Recipes for Success: Knowledge Access–Appropriation Configurations for

d
Innovation Performance

we
Qingliang Menga, Xueyang Wua, Killian J. McCarthyb,c, Ruling Zhanga, , Jiajun Suna

a School of Economics and Management, Jiangsu University of Science and Technology, Zhenjiang, Jiangsu, China
b Nijmegen School of Managements, Radboud University, 6525 AJ Nijmegen, The Netherlands

vie
c Kyiv School of Economics, Kyiv, Ukraine

ABSTRACT: Firms face a well-known paradox of openness: they must access external knowledge to innovate
yet appropriate knowledge to capture value. Most prior research treats access and appropriation as independent

re
and static, overlooking how firms manage this tension over time. We address this gap using a multi-time-period
fuzzy-set qualitative comparative analysis (MTP-QCA) of 150 firms (2020–2024) operating in on the frontiers
of AI to examine which combinations of access and appropriation mechanisms drive innovation performance and
how these combinations evolve as conditions shift. We identify five equifinal configurations—Open Platform
er
Integrators, Collaborative Absorbers, Closed Specialisers, Leading Explorers, and IP-focused Market Makers—
each associated with high innovation performance, and three trajectories—Stable, Adaptive, and Transient—that
pe
capture how firms adjust strategies over time. Our findings show that there is no single best practice but multiple
viable paths, and that superior performance depends not only on choosing the right configuration but on shifting
strategies as technologies, competitors, and institutions change. This study reconceptualises the access–
appropriation tension as a dynamic, configurational challenge, extends paradox theory by integrating temporal
change, and advances theory on knowledge governance under shifting technological and institutional conditions
ot

Keywords: Knowledge access; knowledge appropriation; artificial intelligence; MTP-QCA

tn

Qingliang Meng can be contacted at mengql2007@[Link].

Xueyang Wu can be contacted at wuxueyoung@[Link].
Killian J. McCarthy (corresponding author) is an associate professor at Nijmegen School of Managements, Radboud
University. His research interests include technology innovation, strategy and Mergers and Acquisitions.
rin

Ruling Zhang can be contacted at [Link]@[Link].

Jiajun Sun can be conducted at sjjjust2@[Link].

Tel: 86-(0)511-8440-0186
E-mail address: [Link]@[Link]
ep

Full postal address：

Nijmegen School of Managements, Radboud University, 6525 AJ Nijmegen, The Netherlands
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Recipes for Success: Knowledge Access–Appropriation

ed
Configurations for Innovation Performance

v iew
Abstract: Firms face a well-known paradox of openness: they must access external knowledge to

re
innovate yet appropriate knowledge to capture value. Most prior research treats access and
appropriation as independent and static, overlooking how firms manage this tension over time. We
address this gap using a multi-time-period fuzzy-set qualitative comparative analysis (MTP-QCA) of
er
150 firms (2020–2024) operating in on the frontiers of AI to examine which combinations of access
and appropriation mechanisms drive innovation performance and how these combinations evolve as
conditions shift. We identify five equifinal configurations—Open Platform Integrators, Collaborative
pe
Absorbers, Closed Specialisers, Leading Explorers, and IP-focused Market Makers—each associated
with high innovation performance, and three trajectories—Stable, Adaptive, and Transient—that
capture how firms adjust strategies over time. Our findings show that there is no single best practice
but multiple viable paths, and that superior performance depends not only on choosing the right
configuration but on shifting strategies as technologies, competitors, and institutions change. This
ot

study reconceptualises the access–appropriation tension as a dynamic, configurational challenge,

extends paradox theory by integrating temporal change, and advances theory on knowledge
tn

governance under shifting technological and institutional conditions

Keywords: Knowledge access; knowledge appropriation; artificial intelligence; MTP-QCA

rin
ep
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 1. INTRODUCTION

ed
Artificial intelligence (AI) thrives on access to vast and diverse pools of data, expertise,

and technological inputs. Yet the very collaborations that make such access possible also

iew
expose AI firms to significant risks of knowledge leakage, imitation, and loss of competitive

advantage. This tension—between the need to open up in order to innovate, on the one hand,

v
and the need to protect in order to capture value, on the other—embodies what the literature

re
calls the paradox of openness (Arora et al., 2016; Foege et al., 2019; Laursen & Salter, 2014).

Research on how firms manage this tension has made important advances, showing that
er
some firms privilege access, others privilege appropriation, and still others strive for a
pe
balance between the two (Ayerbe et al., 2024; Ding & Wu, 2022; Laursen & Salter, 2004).

Yet two important gaps remain. First, most studies adopt a static, variable-centered view,

treating access and appropriation as independent levers rather than as interdependent

ot

elements embedded in configurations that co-determine outcomes. Second, few examine the
tn

dynamic nature and evolving trajectories of these configurations, even though technological,

competitive, and regulatory shifts can destabilize strategies that once worked.
rin

The knowledge-based view (KBV) Grant (1996) offers a compelling lens to address

these gaps. From a KBV perspective, firms compete by integrating, protecting, and
ep

leveraging knowledge resources, making both knowledge access and knowledge

appropriation central to strategy. In the AI context, four factors emerge as being particularly
Pr

salient: the breadth and depth of access—which determine the variety and intensity of

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 external knowledge flows (Laursen & Salter, 2006); the strength of the firm’s appropriation

ed
regimes—which shapes value-capture potential and bargaining power (Teece, 1986); the

firm’s industry-layer positioning (base, technology, or application)—which conditions

iew
spillover risks, collaboration requirements, and competitive dynamics; and the firm’s

absorptive capacity (Cohen & Levinthal, 1990) — which governs the ability to recognize,

external knowledge, and can in some cases substitute for formal appropriation regimes.

v
re
In this paper, we therefore ask: How do firms combine knowledge access and knowledge

appropriation conditions to maximize their innovation performance, and how these

er
combinations evolve over time? We examine this question using a unique sample of 150 AI

firms active from 2020 to 2024. We applying a multi-time-period fuzzy-set qualitative

pe
comparative analysis (MTP-QCA) to capture the complex, conjunctural causality (Furnari et

al., 2021), and to track how these configurations persist, adapt, or vanish over time.
ot

Our findings reveal three broad archetypal approaches: access-led strategies (leveraging
tn

high breadth and/or depth of access, reinforced by formal appropriation regimes or absorptive

capacity); capability-led strategies (where high absorptive capacity underpins selective access
rin

and can substitute for formal appropriation regimes); and appropriation-led strategies (where

formal appropriation regimes are central, especially in market-facing layers). Within these
ep

archetypes, we identify five distinct configurations, each associated with high AI innovation

performance. Dynamic analysis shows trajectories these configurations may evolve through.
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 In doing so, this study makes four contributions. First, we reconceptualise the paradox of

ed
openness, or the access–appropriation tension, as a dynamic, configurational challenge,

identifying multiple, temporally sensitive “recipes” for innovation success in AI. Second, we

iew
extend paradox theory by integrating dynamic and changes, showing how certain access–

appropriation balances persist while others adapt or vanish. Third, we enrich the KBV by

specifying how industry-layer positioning and absorptive capacity jointly shape the

v
re
sustainability of knowledge strategies in fast-evolving, knowledge-intensive sectors. Finally,

and methodologically, we introduce a replicable, context-sensitive classification system for

er
industry-layer positioning in AI, applicable across sectors. In that way, we advance the

academic research on the subject while providing actionable guidance for managers
pe
navigating the access–appropriation challenge in knowledge-intensive industries.

2. THEORY
ot

2.1 The Access-Appropiation Challange

tn

The paradox of openness, or the access–appropriation challenge, captures the inherent

tension between the benefits of drawing on diverse external knowledge and the need to
rin

capture value from it (Arora et al., 2016; Foege et al., 2019; Laursen & Salter, 2014).

Knowledge access enables firms to broaden their technological scope, accelerate innovation,
ep

and foster recombination (Chesbrough, 2003; Laursen & Salter, 2006). Yet extensive access

also increases exposure to imitation, spillovers, and opportunistic behaviour by partners

Pr

(Hannah & Eisenhardt, 2018; West, 2003). Research has documented divergent strategic

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 responses to this challenge: some firms privilege access (e.g., open-source communities,

ed
broad collaborations), others privilege appropriation (e.g., strong formal protection regimes,

restricted R&D), and still others aim for a contingent balance between the two (Ayerbe et al.,

iew
2024; Ding & Wu, 2022; Laursen & Salter, 2004). Much of this literature, however, treats

access and appropriation as independent levers, overlooking how they interact within broader

configurations of capabilities, structural positions, and environmental conditions.

v
re
2.2 A Configurational Perspective on Knowledge Strategies

The access–appropriation challenge is inherently multi-dimensional. Access and

er
appropriation mechanism rarely operate in isolation; they are embedded in broader
pe
constellations of organizational capabilities, technological positions, and environmental

conditions. A configurational perspective (Fiss, 2011; Misangyi et al., 2017) captures this

reality by examining how different elements combine to produce outcomes, embracing the
ot

logic of conjunctural causality—where the effect of any one factor depends on its interplay
tn

with others. From this perspective, there is no single “optimal” level of access or

appropriation. Rather, multiple equifinal pathways can lead to high innovation performance,
rin

each representing a distinct strategic “recipe” aligned with specific contextual and capability

conditions (Furnari et al., 2021). For example, broad and deep access may drive performance
ep

when coupled with high absorptive capacity and selective appropriation regimes, whereas the

same access profile could undermine performance if paired with weak internal capabilities or
Pr

high spillover risk. Conversely, strong appropriation regimes may be advantageous in

market-facing contexts but counterproductive in foundational technology layer where

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 collective standard-setting is essential. The configurational approach also accommodates

ed
causal asymmetry—the idea that the conditions leading to high performance may differ from

those leading to low performance—making it well-suited to a context where strategic trade-

iew
offs can yield divergent outcomes. This is particularly relevant for AI, where firms operate

across heterogeneous industry layers with distinct knowledge flows, competitive dynamics,

and appropriation risks. In such a setting, access and appropriation are best understood not as

v
re
binary choices but as relationally contingent elements whose value emerges only within

coherent, mutually reinforcing configurations.

