Does large-scale research infrastructure affect regional knowledge innovation, and how? A case study of the National Supercomputing Center in China

The paper investigates whether and how large-scale research infrastructures (LSRIs), specifically China’s National Supercomputing Centers (NSCs), affect regional knowledge innovation (KI). Motivated by the increasing importance of digital technologies and the high costs and uncertainties of LSRIs, the study addresses gaps in causal identification of LSRI impacts and their regional roles. It frames NSC as both computing (digital) infrastructure and a place-based innovation policy capable of reshaping local innovation resources and interregional knowledge flows. Research objectives include: assessing NSC’s causal effect on KI; unpacking mechanisms (basic resources, network embedding, technology/efficiency); evaluating policy spillovers beyond geography (cooperation and digitization proximities); and testing whether NSCs foster innovation convergence across regions.

The study situates LSRIs within prior work that identifies their scientific, technological, economic, educational, and social impacts (OECD; Michalowski, Qiao et al.). Scientific effects are decomposed into S&T advancement, capability cultivation, networking, and clustering effects. The authors integrate this with the resource-based view (RBV), social network theory, and innovation efficiency: tangible regional resources (financial, human, physical/digital capital) and intangible resources (social capital via network embedding; resource utilization/innovation efficiency). Networks facilitate knowledge diffusion and access to heterogeneous resources; efficiency governs the conversion of inputs to outputs. This mapping yields a conceptual framework linking LSRI scientific effects to regional innovation resources, hypothesizing three mechanisms—basic (R&D funding, talent, digital infrastructure), network (urban centrality in scientific collaboration), and technology (innovation efficiency). The study also draws on literature about digital infrastructure’s effects on innovation and on place-based policies and innovation convergence.

Data: Panel of 283 Chinese cities, 2000–2020. KI measured as per capita number of SCI-indexed papers from Web of Science; robustness uses highly cited papers (top 1% by field-year). Treatment variable: NSC dummy equals 1 from the approval year onward for NSC cities (Tianjin, Shenzhen, Changsha, Guangzhou, Jinan, Wuxi, Zhengzhou). Mechanism variables: (a) Basic effect—government S&T expenditure (R&D_exp, billion yuan), S&T employees (R&D_talent, 10,000 persons), digital infrastructure index (Digital_inf) via entropy method from telecom/internet indicators. (b) Network effect—urban centrality (degree and closeness) computed from a city-level co-publication network built from WoS co-authorships; centrality calculated with Ucinet. (c) Technology effect—innovation efficiency (Innova_effi) via stochastic frontier analysis (SFA) with translog function, inputs: S&T expenditure and S&T employees; output: KI; SFA preferred by LR test (LR chi2=77.76, p=0.000). Controls: per capita GDP (log), industrial structure (secondary industry share), comprehensive growth rate (n+g+δ), financial development (bank deposits+loans, log), human capital (university students per 10,000, log), transport volume (log). Data sources: Web of Science; China Urban Statistical Yearbook; Baidu Map API for inter-city distances. Missing values addressed via interpolation or city means. Empirical strategy: Treat NSC approvals (staggered since 2009) as quasi-natural experiment. Use two-way FE DID and spatial DID (SDM) due to significant spatial autocorrelation (Moran’s I significant at 1% annually). Spatial weight: inverse distance matrix. Baseline SDM: KI_it = vi + μt + ρ W KI_it + β1 NSC_it + β2 W NSC_it + γ Z_it + β3 W Z_it + ε_it. Mechanism tests follow Alesina and Zhuravskaya (2011): (1) regress mechanism on NSC; (2) include mechanism in KI regression and assess NSC attenuation. Spillovers assessed via direct/indirect effects decomposition (LeSage & Pace) using three proximity matrices: geographic inverse distance, cooperation proximity (co-publication frequency), digitization proximity (reciprocal of absolute differences in Digital_inf). Convergence: β-convergence model (Sala-i-Martin) on differenced KI with lagged KI, assessing changes after adding NSC. Robustness: event-study parallel trends; IV (FT per capita in 1984; relief degree of land surface, Rdls; frequency of digital economy policy terms, FDEPT; constructed as interactions per Nunn & Qian) with 2SLS; dynamic SDM with lagged dependent terms; generalized spatial 2SLS (using lags and spatial lags of regressors as instruments); placebo (pre-2009 fake policy dates); PSM-SDID (kernel matching by year on Econ, Industry_sec, n+g+δ, S&T expenditure share); alternative estimators (Tobit, negative binomial); alternative outcomes (per employee KI; highly cited papers); alternative samples (35 large/medium cities; provincial-level aggregation).

