China's AI collaboration network grows through triangle-closing and 'rich-get-richer' dynamics and is driven more by universities and research institutes than by firms; notably, organizations seek complementary technological expertise rather than similar tech, while geographic, cultural and institutional closeness still helps form ties.

Main Finding

Using a 12‑year longitudinal patent co‑authorship network (China, 2013–2024) and Stochastic Actor‑Oriented Models (SAOM), the paper finds that endogenous network dynamics (transitivity and preferential attachment) and actor heterogeneity jointly drive the evolution of China’s AI collaboration network. Universities and research institutes are more central drivers of tie formation than firms; organizations with higher innovativeness attract more partners; geographical, cultural and institutional proximities promote collaboration, but technological proximity has a significant negative effect—suggesting AI collaborations favor technological complementarity rather than similarity.

Key Points

• Data and scope
  • Collaboration measured by jointly authorized invention patents in China’s AI industry, 2013–2024.
  • Nodes: firms, universities, research institutions; ties: co‑inventor/co‑owner patent collaborations.
• Method
  • Stochastic Actor‑Oriented Model (SAOM) via RSIENA used to model tie formation dynamically, integrating endogenous structural effects, exogenous actor attributes, and dyadic proximity measures.
• Endogenous network effects
  • Transitivity (closure) positively influences tie formation — clusters form and facilitate knowledge transfer.
  • Preferential attachment/activity (popular actors attract more ties) is an active mechanism.
  • High overall density tends to inhibit new tie formation (coordination/redunancy costs); isolates are driven to seek ties.
• Actor heterogeneity
  • Universities and research institutes play a disproportionately central/brokering role compared with firms, consistent with open‑science norms and foundational research roles.
  • Higher organizational innovativeness (patenting activity/quality) increases attractiveness as a partner.
• Multidimensional proximities
  • Geographical, cultural, and institutional proximity positively affect collaboration likelihood.
  • Technological proximity has a notable negative effect — collaborations occur more across technologically distant/complementary actors rather than among highly similar technology holders.
• Contribution relative to prior literature
  • Integrates endogenous network dynamics, node attributes, and multiple dyadic proximities in a longitudinal actor‑oriented framework (overcoming cross‑sectional QAP/ERGM limitations).
  • Highlights AI‑specific dynamics (data intensity, rapid obsolescence) that amplify the role of technological complementarity.

Data & Methods

• Data sources and construction
  • Longitudinal panel of jointly authorized invention patents in the Chinese AI sector (2013–2024).
  • Collaboration network constructed at organizational level; organizational type coded (firm / university / research institute); innovativeness proxied by patenting metrics (paper mentions patent output as signal).
  • Dyadic proximity measures included geographical distance, cultural/institutional proximity (likely same region/ownership/state affiliation—paper provides operationalization), and technological proximity (patent IPC/tech class similarity or distance).
• Modeling approach
  • Stochastic Actor‑Oriented Model (SAOM) implemented with RSIENA to model tie formation/dissolution as actor choices over time under a Markov process and utility function.
  • Model simultaneously estimates endogenous structural parameters (density, transitivity, activity/preferential attachment, isolates) and covariate effects (actor attributes, dyadic proximities).
• Strengths and limitations in methods (implicit)
  • Strength: captures dynamics and endogenous dependence that cross‑sectional ERGM/QAP miss.
  • Limitation: reliance on jointly authorized patents as collaboration proxy may undercount informal or non‑patent collaborations; empirical results are China‑specific and may not generalize without replication.

Implications for AI Economics

• For innovation policy and cluster strategy
  • Support and strengthen university–industry–research institute linkages: universities/research institutes act as hubs/brokers and can accelerate system integration and breakthrough research diffusion.
  • Encourage cross‑technological, complementary collaborations (rather than only like‑with‑like partnerships) — policies that lower friction for interdisciplinary/adjacent‑technology teaming can be especially productive in AI.
  • Geographic and institutional proximity still matter: regional policies, incubators, and local networks remain effective levers to promote collaboration.
  • Monitor network density: overly dense, redundant networks may hinder new tie formation; policy should balance cohesion with incentives for novel partnerships.
• For firm strategy
  • Firms should partner with academically rooted actors to access foundational knowledge and broader networks; signaling innovativeness (quality patenting) increases partner inflow.
  • Pursue complementary technological partnerships (partners with different but compatible expertise) to maximize innovation returns in AI.
• For empirical research in AI economics
  • Dynamic actor‑oriented network models (SAOM) are appropriate for studying evolving R&D/collaboration structures in fast‑moving sectors like AI.
  • Measuring technological distance and complementarity matters: treating similarity as uniformly beneficial can mislead analysis in data‑intensive, rapidly changing domains.
  • Future work should triangulate patent data with other collaboration indicators (joint projects, publications, data‑sharing agreements) and test external validity across countries/sectors.

Notes/ Caveats - The manuscript is an Article in Press (accepted 12 March 2026) and the version provided is unedited; operational details (exact measures for proximities, model specifications, robustness checks) are not fully reproduced here and should be checked in the final published version.