2.3 Knowledge-Based View

er
pe
The KBV positions knowledge as the most strategically significant resource of the firm,

shaping its ability to create, appropriate, and sustain competitive advantage (Grant, 1996; Xie

et al., 2023). From this perspective, both knowledge access and appropriation are central to
ot

explaining innovation performance.

tn

In the AI context, four causal conditions are particularly critical for explaining how firms

balance access and appropriation. First, the breadth and depth of access determine the
rin

diversity and intensity of external knowledge inflows, influencing both innovation potential

and exposure to leakage. Second, the strength of the appropriation regimes—formal

ep

protection mechanisms embedded in both legal enforcement environments and firm-level

codification—shapes a firm’s ability to secure returns from innovation and to control the
Pr

terms of collaboration. Third, the firm’s industry-layer positioning (base, technology, or

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 application) conditions both opportunities for knowledge exchange and the risks of spillovers,

ed
serving as a structural anchor of the feasible balance between access and appropriation.

Finally, absorptive capacity (Cohen & Levinthal, 1990) governs the firm’s ability to

iew
recognise, assimilate, and exploit external knowledge, acting as a capability-based enabler

that can either complement or substitute for formal appropriation regimes. Together, these

conditions interact in complex, conjunctural ways, producing multiple equifinal pathways to

v
re
high innovation performance and underscoring that in AI, competitive advantage rests not on

a universal formula, but on contextually aligned knowledge strategies that evolve over time.
er
2.4 Theoretical Archetypes of Access–Appropriation Configurations
pe
The configurational perspective suggests that firms can achieve high innovation

performance through multiple, equally viable combinations of access and appropriation

mechanism, conditioned by their structural position and capabilities (Fiss, 2011; Misangyi et
ot

al., 2017). Drawing on the KBV (Grant, 1996) and paradox theory (Lewis & Smith, 2014),
tn

we theories three broad archetypes are particularly relevant.

First, the access-led strategies. These strategies emphasise high breadth and/or depth of
rin

external knowledge ties as the primary driver of innovation performance (Chesbrough, 2003;

Laursen & Salter, 2006). Formal appropriation regimes may be moderate, but absorptive
ep

capacity (Cohen & Levinthal, 1990) and relational governance mechanisms (Oxley, 1997)

allow firms with this configuration to benefit from openness while managing the threat the
Pr

comes from leaks. Such strategies will be most viable in base-layer positions, where

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 collective standard-setting and interoperability are critical (West, 2003), and the innovation

ed
premium comes from early integration of diverse knowledge sources.

Second, capability-led strategies. In these configurations strong absorptive capacity

iew
enables selective or even limited access to deliver high innovation returns (Zahra & George,

2002). Capability-led strategies allowing firms to internalise and recombine external

v
knowledge rapidly, reducing dependence on strong legal protection, can substitute for weaker

re
formal appropriation regimes under a specific situation (Rothaermel & Alexandre, 2009).

These strategies can arise in both base and technology layers, where the speed and quality of
er
internal knowledge transformation are decisive for capturing innovation benefits.
pe
Third, appropriation-led strategies. In these strategies, formal appropriation regimes—that

is, strong intellectual property (IP) protection, control over proprietary data, or platform

governance—are the dominant mechanism for capturing value (Pisano & Teece, 2007; Teece,
ot

1986). Access mechanism may be narrower in this configuration with partners carefully
tn

selected to minimise spillovers (Grimaldi et al., 2021). Such configurations are relevant in

both technology and application layers, where firms’ competitive advantage depends on their
rin

unique technologies, market proximity and complementary assets allow firms to monetise

tightly controlled innovations (Teece, 2018).

ep

While these archetypes are distinct, they are not mutually exclusive. Hybrid approaches

may emerge when firms combine elements from different archetypes in ways that fit their
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 context (Alexy et al., 2013). However, the structural role of the industry and the enabling role

ed
of absorptive capacity shape the feasibility and sustainability of each archetype over time.

2.5 Dynamic Evolution of Access–Appropriation Configurations

iew
Although the archetypes described above provide a static map of strategic possibilities,

the access–appropriation balance is rarely fixed. Paradox theory emphasises that competing

v
demands can be managed through persistent resolution, dynamic rebalancing, or temporal

re
cycling (Lewis & Smith, 2014; Schad et al., 2016). From a KBV, the relative value of access

and appropriation mechanism evolves with changes in capabilities, technological regimes,

er
and market contexts (Eisenhardt & Martin, 2000; Grant, 1996).
pe
In AI, these dynamics are amplified by rapid shifts in algorithms, data regimes, and

computing infrastructure, as well as evolving regulatory and competitive landscapes

(Lundvall & Rikap, 2022). For example, base-layer firms may need to maintain access in
ot

early stages to shape standards (West, 2003), but tighten appropriation regimes once
tn

dominance is established. Conversely, technology-layer and application-layer firms might

initially rely on formal appropriation regimes but expand access as they diversify into new
rin

technologies or markets (Adner & Kapoor, 2016). Similarly, absorptive capacity can

substitute for formal appropriation regimes in early, fluid ecosystems but may lose relative
ep

effectiveness as codification rises and competitive advantages become harder to sustain

without legal protection (Helfat & Peteraf, 2003).

Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 We theorise three trajectories through which access–appropriation configurations may

ed
evolve. First, stable strategies; that is, alignment between access mechanism, appropriation

mechanism, capabilities, and structural position remains robust despite environmental

iew
changes. Second, adaptive strategies; the overarching archetype persists, but firms adjust

enabling or constraining mechanisms—such as shifting from capability-led to stronger formal

appropriation—to respond to changing conditions (Tushman & O'Reilly III, 1996). Third,

v
re
transient strategies; which yield short-term performance advantages under specific conditions

but lose viability as technological, market, or regulatory contexts shift (Adner & Kapoor,
er
2016). By integrating these evolutionary trajectories with the archetypes above, we extend the

understanding of how paradoxical tensions in knowledge strategies are managed in fast-

pe
evolving, knowledge-intensive sectors, combining configurational and dynamic perspectives

to explain the persistence, adaptation, or vanishment of strategies.

ot

3. METHODS
tn

3.1. Setting

We examine how combinations of access (openness breadth and depth) and appropriation
rin

(IP strength) mechanism, embedded in distinct industry-layer positioning and moderated by

absorptive capacity, shape AI innovation performance and its dynamic evolution over time.
ep

We collect data from China because China is one of the world’s largest and fastest-growing

AI markets, it has a policy environment that strongly incentivizes AI development, it features

Pr

intense domestic competition, and it exhibits distinctive institutional conditions (Chen &

10

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Zhang, 2024; Ding & Wu, 2022; Zeng, 2022; Zhou et al., 2025). These characteristics

ed
generate substantial heterogeneity in how firms combine access and appropriation

mechanism, making China a theoretically rich setting for examining the strategic archetypes

iew
and temporal trajectories outlined in Sections 2.4 and 2.5.

3.2. Sample

v
We used two widely used industry lists to identify firms operating in the AI industry: the

re
AI Concept Stock List (firms with core AI-related products, technologies, or services) and the

AI50 Concept Stock List (50 representative, high-growth AI listed companies). Both lists
er
focus on firms operating in China’s AI industry and listed on Chinese stock exchanges, and
pe
are widely used in academic research (e.g., You and Meng (2018)) and policy analysis (e.g.,

Xinhua News, 2022; Yicai Global, 2025). We collected data from three consecutive editions

of the two lists: 2021–2022, 2022–2023, and 2023–2024. We then refined the list to only
ot

include firms that: (1) operated since 2020, to ensure that all firms were present for the full
tn

period; (2) ranked in the top 200 by market capitalization at the end of 2024, to focus on the

strategies of the most economically significant firms; (3) did not issue a “risk warning” (e.g.,
rin

“Delisting Risk Warning”) during the study period, to exclude financially distressed firms

whose strategies may be atypical; (4) filed at least one AI patent application during the study
ep

period, to ensure all firms have were active in AI innovation; and (5) had complete data for

all study variables, enabling robust cross-case comparison. Doing so let to the creation of a
Pr

list of 150, large, Chinese firms, active in the AI industry.

11

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 3.3. Measures

ed
We examine one outcome variable—AI innovation performance—and a number of causal

conditions grouped under the access–appropriation framework. The access dimension is

iew
operationalised through openness breadth (OB) and openness depth (OD), which capture the

diversity and intensity of external knowledge inflows, and absorptive capacity (AC), which

v
shapes the firm’s ability to exploit those inflows. The appropriation dimension is captured

re
through intellectual property (IP) strength, reflecting both internal and external mechanisms

for securing returns, and industry-layer position (BL, TL, AL), which conditions spillover
er
risks and value-capture opportunities. Table 1 provides an overview of all variables.

