Top 200 universities) with up to 5 million yuan in salary subsidies, 300,000 yuan research funding, and 100,000+ yuan living allowances. "Bounty-as-a-Service" models offer up to $700,000 for specialized chip and AI-adjacent engineers. AI job openings grew 543% year-over-year in 2025.

Four Chinese universities now rank in the global top 10 for AI conference publications: Tsinghua (#2, surpassing MIT/Stanford), Peking, Zhejiang, and Shanghai Jiao Tong. China published 23,695 AI papers in 2024 — 42% of global output, matching the combined output of the US, UK, and EU-27.

Regional Implementation

Beijing

2,200+ AI enterprises (40% of China's total), nearly 50% of national large models. Zhongguancun: 1,900+ AI firms, 37 universities, 96 national research institutes, ZGC AI Park (350,000 m², 1B yuan startup funds), 70% discounted cloud from Alibaba and Huawei.

Guangdong

AI + robotics at industrial scale.

Zhejiang

Advanced chip design targeting 3–7nm nodes.

Guizhou

Leveraging cheap power, tripled compute capacity to 150 EFLOPS in 2025, with 90% AI-focused.

Chongqing

AI + automotive ecosystem (2.788M vehicles produced nationally).

Sources: State Council, Beijing/provincial government statistics, NeurIPS/ICML/CVPR 2024, NSF NCSES 2026. Confidence: 0.78–0.90

Section 3 — Centerpiece

The Incentive Machine

China's private AI market is the same size as Europe's. What sets it apart is a three-pillar incentive architecture that simultaneously funds supply, lowers costs, and guarantees demand. No Western country operates all three.

China's AI incentive system isn't just subsidies or just investment — it's an integrated machine with three interlocking pillars. Each pillar reinforces the others. This is the operational mechanism of the "whole-of-nation" model, and it has no Western equivalent.

1
Investment

Funding the Supply Side

State capital at a scale no other country matches: $8.3B national AI fund, $140B over 20 years, $47.5B IC Fund Phase III. Policy alignment weighted 60% in fund evaluations — financial returns explicitly secondary.

2
Subsidies

Lowering Costs at Every Layer

Tax breaks (15% vs 25%), 175% R&D super deduction, compute vouchers in 17+ provinces ($140K–$1.1M/company), up to 50% electricity subsidies, below-market land, K-12 AI education mandates.

3
Regulatory Protection

Guaranteeing Demand

70% AI penetration by 2027 (mandatory). Domestic chip mandates for state-funded data centers. Co-purchasing rules tying foreign chip imports to domestic procurement. Energy subsidies gated on domestic chips.

The Closed Loop

Investment
builds domestic capacity
Subsidies
make it affordable
Mandates
guarantee adoption
De-risks
further investment

Mandated adoption creates a guaranteed market → de-risks further investment → attracts state and private capital back into the ecosystem. The three pillars don't just coexist — they compound.

Pillar 1: Investment Comparison

Metric🇨🇳 China🇺🇸 US🇪🇺 EU
Private AI investment 2024$9.3B$109.1B$9.3B
Cumulative private AI 2013–2024$119B$470B$50B
AI VC 2025 (partial/est)$6B (574 deals)$174.6B (Nov)$17.5B (+75% YoY)
Source: Stanford HAI AI Index 2025, PitchBook, Crunchbase. Note: EU 2025 figure ($17.5B) is Crunchbase VC data, not directly comparable to the 2024 Stanford HAI figure ($9.3B).

But the private market is only half the story. China supplements with state capital:

  • $8.3B (¥60B) national AI investment fund (March 2025)
  • $140B (1T yuan) government-backed fund for AI and emerging tech over 20 years
  • $28–70B chipmaking incentives package under consideration (Bloomberg, Dec 2025 — unconfirmed)
  • National IC Fund Phase III ("Big Fund III"): $47.5B for chips (70% foundational, 30% advanced tech)

Pillar 2: Multi-Layered Subsidy Architecture

Tax Incentives

15% corporate tax for High/New Tech Enterprises (vs 25% standard). R&D super deduction: 175%. 220% amortization for IC/AI intangibles (2023–2027). 10-year loss carryforward.