- Baseline effects: OLS with FE shows NSC significantly increases KI (coefficients ≈10.191 and 8.958). SDM confirms strong positive local effect (≈10.184 and 8.934, p<0.01) and significant positive spillovers to neighbors via W·NSC (≈47.319 and 37.966, p<0.01). - Basic effect (mechanism): NSC increases R&D_exp (6.216***), R&D_talent (3.228***), and Digital_inf (0.057***). When added to KI regressions, these mechanisms are positive and significant (R&D_exp 1.023***; R&D_talent 1.369***; Digital_inf 90.641***), and the NSC coefficient declines but remains significant, indicating partial mediation. - Network effect (mechanism): NSC raises Degree centrality (9.069***) and Closeness centrality (8.680***). Including centrality in KI regressions yields positive significant effects (Degree 0.736***; Closeness 0.528***), and NSC’s coefficient drops (from 8.934 to 2.226 and 4.329), evidencing mediation via enhanced network embedding. - Technology effect (mechanism): NSC improves innovation efficiency (Innova_effi 0.091***). Innova_effi significantly boosts KI (positive, reported as significant), and NSC’s coefficient declines (to ≈3.884) yet remains significant, indicating mediation through efficiency gains. - Policy spillovers across proximities (direct/indirect effects): • Geographic matrix: Direct 8.918***; Indirect 32.964***. • Cooperation proximity: Direct 8.626***; Indirect 1.917**. • Digitization proximity: Direct 4.413***; Indirect 14.015***. Spillovers are strongest via geographic and digitization proximities. - Convergence: Lagged KI coefficient is significantly negative (e.g., −0.632***). Including NSC changes absolute value only slightly (e.g., −0.634, −0.652), indicating NSC does not materially accelerate regional KI convergence. - Heterogeneity: Cities with higher shares of Mathematics publications experience weaker NSC effects; interaction NSC×Basic (Mathematics share) is negative and significant (−2.956***; −2.573***). Synthetic DID at city level shows significant positive ATT only for Shenzhen (8.938**, Basic=0.024) and Guangzhou (3.685**, Basic=0.033); overall treatment average 1.571***. - Robustness: Event-study supports parallel trends (effects emerge post-NSC). IV-2SLS and spatial G2SLS confirm positive NSC effects (e.g., second-stage NSC coefficients ≈69.125***; 37.768***; 20.310***). Dynamic SDM and alternative estimators (Tobit: 8.958***; NBR: 0.236***) support findings. Alternative outcomes (per employee KI: 25.857***; highly cited papers: 0.491***) and samples (medium cities: 3.020***; provinces: 3.338***) yield consistent positive effects. Placebo (pre-2009) shows no significant effects.

The findings directly address whether LSRI investments in computing infrastructure causally enhance regional knowledge creation: NSCs significantly increase local KI and generate measurable spillovers to surrounding and digitally similar cities. By mapping LSRI scientific effects to regional resources via RBV, the study shows NSCs augment tangible inputs (fiscal S&T support, talent, digital infrastructure), elevate cities’ social capital through higher network centrality, and improve innovation efficiency—each mediating part of the total effect on KI. This validates theorized capability cultivation, networking, and clustering effects at the regional scale and extends digital infrastructure literature by documenting computing power as a driver of scientific outputs. Nonetheless, the modest effect on convergence indicates that while NSCs stimulate local and proximate innovation, they are insufficient alone to reduce interregional disparities, likely due to limited center numbers and uneven digital readiness. Heterogeneity results imply stronger impacts in application-oriented innovation ecosystems (e.g., Shenzhen) than in cities with heavier emphasis on basic mathematical research, informing targeted, place-based policy design.

The study contributes a causal evaluation of LSRI impacts on regional knowledge innovation using China’s NSCs as a quasi-natural experiment. It integrates the resource-based view, social network theory, and innovation efficiency to build and test a mechanism framework. Main conclusions: (1) NSCs significantly increase local KI and induce spillovers via geographic, cooperation, and digitization proximities; (2) mechanisms include enhanced fiscal S&T investment, talent accumulation, digital infrastructure, greater network centrality, and higher innovation efficiency; (3) NSCs do not substantially promote regional KI convergence under current deployment; (4) effects are heterogeneous, being stronger in application-oriented cities. Policy implications: strengthen NSC support for both basic and applied research; expand and interconnect computing infrastructure to reduce “computing power islands”; leverage city-cluster strategies to amplify spillovers across proximities; for non-NSC cities, upgrade digital infrastructure and embed in collaboration networks; consider local knowledge orientation when siting and supporting NSCs. Future research could broaden coverage to provincial/micro-level computing centers, explore additional innovation outcomes and longer-term dynamics, and conduct cross-country comparisons to generalize LSRI computing infrastructure effects.

The number of NSC cities is limited, and the analysis does not include provincial or more microscopic-level supercomputing centers, which may underestimate radiation and spillover effects. This constraint likely contributes to the limited evidence on accelerating regional innovation convergence.