Following Meng et al. (2024), we apply a one-year time lag between the measurement of
pe
causal conditions and the outcome to reduce simultaneity bias.

-- Insert Table 1 About Here --

ot

3.3.1. AI innovation performance

tn

We follow prior work using patent counts to capture a firm’s AI innovation performance

(McCarthy & Aalbers, 2022; Wu et al., 2025; Zhou et al., 2025). Specifically, we measure the
rin

number of AI patent applications filed by each firm each year. We identified AI patents using

the Key Digital Technology Patent Classification System (2023) issued by the State
ep

Intellectual Property Office of China (SIPO), hereafter PCS 2023. PCS 2023 standardises AI-

related invention categories for statistical and policy purposes, spanning foundational
Pr

algorithms to sector-specific applications. Patent data were obtained from IncoPat. Each

12

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 patent’s International Patent Classification (IPC) codes were matched to PSC 2023

ed
categories, distinguishing AI from non-AI patents. Figure 1 illustrates how. Coding was

conducted by multiple researchers, with discrepancies resolved through discussion and cross-

iew
checking against PCS 2023 definitions. The outcome variable is the total number of AI

patents, across all PCS 2023 classes, for each firm-year. This outcome measure captures the

innovative output that access and appropriation strategies jointly aim to influence.

v
re
— Insert Figure 1 About Here —

3.3.2. Access: Openness Breadth and Depth

er
We follow Laursen and Salter (2006) and Kobarg et al. (2019) in operationalising
pe
openness in two dimensions. Openness breadth (OB) captures the diversity of external

knowledge sources, measured as the number of distinct partners (e.g., firms, universities,

research institutes, social institutions, individuals) with whom the focal firm has jointly
ot

applied for patents. Openness depth (OD) captures the intensity of these relationships,
tn

measured as the average number of joint patent applications with each partner (Devarakonda

& Reuer, 2018; Wu & Zhou, 2021). Patent co-ownership data were retrieved from IncoPat,
rin

which records applicant names. Partner identities were standardised (removing legal suffixes,

consolidating name variants, merging subsidiaries) and classified by type using firm
ep

websites, annual reports, ESG/sustainability reports, and independent industry directories

(e.g., iResearch AI Industry Report). Coding was done independently by multiple coders,
Pr

with discrepancies resolved via cross-checking. OB was calculated as the count of unique

partners per firm-year; OD as the mean number of joint patent applications per firm-year.

13

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 These measures directly reflect the firm’s access profile: breadth reflects the scope of

ed
external search, while depth reflects the relational strength of knowledge inflows.

3.3.3. Appropriation: Intellectual property (IP) strength

iew
Following Meng et al. (2024), IP strength combines two components: (1) the strength of

IP enforcement in the firm’s home province, and (2) the firm’s R&D intensity. This dual

v
measure reflects both internal and external drivers of appropriation (Ayerbe et al., 2024;

re
Teece, 1986). The internal component—R&D intensity—is the ratio of R&D expenditure to

main operating revenue, benchmarked against the industry average for the same year (CSRC
er
taxonomy used for sector classification). R&D expenditure and revenue data were drawn

from audited annual reports and cross-verified with CSMAR. Benchmarking ensures the
pe
measure reflects relative rather than absolute intensity. The external component—IP

enforcement—is measured as the proportion of patents granted and the number of IP

ot

enforcement cases (including infringement and counterfeiting disputes) in the firm’s home

province, relative to national totals in the same year. Data were taken from SIPO’s annual
tn

statistical reports. Together they construct a composite IP strength index. This composite

captures the firm’s capacity to appropriate value from innovation through internal R&D
rin

investment and the protection afforded by its institutional environment.

ep

3.3.4. Structural Anchor: Industry-Layer Positioning

We assign each firm to one of three mutually exclusive AI industry layers—base (BL),
Pr

technology (TL), or application (AL)—reflecting its primary role in the AI value chain. This

follows the Guidelines for the Construction of a Comprehensive Standardization System for

14

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 the National Artificial Intelligence Industry (MIIT, 2024), the Guidelines for Identification

ed
and Classification Grading of Artificial Intelligence Enterprises (Shanghai Government,

2024), and the Artificial Intelligence Standardization White Paper (China Electronics

iew
Standardization Institute, 2021). The classification covers foundational algorithms and

computing frameworks (base layer), enabling technologies (technology layer), and end-user

products/services embedding AI (application layer). Criteria, subdimensions, and examples

v
re
are summarised in Table 2. Firms’ primary layer was determined via structured content

analysis of annual reports, official websites, ESG reports, and independent rankings (e.g.,
er
iResearch AI Industry Report). Multiple coders applied the frameworks, with ambiguous

cases discussed and resolved. All firms were assigned to exactly one layer (MECE principle)
pe
(Harwood & Garry, 2003). Representative coding examples are in Table 3. Industry-layer

positioning shapes appropriation conditions: midstream and downstream actors tend to

ot

protect value through IP, while upstream actors often rely on openness to shape standards.

— Insert Tables 2 and 3 About Here —

tn

3.3.5. Enabler: Absorptive capacity (AC)

rin

We measure absorptive capacity (AC) as the natural logarithm of the number of R&D

personnel employed by the firm (Flatten et al., 2011; Meng et al., 2024; Zahra & George,
ep

2002). Data were collected from annual reports, ESG/sustainability disclosures, and

supplemented with CSMAR where needed. Where multiple sources existed, figures were
Pr

cross-checked for consistency; the most complete set was retained. Log transformation

normalises the distribution and reduces the influence of extreme values, important given the

15

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 size heterogeneity in China’s AI industry. This operationalisation captures the firm’s ability

ed
to leverage access—turning external knowledge into internal innovation through human

capital.

iew
3.4. Approach

We adopt a multiple-time-period fuzzy-set qualitative comparative analysis (MTP-QCA)

v
to examine how configurations of access-oriented and appropriation-oriented mechanism, in

re
combination with contextual factors, lead to high AI innovation performance.

QCA is well suited to this context because the paradox reflects complex causality, where
er
different combinations of conditions can lead to the same outcome (equifinality) and the
pe
presence or absence of a given condition can have asymmetric effects. Unlike regression,

which estimates net effects, QCA models the interdependencies between variables—treating

access and appropriation not as isolated predictors but as interacting strategic choices
ot

embedded in specific contexts (Du & Kim, 2021; Furnari et al., 2021).
tn

We use MTP-QCA, which applies QCA separately to multiple time periods, enabling us

to capture dynamics evolution in strategic configurations (Schneider & Wagemann, 2010).

rin

This is particularly relevant for AI firms, where technological, regulatory, and competitive

conditions evolve rapidly (Ming et al., 2025; Puntoni et al., 2021). By examining three
ep

consecutive periods (2021–2022, 2022–2023, 2023–2024), we identify stable configurations

that persist as well as adaptive or transient configurations.

Pr

16

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Finally, we integrate content analysis (CA) into our approach to classify firms into

ed
mutually exclusive AI industry-layers (base, technology, application). This captures how the

effectiveness of access–appropriation combinations are context-dependent (Arora et al.,

iew
2016; Ayerbe et al., 2024; Grimaldi et al., 2021) and ensures our classification is anchored in

policy-recognised standards (Harwood & Garry, 2003).

v
3.5 Analysis

re
We calibrated the outcome and causal conditions into fuzzy-set membership scores

following established fsQCA guidelines (Fiss, 2011; Rihoux & Ragin, 2009). For continuous
er
variables, the upper quartile, median, and lower quartile of the distribution served as anchors
pe
for full membership, the crossover point, and full non-membership, respectively (Du & Kim,

2021). Calibration was performed separately for each time period, as recommended for MTP-

QCA (Meng et al., 2024), to avoid bias from shifts in the underlying distributions.
ot

Categorical variables—the three industry layers—were coded as crisp sets, with 1

tn

assigned if a firm belonged to a given layer and 0 otherwise. The layers are MECE. All

coding was performed independently by multiple researchers, with inter-coder reliability

rin

assessed using Cohen’s kappa (κ > 0.80). Discrepancies were resolved by cross-checking

against primary sources (e.g., annual reports)

ep

The analysis proceeded in two stages. First, we conducted fsQCA for the first

observation period (T1: 2021–2022) to identify configurations of access, appropriation, and

Pr

contextual factors associated with high AI innovation performance. The minimum case

17

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 frequency was set to 2 (retaining ≥75% of observed cases), the consistency threshold for

ed
sufficiency at 0.80, and the proportional reduction in inconsistency (PRI) threshold at 0.70,

following Fiss (2011). All analyses were performed in fsQCA 4.1.

iew
Second, we extended single-period fsQCA to the second (T2: 2022-2023) and third (T3:

2023-2024) observation period, to reveal trajectories configurations may evolve through. This

v
dynamic design captures both persistent strategic trajectories and adaptive shifts in response

re
to environmental or technological changes (Schneider & Wagemann, 2010). We report the

intermediate solution as the main result for each period, supplemented with the parsimonious
er
solution for robustness. Sensitivity tests adjusting the consistency cutoff and the PRI

thresholds yielded substantively unchanged results. Given the absence of a priori assumptions
pe
about the directionality of all conditions, we applied the “presence or absence” counterfactual

approach (Fiss, 2011).

ot

4. RESULTS
tn

4.1 Descriptive Statistics and Calibration

Table 4 reports the descriptive statistics for all causal conditions and the outcome
rin

variable across the three observation periods (T1: 2021–2022; T2: 2022–2023; T3: 2023–

2024). The means and standard deviations indicate moderate-to-substantial variation across
ep

firms, reflecting heterogeneity in strategic orientation, resource profiles, and industry

positioning. The table also reports the calibration thresholds used in the fsQCA. Following
Pr

(Ragin, 2009), we set the fully in, crossover, and fully out thresholds for continuous

18

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 conditions based on theoretical expectations, empirical distributions, and prior work on

ed
access–appropriation strategy in innovation contexts (Arora et al., 2016; Laursen & Salter,

2014). Conditions capturing knowledge access — OB and OD — were calibrated using the

iew
upper quartile, median, and lower quartile as anchors. Conditions capturing knowledge

appropriation — IP strength — were calibrated in the same manner. AC was treated as a

complementary capability condition and calibrated using identical thresholds. The industry-

v
re
layer positioning (base, technology, application) derived from the CA was coded as crisp sets,

with a value of 1 assigned if a firm belonged to the given layer and 0 otherwise. These layers
er
are mutually exclusive and collectively exhaustive. Calibration was conducted separately for

each time period to account for dynamic shifts in the empirical distribution of variables, as
pe
recommended for MTP-QCA (Meng et al., 2024). These calibrated set memberships form the

basis for the subsequent analyses of necessary and sufficient conditions for high AI
ot

innovation performance.