Compute Vouchers (算力券)

17+ provinces, $140K–$1.1M per company. Shanghai: ¥600M program covering up to 80% of AI compute rental costs, plus ¥100M in data vouchers. Chengdu: ¥100M. Shandong: ¥30M vouchers + ¥1B infrastructure. Plus "model vouchers" for SME access to foundation models.

Energy & Infrastructure

Up to 50% electricity subsidies for major data centers (conditional on domestic AI chips). Below-market land pricing (70% minimum for encouraged industries). Up to RMB 10M per intelligent computing center. Beijing invested >$6B in computing hubs in western provinces.

Talent & Education

Qiming Plan and bounty programs. K-12 AI mandates: Tianjin compulsory weekly classes, Beijing 1,400+ schools reaching 1.83M students. 99% university AI adoption. 30M worker reskilling target by 2027. 20+ AI-linked occupations formally recognized.

Pillar 3: Regulatory Protection — The Pillar the West Doesn't Have

Adoption mandates: 70% AI penetration in intelligent terminals by 2027 — mandatory for SOEs and local governments. 90% by 2030. MIIT tracking: 3–5 large models, 1,000 intelligent agents, 100 datasets, 500 application scenarios by 2027.

Domestic chip mandates: Nov 2025 MIIT directive — all state-funded data centers must use only domestic AI chips (Huawei, Cambricon), retroactive to projects less than 30% complete. Dec 2025: Huawei and Cambricon added to government procurement list. Jan 2026: Draft co-purchasing rules tying each Nvidia H200 purchase to proportional domestic chip procurement. Jan 2026: China blocked H200 imports despite US export approval.

Energy-subsidy gating: The 50% electricity subsidies are conditional on using domestic AI chips — linking Pillar 2 directly to Pillar 3. You get cheap power, but only if you buy Chinese.

The Western Contrast

The EU offers programmatic funding: Horizon Europe + Digital Europe (~€1B/year), InvestAI Facility (aims to mobilize €200B), €600M for compute access, and plans to double annual AI investments to over €3B. Meaningful — but no compute vouchers, no energy subsidies for AI, no mandatory AI education, no state-directed talent recruitment, no compliance targets. The EU has Pillar 1 (partially) but lacks Pillars 2 and 3 entirely.

The US has even less on the incentive front — no direct equivalent to any of China's three pillars. US AI leadership rests on private capital, market dynamics, existing infrastructure advantages, and the sheer scale of its tech ecosystem. The US approach works through market pull, not state push.

Sources: Stanford HAI, PitchBook, China Briefing, Lawfare, RAND, Bloomberg, EC digital-strategy, Reuters, MIIT. Confidence: 0.75–0.95

Section 4

The Compute Gap

In raw GPU power, the US is 8x China and 17x the EU. But the numbers are muddier than they look.
Metric🇨🇳 China🇺🇸 US🇪🇺 EU
H100-equivalent GPUs~110,000~850,000~50,000
Global AI supercomputer perf share14.1%74.5%4.8%
Source: Sanchez.vc GeoCoded Report Aug 2025, Epoch AI, State of AI Report 2025. Methodology: Epoch AI tracks 501 AI clusters operational since 2019.
⚠ Beware the EFLOPS Confusion

Chinese state media reports "1,000+ EFLOPS of intelligent computing power" as of end-2025. This includes all computing — general purpose, industrial, scientific — not just AI-specific capacity. AI-specific compute is estimated at 105–230 EFLOPS (sources vary). These EFLOPS figures are not directly comparable across regions. The H100-equivalent GPU count and Epoch AI supercomputer performance share are the most consistent cross-region metrics.

The Chip War

China remains ~75% dependent on Nvidia export-compliant GPUs. The domestic alternative — Huawei's Ascend 910C — tells a story of real progress falling short of marketing claims:

BenchmarkAscend 910C Performance
Huawei marketing claim8–12% behind H100
DeepSeek researchers' assessment~60% of H100
vs Nvidia H200 (Counterpoint)76%
vs Nvidia GB200 for inferenceonly 20–30% per chip
System-level workaroundCloudMatrix 384 (384×910C) ≈ GB200 NVL72 — using 5.3× more chips

Production constraints remain severe: 200,000–800,000 units estimated for 2025, with SMIC 7nm yields below 30%. The Ascend 950, targeting H100-level competition, enters mass production in H1 2026.