-- Insert Table 4 About Here –

tn

4.2 Necessary Condition Analysis

rin

Before examining combinations of conditions, we first assess whether any single

condition is necessary for achieving high AI innovation performance. A condition is

ep

considered necessary if its consistency exceeds the conventional threshold of 0.90 (Schneider

& Wagemann, 2012). Table 5 reports the results for each causal condition — including
Pr

access-oriented (OB, OD), appropriation-oriented (IP strength), industry-layer positioning

19

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 (Base, Tech, App), and complementary capability (AC) factors — as well as their negations

ed
(~), across all three observation periods (T1–T3). None of the conditions, whether in their

presence or absence, reach the 0.90 consistency threshold. The highest observed consistencies

iew
are for OB (0.69 in T2) and OD (0.69 in T2), but these values remain well below the

necessity benchmark. Similarly, AC shows moderately high consistency (around 0.68 across

periods), yet still falls short. These results indicate that no single access or appropriation

v
re
factor alone can explain high AI innovation performance in our sample. Instead, outcomes

emerge from specific configurations of multiple conditions — reinforcing the configurational

er
premise of our study and the rationale for the sufficiency analysis that follows.

-- Insert Table 5 About Here –

pe
4.3 Configurations for High AI Innovation Performance

We conducted the sufficiency analysis in fsQCA 4.1 using the truth table algorithm,
ot

setting the frequency threshold to 2 and the consistency cutoff to 0.80. Intermediate solutions

were derived by incorporating theoretical expectations on how access mechanism (openness

tn

breadth and depth), appropriation mechanism (IP strength), industry-layer positioning (base,

technology, and application), and absorptive capacity jointly contribute to AI innovation

rin

performance. Table 6 presents the configurations linked to high AI innovation performance

for the first observation period (T1: 2021–2022). The results reveal five configurations (M1–
ep

M5), summarised below and in Table 7. Consistency scores across the five configurations

range from 0.828 to 1.000, with an overall solution consistency of 0.858 and coverage of
Pr

0.394, indicating strong empirical fit.

20

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 The first configuration, M1, we term the Open Platform Integrators. This configuration is

ed
typically found in the base layer. These firms open their access boundaries widely, both in

breadth and depth, drawing in diverse streams of external knowledge while simultaneously

iew
maintaining strong IP strength. This balance of generous inflows and firm appropriation

ensures that platform effects are amplified without eroding the firm’s ability to capture value

— a classic appropriability logic (Teece, 1986).

v
re
The second configuration, M2, we term the Collaborative Absorbers. Like the integrators,

these firms pursue broad and deep access, but instead of relying on strong IP strength, they
er
lean heavily on exceptional absorptive capacity. This allows them to internalise and

recombine partner knowledge rapidly, capturing value through superior learning and
pe
integration processes rather than formal exclusion mechanisms — a pathway consistent with

the KBV and absorptive capacity theory (Cohen & Levinthal, 1990).
ot

The third configuration, M3, we term the Closed Specializers. This configuration is

common in the technology layer, and the firms are assumed to rather focus on a closed
tn

innovation, combining narrow access breadth with shallow depth. By building, accumulating

and optimising unique internal resources, they develop their own sustainable competitive
rin

advantage, rely on strong IP strength and absorptive capacity — a classic market segment
ep

securing defender strategy (Enkel & Bader, 2012).

The fourth configuration, M4, we term the Leading Explorers. Also concentrated in the
Pr

technology layer, these firms cast a wide net, engaging with a large and diverse set of

partners as well as tight and frequent collaboration. This access breadth and depth, with

21

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 strong absorptive capacity and IP strength enables them to identify novel opportunities and

ed
experiment with emerging technologies quickly and exchange tacit knowledge, consistent

with exploration logics (March, 1991).

iew
The fifth and final configuration, M5, we term the IP-Focused Market Makers.

Predominantly in the application layer, these firms build their strategies on strong IP

v
protections, complemented by targeted, selective access. Their proximity to end markets

re
allows them to monetise downstream applications effectively, securing value capture in line

with value capture theory (Pisano & Teece, 2007).

er
Taken together, these configurations highlight that there is no single “best recipe” to

organise access and appropriation in AI innovation. Industry layer positioning shapes

pe
strategic options, complementary capabilities — whether IP strength or absorptive capacity

— underpin performance, and multiple configurations can lead to similar outcomes. The
ot

result is a rich set of strategic pathways, each suited to different resource profiles,

technological positions, and collaboration styles.

tn

-- Insert Table 6 & 7 About Here –

rin

4.4. Dynamic Evolution of Configurations

Extending the analysis to the second (T2: 2022–2023) and third (T3: 2023–2024)
ep

observation periods reveal that some configurations persist as enduring strategic archetypes,

while others are more transient, classifying them as three evolution trajectories (i) stable
Pr

strategies, (ii) adaptive strategies, or (iii) transient strategies.

22

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 The Leading Explorers (M4) is the most robust, appearing in all three periods with hardly

ed
any variations in the supporting conditions. Their combination of broad-and-deep access with

strong IP strength and high absorptive capacity seems resilient to changes in technological,

iew
regulatory, and market environments, making it a stable strategy.

The IP-focused Market Makers (M5) also shows signs of stability but with adaptation

v
over time — typically towards IP strength — gradually weaken. Agility in appropriation is a

re
viable strategy for market-facing AI business to distinguish themselves amidst uncertainty,

making it an adaptive strategy.

er
The Open Platform Integrators (M1) and the Collaborative Absorbers (M2) both reappear

in two periods and exist alternative conditions. This suggests that absorptive capacity or
pe
relational embeddedness can sustain performance for some firms, environmental shifts such

as increased competition or tightening IP regimes can push others to adjust their

ot

appropriation strategies, underscores the dynamic nature of strategies, making them adaptive

strategies.
tn

In contrast, the Closed Specializers (M3) is short-lived. It appears only in the first period
rin

and does not re-emerge in later years. This trajectory indicates that the configuration may be

contingent on temporary opportunity windows, favourable regulatory contexts, or

ep

experimental strategies that proved difficult to sustain as conditions evolved. Besides, two

additional short-lived configurations have been observed for the first time, both for base
Pr

layer, making them transient strategies.

23

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Taken together, the dynamic analysis suggests that the configurations yielding high AI

ed
innovation performance are not equally durable. Some can be institutionalised as repeatable,

persistent strategies, while others function as opportunistic plays suited to particular

iew
moments. The relationship between access and appropriation mechanisms exhibits dynamic

and evolution characteristics, for managers, this underscores the need to periodically reassess

whether their current access–appropriation balance remains aligned with the shifting contours

v
re
of the AI innovation landscape.

-- Insert Tables 8 About Here –

4.5. Robustness
er
pe
We adopted the method of adjusting the consistency cutoff (from 0.8 to 0.85) and the

PRI thresholds (from 0.70 to 0.80) for robustness testing. We judged it with the help of the

two criteria proposed by Schneider and Wagemann (2010) for the robustness of the QCA
ot

results. In conclusion, the findings of this study remained robust, see in Tabel 9. These
tn

checks would demonstrate that the identified configurations are not artefacts of specific

parameter choices.
rin

-- Insert Tables 9 About Here –

5. DISCUSSION
ep

5.1 Key Findings

Pr

Our results show that firms navigate the paradox of openness not through a single “best

recipe”, but via multiple, equally viable strategic configurations——Open Platform

24

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Integrators, Collaborative Absorbers, Closed Specialisers, Leading Explorers, and IP-focused

ed
Market Makers. Each combines strategies for knowledge access and knowledge appropriation

in a way that reflects the firm’s resource base, competitive posture, and positioning in the AI

iew
value chain. Industry-layer positioning emerges as a structural anchor. Whether a firm

operates in the base, technology, or application layer shapes both the opportunities and the

risks associated with openness. Within these structural constraints, complementary

v
re
capabilities — notably strong intellectual property (IP) protection or high absorptive capacity

— determine how effectively firms can benefit from external knowledge flows without

undermining value capture.

er
Dynamic analysis shows three trajectories these configurations may evolve through —
pe
Stable, Adaptive and Transient strategies. Some configurations, particularly Leading

Explorers anchored in the technology layer, prove a robust, long-time strategy. Others persist
ot

only by adapting, rebalancing access and appropriation as technological maturity, regulation,

or market conditions evolve — an indication that strategies in AI innovation are dynamic in

tn

nature, such resilience requires adjustment to enabling mechanisms while maintaining

structural fit. A third configuration (M3) appears only briefly, exploiting time-bound
rin

opportunities or responding to short-term pressures before vanishing as conditions shift,

reveals the limitations of transient strategies.

ep

Viewed together, these findings reframe the access — appropriation tension as a

Pr

dynamic, configurational challenge rather than a static trade-off. Sustained AI innovation

performance is not achieved by privileging access or appropriation in isolation, but by

25

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 orchestrating their alignment with the firm’s positioning in the value chain and adapting

ed
absorptive capacity over time. Strategic agility — the capacity to adjust enabling capabilities

while recognising when temporary approaches have reached their limit — emerges as a

iew
central driver of long-term success.

5.2 Theoretical contributions

v
This study advances research on artificial intelligence (AI), open innovation, and the

re
management of knowledge tensions in four ways.

First, we reconceptualize the paradox of openness (Arora et al., 2016; Kobarg et al.,
er
2019; Laursen & Salter, 2014) as a dynamic, configurational challenge rather than a static
pe
trade-off. While prior studies often treat access and appropriation as mutually exclusive (Cai

& Ma, 2022; Ding & Wu, 2022; Miozzo et al., 2016; Yacoub et al., 2020), our analysis shows

they can co-exist in distinct, high-performing “recipes” that vary systematically with a firm’s
ot

industry-layer positioning and complementary capabilities. Using multi-time-period fuzzy-set

tn

qualitative comparative analysis (MTP-QCA), we identify multiple, dynamic evolve

configurations to success, highlighting that no universal best practice exists.

rin

Second, we extend paradox theory (Lewis & Smith, 2014) by explicitly integrating

dynamic into the study of access–appropriation tensions. We show that some configurations,
ep

such as Leading Explorers remains stable, while others adapt their enabling mechanisms—or

vanish entirely—in response to shifts in technology, competition, or regulation. This

Pr

26

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 perspective advances understanding of how paradoxical tensions are managed not only

ed
through structural balance, but also through strategic sequencing and dynamic adaptation.

Third, we enrich the KBV (Grant, 1996) by demonstrating that knowledge access and

iew
appropriation are interdependent levers whose viability is shaped by structural positioning

and capability base. Industry-layer positioning (base, technology, or application) and

v
absorptive capacity (Cohen & Levinthal, 1990) jointly determine which access–appropriation

re
configurations are both feasible and sustainable in the AI industry. We show that absorptive

capacity can substitute for IP strength in enabling openness without excessive vulnerability,
er
while strong IP can act as a trust-enabler in other contexts. Embedding the access–

appropriation tension within these conditions allows us to specify boundary conditions for the
pe
robustness of different innovation “recipes” over time.

Finally, we make a methodological contribution by developing a replicable classification

ot

and coding system for industry-layer positioning in AI firms. This addresses a measurement
tn

gap: although government and industry reports (e.g., in China) recognize distinct AI value

chain segments, scholarly work has lacked a standardized operationalization. Drawing on

rin

official standards and prior applications of configurational analysis, we propose a coding

scheme that can be applied in future empirical studies of AI and other emerging industries,
ep

enabling more precise and context-sensitive theorizing.

5.3 Practical implications

Pr

27

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Importantly, our findings offer actionable guidance for managers navigating the paradox

ed
of openness in AI and other fast-evolving, knowledge-intensive sectors. First, there is no

single “best recipe” for balancing access and appropriation. Multiple pathways can deliver

iew
strong AI innovation performance, but the right one depends on a firm’s industry-layer

positioning—whether in the base, technology, or application layer. This industry structural

anchor shapes spillover risks, bargaining power, and the types of partnerships a firm can

v
re
sustain. Second, access must be tuned along two dimensions—breadth (the number and

diversity of partners) and depth (the extent of resource and information sharing)—and
er
matched with the right enabling mechanisms. Where absorptive capacity is high, firms can

afford greater access without excessive vulnerability. In more appropriation-sensitive

pe
environments, robust IP strength can actually facilitate trust and collaboration without

eroding competitive advantage. Third, strategies are dynamic in nature. Some, such as
ot

technology layer explore, tend to be structurally durable and can be pursued consistently.

Others, like closed innovation in the technology layer, may yield quick gains but are more
tn

vulnerable to shifts in regulation, market demand, or competitive intensity. Leaders should

actively monitor technological and regulatory changes, ready to rebalance access and
rin

appropriation as conditions evolve. Finally, treating access and appropriation as

interdependent elements of a coherent configuration—rather than independent levers—helps

ep

managers anticipate complementarities and trade-offs. The most effective strategies align

knowledge access and appropriation with structural positioning and capability base, ensuring
Pr

that each reinforces the other rather than pulling in opposite directions.

28

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 5.4 Limitations and future research

ed
As with all research, the design and methodological choices that we made in this study

inevitably involve limitations, which also open promising avenues for further research.

iew
First, our analysis is based on a sample of 150 AI firms active between 2020 and 2024.

While this window captures a formative period for the industry and allows for meaningful

v
temporal comparisons, it is necessarily constrained in both scope and duration. We prioritised

re
a manageable sample that enabled deep, multi-period data collection across multiple

constructs, but a broader and longer time frame would allow for testing whether the identified
er
configurations hold under different industry maturity stages or macroeconomic conditions.
pe
Future studies could extend this approach to larger, cross-industry samples or longitudinal

horizons to examine the generalisability and evolution of these configurations.

Second, we operationalised knowledge access and appropriation through measures of

ot

openness breadth, openness depth, and IP strength, complemented by absorptive capacity and
tn

industry-layer positioning. While these measures draw on established definitions (Cohen &

Levinthal, 1990; Laursen & Salter, 2006; Teece, 1986) they inevitably simplify complex
rin

underlying constructs. Our approach balanced theoretical validity with data availability, but

richer multi-source or qualitative indicators—such as partner-level openness measures or IP

ep

enforcement practices—could capture more nuanced mechanisms. Future research could

develop finer-grained, process-oriented measures to test and refine our configurations.

Pr

29

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Third, the industry-layer coding system, developed from government and industry

ed
classifications, provides a novel contextual variable but also introduces the potential for

classification error or ambiguity. We adopted a transparent coding protocol to ensure

iew
replicability and relevance to policy and managerial contexts. We recogonise, however, that

the the rapidly evolving nature of AI technologies means, in reality, some firms straddle

multiple layers. Future studies could explore more dynamic or multidimensional ways of

v
re
capturing industry positioning, possibly integrating network analyses of value-chain roles.

Fourth, we used MTP-QCA to identify equifinal configurations and their temporal

er
stability. QCA is well suited to uncovering conjunctural causality in small-to-medium

samples, but it necessarily abstracts from stochastic variation and continuous functional
pe
forms. We chose QCA because it aligns with our configurational research question and

enables temporal trajectories to be studied in a systematic way. Future research could

ot

complement QCA with econometric or simulation approaches to test the robustness of these

pathways under alternative assumptions.

tn

Finally, our measure of AI innovation performance focuses on outcomes observable in

rin

the short-to-medium term, such as patents, which captures innovation-relevant information

but may not reflect longer-term, systemic contributions to the ecosystem. This focus was
ep

necessary to ensure comparability across firms and time periods, but it leaves open the

question of how access–appropriation configurations influence broader innovation

Pr

trajectories. Future research could integrate alternative performance metrics, such as societal

or ecosystem-level outcomes, to assess the wider implications of knowledge strategies.

30

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 6. CONCLUSIONS

ed
This study examined how AI firms navigate the paradox of openness—the tension

between accessing external knowledge to innovate and protecting proprietary knowledge to

iew
capture value. Adopting a dynamic, configurational perspective and applying MTP-QCA, we

identified five equifinal strategic configurations to high AI innovation performance. These

v
configurations—spanning access-led, capability-led, and appropriation-led archetypes—

re
demonstrate that there is no universal “optimal” balance between access and appropriation.

Instead, effective strategies emerge from context-specific combinations of access mechanism,

er
appropriation mechanism, absorptive capacity, and industry-layer positioning. Dynamic

analysis shows that while some configurations, such as Leading Explorers in technology-
pe
layer, remains a stable strategy, others adapt their enabling mechanisms or fade entirely in

response to technological, market, or regulatory change. This underscores that the paradox of
ot

openness is not a static trade-off but an evolving challenge, requiring ongoing alignment

between knowledge strategies and shifting environmental conditions. Theoretically, we

tn

reconceptualize the access–appropriation tension as a configurational, dynamic challenge

rather than a static trade-off, extend paradox theory by showing how access–appropriation
rin

balances can persist, adapt, or vanish depending on contextual contingencies. We enrich the

KBV by specifying how industry-layer positioning and absorptive capacity jointly shape the
ep

sustainability of knowledge strategies in fast-evolving, knowledge-intensive sectors.