DeepSeek's experience is instructive: they initially tried Huawei chips for their flagship model but found "results were unacceptable," switching to Nvidia H20 GPUs. Zhipu's GLM-Image, by contrast, trained successfully on the Ascend stack — suggesting domestic chips work for some workloads but not yet for frontier training.

The Policy Response

Domestic chip mandates for state-funded data centers (Nov 2025), government procurement lists (Dec 2025), co-purchasing rules tying foreign chip imports to domestic procurement (Jan 2026 draft), and an outright block on H200 imports despite US export approval (Jan 2026). These are demand guarantees for the domestic chip ecosystem — ensuring Huawei's 60%-of-H100 chips have a captive market while the performance gap closes.

The US response: Jan 2026 BIS rule shifts from "presumption of denial" to case-by-case licensing. HBM remains the key chokepoint. The overseas workaround (Alibaba and ByteDance training in Singapore/Malaysia) is targeted: "presumption of denial" maintained for China-owned overseas data centers.

The 2027 chip independence target — Beijing's stated goal of AI hardware independence — carries low confidence (0.58). Current realities (75% Nvidia dependence, 60% performance gap, sub-30% yields) suggest the target will likely be redefined rather than achieved.

Sources: Sanchez.vc, Epoch AI, Counterpoint Research, CFR, Reuters, WSJ, BIS Federal Register. Confidence: 0.74–0.98

Section 5

The Patent Paradox

China files 70% of the world's AI patents. The US gets 7x more citations per patent. Both facts matter.
Metric🇨🇳 China🇺🇸 US🇪🇺 EU (select)
Granted AI patents (10mo 2024)12,9458,609DE: 784, UK: 369, NL: 249, SE: 243
Citations per patent1.9013.18DE: 6.12
Global share of granted AI patents (2023)69.7%14.2%~2.8%
AI patent filings 2024300,51067,773
GenAI inventions (2014–2023 cumul.)38,2106,276n/a
Source: RDWorldonline (WIPO PATENTSCOPE), Stanford HAI AI Index 2025, EPO Patent Index 2024. Note: filings ≠ grants — 300,510 are filings; 12,945 are granted patents.
US citation advantage
per patent vs China
69.7%
China's share of global
granted AI patents
~2.8%
EU share of global
granted AI patents

China dominates AI patent volume at a staggering scale — nearly 70% of global granted patents, a 4.4:1 filing ratio over the US, and 6x more GenAI inventions cumulatively. But citation rates tell a different story. US patents receive 13.18 citations each on average versus China's 1.90 — a 7x quality gap that suggests many Chinese patents have limited downstream impact.

The EU is marginal in volume (~2.8% of global grants) but Germany's citation rate of 6.12 — more than 3x China's — suggests a quality-over-quantity model.

The Unresolved Question

Is China's citation gap a temporary lag (citations take years to accumulate, and China's patent surge is recent) or a structural difference in research impact? If the gap is closing, China's volume advantage becomes overwhelming. If it's structural, quantity without quality may not translate to technological leadership.

Sources: WIPO, Stanford HAI, EPO. Confidence: 0.78–0.85

Section 6

The Governance Divide

Three entirely different philosophies. Each claims to be the right approach.
Dimension🇨🇳 China🇺🇸 US🇪🇺 EU
Framework typeSector-specific, iterativeDeregulation / self-regulationHorizontal, risk-tiered
Key legislationPIPL (2021), GenAI rules (2023), human-interactive AI (2025 draft)Executive orders (rescinded under Trump)EU AI Act (Aug 2025), GDPR (2018)
EnforcementCAC direct authorityMarket self-regulationNational authorities under AI Office
Unique features2hr anti-addiction pauses, 1M user security thresholdNo comprehensive frameworkRisk tiers: prohibited/high/limited/minimal
Investment governance60% policy-weighted fund evaluationNoneNone

China regulates iteratively — sector by sector, adjusting quickly. The December 2025 draft rules on human-interactive AI target chatbots and emotional AI specifically, requiring conspicuous alerts when users interact with AI, dynamic reminders for new logins or excessive dependence, and mandatory 2-hour pause prompts. Security assessments trigger at 1 million registered users or 100,000 monthly actives. No other country regulates AI interaction patterns at this level.