Pr

Methodologically, we advance a replicable classification system for industry-layer

positioning in AI, offering a foundation for comparative research in other technology-

31

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 intensive domains. For managers, our results provide a menu of viable “recipes” for

ed
reconciling access and appropriation, highlighting when and why certain strategies work—

and when they risk obsolescence. Ultimately, we suggest that thriving in AI innovation

iew
requires not only choosing the right balance between access and appropriation, but also

recognising when that balance must change.

v
CRediT authorship contribution statement

re
Qingliang Meng: Writing – original draft, Visualization, Methodology, Investigation,

Formal analysis, Data curation, Conceptualization. Xueyang Wu: Writing – original draft,
er
Visualization, Investigation, Formal analysis, Data curation, Conceptualization. Killian J

McCarthy: Writing – review & editing, Supervision, Investigation, Conceptualization.

pe
Ruling Zhang: Writing – review and editing, Supervision, Funding acquisition. Jiajun Sun:

Validation.
ot

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal
tn

relationships that could have appeared to influence the work reported in this paper.

Acknowledgments
rin

This work was supported by National Office for Philosophy and Social Sciences (No.

23AGL020), Jiangsu Office for Philosophy and Social Sciences (No. 22HQA1), Jiangsu
ep

Education Department (SJCX_1885). These financial supports are gratefully acknowledged.

Pr

Data availability

Data will be made available on request.

32

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 REFERENCE

ed
Adner, R., Kapoor, R., 2016. Innovation ecosystems and the pace of substitution:
Re‐examining technology S‐curves. Strateg. Manage. J. 37, 625-648.
Alexy, O., George, G., Salter, A.J., 2013. Cui bono? The selective revealing of

iew
knowledge and its implications for innovative activity. Acad. Manage. Rev. 38,
270-291.
Arora, A., Athreye, S., Huang, C., 2016. The paradox of openness revisited:
Collaborative innovation and patenting by UK innovators. Res. Policy 45, 1352-
1361.
Ayerbe, C., Boulos, C., Castellaneta, F., 2024. Navigating protection mechanisms and

v
innovation models: A literature-based configurational framework of intellectual

re
property strategies. Technovation 137, 103101.
Cai, S., Ma, H., 2022. Open paradox, Dominant design scenario and ambidextrous
innovation performance. J. Econ. Surv. 39, 98-107.
Chen, Y., Zhang, P., 2024. The innovation promotion effect of artificial intelligence
industry policies. Modern. Econ. Res., 69-79+132.
er
Chesbrough, H.W., 2003. Open innovation: The new imperative for creating and
profiting from technology. Harvard Business Press.
Cohen, W.M., Levinthal, D.A., 1990. Absorptive capacity: A new perspective on
pe
learning and innovation. Admin. Sci. Quart. 35, 128-152.
Devarakonda, S.V., Reuer, J.J., 2018. Knowledge sharing and safeguarding in R&D
collaborations: The role of steering committees in biotechnology alliances.
Strateg. Manage. J. 39, 1912-1934.
Ding, Y., Wu, J., 2022. Overcoming openness paradox in open networks: A
ot

configurational approach. J. Bus. Res. 150, 528-537.

Du, Y., Kim, P.H., 2021. One size does not fit all: Strategy configurations, complex
environments, and new venture performance in emerging economies. J. Bus. Res.
tn

124, 272-285.
Eisenhardt, K., Martin, J., 2000. Dynamic capability: What are they. Strateg. Manage.
J. 21, 1105-1121.
Enkel, E., Bader, K., 2012. Balancing open and closed innovation: Strategy and culture
rin

as determinants, ISPIM Conference Proceedings. The International Society for

Professional Innovation Management (ISPIM), p. 1.
Fiss, P.C., 2011. Building better causal theories: A fuzzy set approach to typologies in
organization research. Acad. Manage. J. 54, 393-420.
ep

Flatten, T.C., Engelen, A., Zahra, S.A., Brettel, M., 2011. A measure of absorptive
capacity: Scale development and validation. Eur. Manag. J. 29, 98-116.
Foege, J.N., Lauritzen, G.D., Tietze, F., Salge, T.O., 2019. Reconceptualizing the
paradox of openness: How solvers navigate sharing-protecting tensions in
Pr

crowdsourcing. Res. Policy 48, 1323-1339.

33

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Furnari, S., Crilly, D., Misangyi, V.F., Greckhamer, T., Fiss, P.C., Aguilera, R.V.,
2021. Capturing causal complexity: Heuristics for configurational theorizing.

ed
Acad. Manage. Rev. 46, 778-799.
Grant, R.M., 1996. Toward a knowledge‐based theory of the firm. Strateg. Manage. J.
17, 109-122.
Grimaldi, M., Greco, M., Cricelli, L., 2021. A framework of intellectual property

iew
protection strategies and open innovation. J. Bus. Res. 123, 156-164.
Hannah, D.P., Eisenhardt, K.M., 2018. How firms navigate cooperation and
competition in nascent ecosystems. Strateg. Manage. J. 39, 3163-3192.
Harwood, T.G., Garry, T., 2003. An overview of content analysis. Marketing. Rev. 3,
479-498.
Helfat, C.E., Peteraf, M.A., 2003. The dynamic resource‐based view: Capability

v
lifecycles. Strateg. Manage. J. 24, 997-1010.

re
Kobarg, S., Stumpf-Wollersheim, J., Welpe, I.M., 2019. More is not always better:
Effects of collaboration breadth and depth on radical and incremental innovation
performance at the project level. Res. Policy 48, 1-10.
Laursen, K., Salter, A., 2004. Searching high and low: what types of firms use
universities as a source of innovation? Res. Policy 33, 1201-1215.
er
Laursen, K., Salter, A., 2006. Open for innovation: The role of openness in explaining
innovation performance among UK manufacturing firms. Strateg. Manage. J. 27,
131-150.
pe
Laursen, K., Salter, A.J., 2014. The paradox of openness: Appropriability, external
search and collaboration. Res. Policy 43, 867-878.
Lewis, M.W., Smith, W.K., 2014. Paradox as a metatheoretical perspective: Sharpening
the focus and widening the scope. J. App. Behav. Sci. 50, 127-149.
Lundvall, B.-Å., Rikap, C., 2022. China's catching-up in artificial intelligence seen as
ot

a co-evolution of corporate and national innovation systems. Res. Policy 51,

104395.
March, J.G., 1991. How decisions happen in organizations. Human-computer
tn

Interaction 6, 95-117.
McCarthy, K.J., Aalbers, H.L., 2022. Alliance-to-acquisition transitions: The
technological performance implications of acquiring one's alliance partners. Res.
Policy 51, 104512.
rin

Meng, Q., Wu, X., Hang, Y., 2024. Study on the matching model of innovation
openness and intellectual property protection considering innovation context——
A fuzzy set qualitative comparative analysis based on multiple time periods.
Manage. Rev. 36, 157-170.
ep

Ming, X., Wang, Q., Liu, Y., 2025. The performance implications of R&D
collaborations on artificial intelligence. Technovation 145, 103271.
Miozzo, M., Desyllas, P., Lee, H.-f., Miles, I., 2016. Innovation collaboration and
appropriability by knowledge-intensive business services firms. Res. Policy 45,
Pr

1337-1351.

34

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Misangyi, V.F., Greckhamer, T., Furnari, S., Fiss, P.C., Crilly, D., Aguilera, R., 2017.
Embracing causal complexity: The emergence of a neo-configurational

ed
perspective. J. Manag. 43, 255-282.
Oxley, J.E., 1997. Appropriability hazards and governance in strategic alliances: A
transaction cost approach. J. Law. Econ. Organ. 13, 387-409.
Pisano, G.P., Teece, D.J., 2007. How to capture value from innovation: Shaping

iew
intellectual property and industry architecture. Calif. Manage. Rev. 50, 278-296.
Puntoni, S., Reczek, R.W., Giesler, M., Botti, S., 2021. Consumers and artificial
intelligence: An experiential perspective. J. Marketing. 85, 131-151.
Ragin, C.C., 2009. Redesigning social inquiry: Fuzzy sets and beyond. University of
Chicago press.
Rihoux, B., Ragin, C.C., 2009. Configurational comparative methods: Qualitative

v
comparative analysis (QCA) and related techniques. Sage.

re
Rothaermel, F.T., Alexandre, M.T., 2009. Ambidexterity in technology sourcing: The
moderating role of absorptive capacity. Organ Sci. 20, 759-780.
Schad, J., Lewis, M.W., Raisch, S., Smith, W.K., 2016. Paradox research in
management science: Looking back to move forward. Acad. Manag. Ann. 10, 5-
64.
er
Schneider, C.Q., Wagemann, C., 2010. Standards of good practice in qualitative
comparative analysis (QCA) and fuzzy-sets. Comp. Sociol. 9, 397-418.
Schneider, C.Q., Wagemann, C., 2012. Set-theoretic methods for the social sciences: A
pe
guide to qualitative comparative analysis. Cambridge University Press.
Teece, D.J., 1986. Profiting from technological innovation: Implications for
integration, Collaboration, Licensing and public policy. Res. Policy 15, 285-305.
Teece, D.J., 2018. Profiting from innovation in the digital economy: Enabling
technologies, standards, and licensing models in the wireless world. Res. Policy
ot

47, 1367-1387.
Tushman, M.L., O'Reilly III, C.A., 1996. Ambidextrous organizations: Managing
evolutionary and revolutionary change. Calif. Manage. Rev. 38, 8-29.
tn

West, J., 2003. How open is open enough?: Melding proprietary and open source
platform strategies. Res. Policy 32, 1259-1285.
Wu, H., Zhou, Y., 2021. Optimal degree of openness in open innovation: A perspective
from knowledge acquisition & knowledge leakage. Technol. Soc. 67, Article
rin

101756.
Wu, Q., Qalati, S.A., Tajeddini, K., Wang, H., 2025. The impact of artificial
intelligence adoption on Chinese manufacturing enterprises’ innovativeness: new
insights from a labor structure perspective. Ind. Manage. Data. Sys. 125, 849-874.
ep

Xie, X., Wu, Y., Martínez, J.M.G., 2023. More is not always better: Reconciling the
dilemma of R&D collaboration in high-tech industries in transition economies.
Technol. Forecast. Soc. Chang. 190, 122422.
Yacoub, G., Storey, C., Haefliger, S., 2020. Appropriability mechanisms for
Pr

manufacturing and service firms: the contingencies of openness and knowledge

intensity. R&D. Manage. 50, 551-572.

35

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 You, H., Meng, Q., 2018. Research on the “Stickiness” Relationship between R & D
and Profitablity of “Unicon” Company—Data Based on the AI Concept Stocks.

ed
Open. J. Bus. Manage. 6, 938-943.
Zahra, S.A., George, G., 2002. Absorptive capacity: A review, reconceptualization, and
extension. Acad. Manage. Rev. 27, 185-203.
Zeng, J., 2022. Artificial intelligence with Chinese characteristics: National strategy,

iew
Security and authoritarian governance. Springer.
Zhou, K., Qu, Z., Liang, J., Wang, S., 2025. Simplification or diversification: artificial
intelligence technological innovation and firm supply chain concentration. Econ.
Rev., 77-92.

v
re
er
pe
ot
tn
rin
ep
Pr

36

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Table 1 - Variable definition and measurement

ed
Variables Components Labels Measurements

AI Innovation Performance AI_Inno Number of AI patents applications

Openness Breadth OB Number of jointly patents applications partners

iew
Access Average times of jointly patents applications with
Openness Depth OD
partners

Composite measure of IP enforcement and R&D

Appropriation Intellectual property Strength IP

v
intensity

Industry-layer positioning Base Layer BL

re
Classification coding based on companies' most
Technology Layer TL
core business
Application Layer AL

Absorption Capacity
er AC Natural logarithm of the number of R&D staff
pe
ot
tn
rin
ep
Pr

37

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 e d
Figure 1 - Route to AI patent identification.
i ew
e v
r
AI Companies Patents Key Digital Technology Patent Classification System (2023)
Num:150 AD: 2022/1/1-2024/12/31 Artificial Intelligence Technology

iFlYTEK

Hikvision
ID: CN119517006A
IPC
G10L15/00
IPC
G10L15*
er Technical Branch
3.2 Intelligent Speech (3.2.1 Speech Recognition)

e
3.4 Biometrics (3.4.4 Voiceprint Recognition)
Cambricon 3.2 Intelligent Speech (3.2.2 Text to Speech)

p
G10L15/18 G10L25* 3.4 Biometrics (3.4.4 Voiceprint Recognition)
ZTE

BOE

FiberHome
G10L25/51

G10L15/06

o t G06F16*

…
1.1 Intelligent Chip
2.1 Machine Learning (2.2.1 Traditional Machine Learning)
2.2 Knowledge Graph
2.3 Brain-Inspired Intelligence Computing

n
3.1 Natural Language Processing (3.1.1 Machine
QAX Translation, 3.1.2 Semantic Understanding)

t
G06F16/334 … 3.4 Biometrics

ir n
… … …

e p
P r 38

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 e d
w
Table 2 - AI industrial chain classification coding system

Layers
Base
(Bas)
Descriptions
Providing AI infrastructure and
underlying technology support such as
Subdimension
Computing Power
Code
CP
Content

v i e
AI Chip, Intelligent Sensors, AI Computing Centers, Cloud Computing Services, Edge
Computing Nodes, Computing Center Network Architecture, etc.

e
hardware facilities, arithmetic Algorithms Alg Fundamental Algorithm Research (Supervised Learning, Reinforcement Learning

r
platforms, data foundation, etc. Frameworks), Distributed Algorithms (General Distributed, Distributed Machine Learning
Algorithms), etc.

r
Data Dat Data Collection, Data Annotation, Data Governance, Dataset and Database, etc.
Technology Developing AI technologies, large General-Purpose GPT Machine Learning, Knowledge Graphs, Computational Architecture, etc.

e
(Tec) models, and toolchains and platforms. Technology
Providing AI technology solutions for Domain-Specific DST Natural Language Processing, Intelligent Speech, Computer Vision, Biometric Feature

e
various application scenarios with Technology Recognition, Virtual Reality, Augmented Reality, Human-machine Hybrid Augmented
‘AI+’ as the business model. Intelligence, Multimodal, Swarm Intelligence, Embodied Intelligence, etc.

p
Large Model LM General-Purpose Large Models, Domain-Specific Large Models and Industry-Specific Large
Models

o t
Toolchain and Platform T&P Model Training Acceleration Tools, MLOps Platforms, Heterogeneous Computing
Framework Adaptation Tools, Model Compression Tools, AI Development Platforms, AI
Training Platforms, etc.

n
Application Building AI intelligent functions or Products and Service P&S Intelligent Robot, Smart Delivery Vehicles and Systems, Intelligent Mobile Terminal, Smart
(App) experiences on the basis of the original Services, Digital Human, etc.

t
scenarios to achieve business needs. Industrial Application IA Smart Manufacturing, Smart Homes, Smart Agriculture, Smart Transportation, Smart
Providing AI enabled products, Healthcare, Smart Education, Smart Office, Smart Energy, Smart Logistics, Smart Finance,

ir n
services or solutions with ‘+AI’ as a Smart Governance, Smart Cities, Smart Cultural and Tourism, Smart Security, Smart
service or innovation model. Building, etc.

e p
P r 39

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Table 3 - Coding examples and evidence display.

ed
Number Sample Citation of case evidence Code Layer

1 Fii "The main business includes cloud computing…", "...R&D capabilities CP Bas
for AI servers and data center cabinets..."(AR); "...five major
categories of the digital economy industry - cloud and edge

iew
computing..."(OW); "Significant revenue growth in cloud service
provider products, AI servers..."(SUS)
2 Hikvision "...computer vision, large models..., Hikvision helps digital DST Tec
applications in various industries...", "...won more than thirty T&P
international academic competitions in AI-related fields", "...Hikvision
AI Open Platform..."(AR); "The company started with video
technology..."(OW); "Our employees have published more than twenty

v
papers in top conferences in the field of computer vision and artificial
intelligence, such as CVPR, ECCV, and AAAI."(ESG)

re
3 Cambricon "...leading technology advantage of smart chip products..."(AR); CP Bas
"...R&D of AI chip products ... core processor chips in the field of
AI..."(OW); "AI chip TOP 10 by EE TIMES"(ESG); "AI chip"(Hurun
50); "AI chip"(CCID 100); "AI chip"(iResearch)
…
14 iFLYTEK "... intelligent speech, natural language processing, multimodal DST Tec
er
intelligence, cognitive intelligence...", "Spark Desk"(AR); "...core AI
technologies, such as intelligent speech, computer vision, natural
LM

language processing, cognitive intelligence..."(OW); "...standards in

the field of intelligent speech and artificial intelligence."(ESG);
pe
"NLP/Intelligent Speech"(Hurun 50); "Intelligent Speech"(CCID 100)
…
82 Qianxin "...AI security solutions ... AI assessment services...", "...industrial- P&S App
grade QAX-GPT Security Robot and large model guards...", "AI + IA
security"(AR); "...data security, AI + security..."(OW); " ...first
Holistic AI Security Response Program and QAX-GPT Security Robot
in China"(ESG)
ot

…
150 GEGTEST "big data AI inspection platform","...AI testing (AI ecological P&S App
validation, human-computer efficacy evaluation) ..."(AR);"intelligent IA
detection technology for industrial defects in high-end arithmetic
tn

chips"(ESG)
Note: AR for annual reports; OW for official website; ESG for ESG report; SUS for sustainability report; Hurun 50 for 2024
Hurun China Artificial Intelligence Enterprises Top 50; CCID 100 for 2024 CCID China Artificial Intelligence Enterprises
Comprehensive Strength Top 100; iResearch for 2023 iResearch Artificial Intelligence Industry Research Report of China.
rin
ep
Pr

40

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Table 4 - Descriptive statistics and calibration.

ed
Stages Conditios Max Min Mean SD Obs Fully in Crossover point Fully out
BL 1 0 150 1 - 0
TL 1 0 150 1 - 0
AL 1 0 150 1 - 0

iew
AC 10.417 3.219 7.134 1.258 150 7.9373 7.1639 6.2730
T1
OB 45.00 0.00 2.88 5.93 150 3.00 1.00 0.00
OD 129.00 0.00 6.72 20.45 150 3.31 1.00 0.00
IP 24.061 0.000 2.109 3.739 150 2.0873 0.9154 0.4465
AI_Inno (2022) 2561.00 0.00 118.13 358.66 150 65.25 18.00 3.00
BL 1 0 150 1 - 0
150

v
TL 1 0 1 - 0
AL 1 0 150 1 - 0
AC 10.500 3.497 7.210 1.275 150 8.0524 7.2162 6.3509

re
T2
OB 48.00 0.00 3.28 6.20 150 3.00 1.00 0.00
OD 149.33 0.00 5.53 18.54 150 2.76 1.00 0.00
IP 40.759 0.012 2.626 5.135 150 2.3878 1.1675 0.5240
AI_Inno (2023) 3403.00 0.00 113.98
er 419.95 150 61.00 18.50 4.00
BL 1 0 150 1 - 0
TL 1 0 150 1 - 0
AL 1 0 150 1 - 0
AC 10.474 3.689 7.219 1.267 150 8.0814 7.2363 6.3417
pe
T3
OB 47.00 0.00 3.15 6.12 150 4.00 1.00 0.00
OD 104.17 0.00 5.19 15.54 150 3.00 1.00 0.00
IP 74.348 0.000 3.440 7.381 150 3.2897 1.5809 0.6528
AI_Inno (2024) 1508.00 0.00 62.52 199.23 150 40.25 10.00 3.00
ot
tn
rin
ep
Pr