The US under the current administration has moved toward deregulation and self-regulation. Previous executive orders were rescinded. There is no comprehensive AI framework.

The EU built a comprehensive horizontal framework with the AI Act (effective August 2025), classifying AI systems into risk tiers. The approach is predictable but slow — the AI Act took years to negotiate while China iterated through multiple sector-specific regulations.

Sources: CAC, BIS, EU AI Office. Confidence: 0.85–0.95

Section 7

The Deployment Paradox

China has 46% of the world's AI users — and captures 1.23% of the revenue. That tension defines the next decade.
46%
of global AI MAUs
(Chinese platforms)
1.23%
of top 100 AI company
revenue
~$1B
China LLM market
vs $4–6B US/Europe

Enterprise vs Population

Metric🇨🇳 China🇺🇸 US🇪🇺 EU
Agentic AI piloting/deploying46%40%30%
Population-level AI usage (H2 2025)~25%28.3%Norway 46.4%, France 44%
Major AI models launcheddozens403
Source: Microsoft AI Diffusion Report H2 2025 (population, confidence 0.95). Agentic AI: Computing.co.uk (confidence 0.70) — treat with caution.
The European Paradox

European citizens adopt AI faster than Americans — Norway 46.4%, France 44% vs US 28.3%. But European enterprises lag: only 30% piloting agentic AI vs 46% in China.

Industrial Deployment Strength

China's deployment strength is industrial, not digital. The country installs over 50% of the world's industrial robots (2024, state media — caveat: source reliability). Automated ports provide concrete evidence:

  • Zhoukou Port (Dec 2025): 12 automated cranes, 1.075M TEU, 30% productivity gains
  • Qingdao Port (Jan 2026): Vacuum-based automated mooring — berthing reduced from 20–30 minutes to under 30 seconds
  • Tianjin Port: Fully unmanned with Beidou navigation, 5G, and AI robots

The Commercialization Gap

Chinese platforms account for 46% of global AI monthly active users — but only 1.23% of top 100 AI company revenue. Revenue splits 89% enterprise, 11% consumer.

The data on actual enterprise uptake is contradictory: one survey shows 70% of Chinese enterprises piloting or deploying AI, another claims 90% haven't started. The discrepancy likely reflects definitional differences.

The state's answer is Pillar 3: regulatory mandates that guarantee demand. The 70% penetration target by 2027 is a compliance requirement. The question is whether mandated adoption creates sustainable usage or hollow metrics.

The Global Pilot-to-Production Problem

Worldwide, 88% of organizations use AI in at least one function, but only 7% have fully scaled across operations. The pilot failure rate: 95% of GenAI pilots fail to show financial returns within 6 months. This isn't uniquely Chinese — but China's response is unique: use state mandates and subsidies to force adoption past the gap.

The Open Source Monetization Question

If models are freely available, how do you build a commercial ecosystem? The 17.1% global download share creates developer dependency — but dependency without payment doesn't build an industry. This tension between open-source positioning and the 2035 economic targets (core AI industry reaching 1.73 trillion yuan / 30.6% global market share, per CCID forecast) remains unresolved.

Sources: Microsoft AI Diffusion Report H2 2025, McKinsey State of AI 2025, Deloitte State of AI 2026, People's Daily, China Daily. Confidence: 0.70–0.95 (varies by metric)

Section 8 — Case Study

OCR: From 75% to 94.5% in 24 Months

In 24 months, Chinese labs took OCR from 75% to 94.5% accuracy — and did it with a model 14x smaller than GPT-4o. This is what state incentives plus open-source culture look like in practice.

OCR offers the cleanest illustration of China's deployment thesis. In 2024, the best models hit roughly 75% accuracy on complex documents. Western VLMs (GPT-4o, Gemini 1.5) improved this to 77–88% through H1 2025. Then Chinese labs took over.

By late 2025, Chandra (83.1%) and HunyuanOCR (860 OCRBench score at under 3 billion parameters) signaled a shift. In January 2026, the field exploded: DeepSeek-OCR-2 hit 91.09%, and PaddleOCR-VL-1.5 (Baidu) reached 94.5% — an 11.5 percentage-point improvement in a single month, more than the entire previous year.