41

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Table 5 - Result of necessary analysis.

ed
High AI innovation performance

Conditions T1 T2 T3

Consistency Coverage Consistency Coverage Consistency Coverage

iew
BL 0.2757 0.5608 0.2687 0.5536 0.2579 0.5347
~BL 0.7243 0.4653 0.7313 0.4758 0.7421 0.4858
TL 0.1053 0.5931 0.0980 0.5592 0.0970 0.5569
~TL 0.8947 0.4782 0.9020 0.4883 0.9030 0.4919
AL 0.6190 0.4488 0.6333 0.4651 0.6450 0.4766
~AL 0.3810 0.5694 0.3667 0.5551 0.3550 0.5406

v
AC 0.6863 0.6642 0.6873 0.6716 0.6784 0.6674
~AC 0.4206 0.4144 0.4163 0.4168 0.4039 0.4065

re
OB 0.6436 0.6864 0.6926 0.7275 0.6239 0.6844
~OB 0.4584 0.4127 0.3937 0.3678 0.4707 0.4286
OD 0.6459 0.6750 0.6902 0.7241 0.6528 0.6901
~OD 0.4506 0.4129 0.3940 0.3684 0.4293 0.4035
IP 0.6000 0.6110 0.5878 0.6015 0.5913 0.6081
~IP 0.5070 0.4755
er
0.5235 0.5009 0.5036 0.4853

Note: ~ indicates the non-set of a logical operation.

pe
ot
tn
rin
ep
Pr

42

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Table 6 – Configurations for High AI innovation performance (T1)

ed
High AI innovation performance (T1)
Conditions
M1 M2 M3 M4 M5
BL ● ● U U U

TL U U ● ● U

iew
AL U U U U ●

AC ● ● ● ●
OB ● ● U ● ●
OD ● ● U ● ●

IP ● ● ● ●
BL*~TL*~AL BL*~TL*~AL ~BL*TL*~AL ~BL*TL*~AL ~BL*~TL*AL

v
Configurations
*OIB*OID*IPR *AC*OIB*OID *AC*~OIB*~OID*IPR *AC*OIB*OID*IPR *AC*OIB*OID*IPR

re
Consistency 0.8917 0.9038 0.8736 1.0000 0.8283
Raw coverage 0.0855 0.1065 0.2076 0.0410 0.2003
Unique coverage 0.0287 0.0597 0.0172 0.0374 0.2003
Solution coverage 0.3937
Solution er 0.8576
consistency

Symbol description: "●" represents the presence of conditions, "U" represents the absence conditions; A large circle represents the
core conditions, a small circle represents the edge conditions, and a space represents that the presence and absence of conditions
pe
do not affect the results.
ot
tn
rin
ep
Pr

43

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Table 7 – Configuration Descriptions

ed
Number Name Key Characteristics Theoretical Logic / Strategic Implication
Openness fosters knowledge inflows, while strong IP
BL: base layer firms safeguards enable the firm to capture returns from shared
The Open
OB: high openness in breadth innovations (appropriability theory; Teece, 1986). Often

iew
M1 Platform
OD: high openness in depth platform-based providers leveraging wide and deep
Integrators.
IP: reinforced by strong IP collaborations while protecting core algorithms or
architectures.
BL: base layer firms Openness benefits are realized through the ability to
The OB: high openness in breadth assimilate and exploit external knowledge (Cohen &

v
M2 Collaborative OD: high openness in depth Levinthal, 1990). Common in fast-evolving technology
Absorbers AC: paired with high absorptive ecosystems where learning speed outweighs formal IP

re
capacity protection.
TL: technology layer firms
Focus on a closed innovation model to drive specialized AI
OB: high openness in breadth
innovation (market segment securing defender strategic
The Closed OD: shallow openness in depth
M3
Specializers
er
IP: reinforced by strong IP
archetype; Enkel & Bader, 2012). Seen in AI technological
firms whose competitive advantage depends on exclusive
AC: paired with high absorptive
technologies, unique algorithms, or specialized datasets.
capacity
pe
TL: technology layer firms
Scans a wide range of external sources for ideas and
OB: high openness in breadth
technologies, a tight collaboration relationship for tacit
The Leading OD: high openness in depth
M4 knowledge, consistent with exploration logic (March,
Explorers IP: reinforced by strong IP
1991). Suitable for firms testing and integrating emerging
AC: paired with high absorptive
ot

AI technology across multiple industries.

capacity
AL: application layer firms
tn

OB: high openness in breadth Leverages proximity to markets and proprietary protection
The IP-Focused OD: high openness in depth to monetize downstream AI products (Pisano & Teece,
M5
Market Makers IP: reinforced by strong IP 2007). Common in applied AI providers targeting end-user
AC: paired with high absorptive markets such as healthcare diagnostics.
rin

capacity
ep
Pr

44

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Table 8 – Configurations and types in multiple time periods analysis (T1, T2, T3)

ed
Raw Unique
Types Periods Configurations Consistency
coverage coverage
2021-2022 ~BL*TL*~AL*AC*OB*OD*IP (M4) 0.0410 0.0374 1.0000
2022-2023 ~BL*TL*~AL*AC*OB*OD*IP 0.0414 0.0414 1.0000

iew
2023-2024 ~BL*TL*~AL*AC*OB*OD*IP 0.0413 0.0413 0.9935
T1 T2 T3 Typical cases: iFLYTEK, Hikvision, Dahua Tech
Stable
2021-2022 ~BL*~TL*AL*AC*OB*OD*IP (M5) 0.2003 0.2003 0.8283
2022-2023 ~BL*~TL*AL*AC*OB*OD 0.3368 0.3368 0.8517
2023-2024 ~BL*~TL*AL*AC*OB*OD*~IP 0.1919 0.1919 0.8360
T1 T2 T3 Typical cases: QAX, Kingsoft Office

v
2021-2022 BL*~TL*~AL*OB*OD*IP (M1) 0.8548 0.0287 0.8917
2022-2023 - - - -

re
2023-2024 BL*~TL*~AL*OB*OD*IP 0.1000 0.0860 0.8373
T1 T2 T3 Typical cases: ZTE, Jingjia Micro, VeriSilicon
Adoptive
2021-2022 BL*~TL*~AL*AC*OB*OD (M2) 0.1065 0.0497 0.9038
2022-2023 BL*~TL*~AL*AC*OB*OD 0.1273 0.0503 0.8621
2023-2024 - - - -
T1 T2 T3
er
Typical cases: FiberHome, Chinsunicom, BOE, SMIC
2021-2022 ~BL*TL*~AL*AC*~OB*~OD*IP (M3) 0.0208 0.0172 0.8736
2022-2023 - - - -
pe
2023-2024 - - - -
T1 T2 T3 Typical cases: Newland, GOODiX
2021-2022 - - - -
2022-2023 BL*~TL*~AL*OB*OD*~IP 0.1006 0.0236 0.8153
Transient
2023-2024 - - - -
T1 T2 T3 Typical cases: Sugon, E-hualu, fii-foxconn, Inspur, [Link]
ot

2021-2022 - - - -
2022-2023 - - - -
2023-2024 BL*~TL*~AL*AC*~OB*~OD*IP 0.0362 0.0222 0.8308
tn

T1 T2 T3 Typical cases: HT-tech, greatwall

rin
ep
Pr

45

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Table 9 – Robustness Checks: Alternative Threshold Settings

ed
Periods Configurations Consistency > 0.85 Consistency PRI > 0.8 Consistency
T1 BL*~TL*~AL*OB*OD*IP (M1)
BL*~TL*~AL*AC*OB*OD (M2) P 0.9038 P 0.9038
~BL*TL*~AL*AC*~OB*~OD*IP (M3) P 1.0000 P 1.0000

iew
~BL*TL*~AL*AC*OB*OD*IP (M4) P 0.8736 P 0.8736
~BL*~TL*AL*AC*OB*OD*IP (M5)
T2 BL*~TL*~AL*AC*OB*OD (P) 0.9623 (P) 0.9623
BL*~TL*~AL*OB*OD*~IP (P) 0.9623
~BL*TL*~AL*AC*OB*OD*IP P 1.0000 P 1.0000

v
~BL*~TL*AL*AC*OB*OD P 0.8517 (P) 0.8818
T3 BL*~TL*~AL*OB*OD*IP (P) 0.9053 (P) 0.9053

re
BL*~TL*~AL*AC*~OB*~OD*IP
~BL*TL*~AL*AC*OB*OD*IP P 0.9935 P 0.9935
~BL*~TL*AL*AC*OB*OD*~IP
er
Note: ✓ is placed in cases where the same solution occurs. (✓) is placed where the solution term is comparable, but one
condition is added or removed vs. initial analysis.
pe
ot
tn
rin
ep
Pr

46

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]