94.5%
PaddleOCR-VL-1.5 accuracy
(Baidu, Jan 2026)
0.9B
parameters — 14× smaller
than GPT-4o
Top 3
OCR models are
all Chinese

This isn't just about talent or funding — it reflects a pattern: the combination of state compute subsidies (lowering the cost of experimentation), open-source culture (rapid iteration and knowledge sharing), and a massive domestic market for document processing created conditions where Chinese labs could specialize and dominate a specific AI application faster than anyone else.

For printed documents, OCR is now solved infrastructure. For the whole-of-nation thesis, it's proof of concept.

Sources: OmniDocBench v1.5, olmOCR-Bench, ICDAR proceedings. Confidence: 0.85–0.90

Synthesis

What's Working, What's Not

Adapted from the strategic coherence assessment in the primary research.

Working

  • Academic excellence (4 universities in global top 10 for AI publications)
  • Algorithmic innovation (DeepSeek proving efficiency offsets hardware gaps)
  • Open-source global reach (17.1% download share, growing)
  • Talent pipeline (53,400+ S&E PhDs annually, aggressive overseas recruitment)
  • Coordinated multi-agency governance (Jan 2026 framework)
  • Regional specialization (provinces adapting AI+ to local strengths)
  • Multi-layered subsidy architecture (tax + compute + energy + talent + mandates)

Struggling

  • Monetization (46% of global AI MAUs, 1.23% of revenue)
  • Hardware per-chip gap (Ascend 910C at ~60% of H100, DeepSeek couldn't use it)
  • Patent impact (7× citation gap vs US)
  • Enterprise adoption clarity (conflicting 70% vs 10% surveys)
  • Production capacity (SMIC yields below 30%, 200K–800K chip output)
  • HBM chokepoint (US controls on high-bandwidth memory remain effective)
The Strategic Question

China has built a remarkably coherent policy architecture — clear milestones, aligned institutions, multi-layered incentives. The question is whether market and technology realities will bend to fit the plan, or whether the plan will quietly adjust its targets. Early evidence (DeepSeek, open-source success, academic dominance, OCR breakthrough) suggests China is finding asymmetric paths around constraints. But the commercialization gap and hardware realities remain fundamental challenges that may require redefining what "success" means.

Methodology

Sources & Confidence

High Confidence (0.85+)

  • Stanford HAI AI Index 2025 (investment, patents)
  • State Council Guo Fa [2025] No. 11 (AI+ targets)
  • NDRC official notices Jan 2026 (fund governance)
  • NSF NCSES 2026 (STEM PhDs)
  • Microsoft AI Diffusion Report H2 2025 (population AI usage)
  • US DoD Annual Report on China (military-civilian fusion)
  • Reuters (export controls)
  • EPO Patent Index 2024 (EU patent filings)
  • BIS Federal Register Jan 2026 (export rules)

Moderate Confidence (0.70–0.84)

  • PitchBook (VC data)
  • Sanchez.vc GeoCoded Report (compute capacity)
  • Epoch AI (GPU cluster tracking)
  • CSET Georgetown (talent projections — based on 2021 data)
  • Counterpoint Research (Huawei chip benchmarks)
  • Computing.co.uk (agentic AI adoption — 0.70, source unclear)

Low Confidence / Disputed

  • "95% LLM developer mindshare" — industry observer, unverified (0.60)
  • ">90% Chinese enterprises use GenAI" — DISPUTED, excluded from analysis
  • "China overtakes US+EU in adoption" (Sep 2025) — NOT VERIFIED by Microsoft data
  • "1,000+ EFLOPS intelligent computing" — misleading (includes non-AI computing)
  • Huawei "8–12% behind H100" — marketing claim contradicted by independent testing
  • 2027 chip independence target — LOW confidence (0.58), likely to be redefined

Known Data Gaps

  • No EU-wide STEM PhD aggregate
  • No clear China enterprise software (non-industrial) GenAI adoption rate
  • Chipmaking incentives ($28–70B) — Bloomberg report, unconfirmed
  • Regional ROI timelines by geography
  • China AI patent citation trajectory (improving or structurally different?)
  • Core AI industry discrepancy: official figures cite both 1T yuan and 1.2T yuan for 2025