Private Capital or Public Policy Support? Firm-Level AI Infrastructure Choice in the Sovereign AI Era

• In Event: Technology Diffusion, Industry, and Global Competition

Abstract

As governments from Washington to Brussels to Seoul assert sovereignty over domestic AI processing capacity, a foundational question remains untested. Does public policy actually shift how firms invest in AI infrastructure? The question confronts multiple levels of government, from national sovereign AI initiatives to U.S. states and EU member states developing their own frameworks. Firm-level choices among proprietary GPU facilities, global cloud platforms, and hybrid configurations collectively determine whether AI capacity stays under domestic control or migrates to foreign-operated data centers. Yet sovereign AI initiatives worldwide have focused almost exclusively on frontier model developers. Mid-tier and service firms remain largely unexamined, even though they constitute the majority of AI infrastructure demand and determine where computational sovereignty actually resides.

This study asks whether dispersed public support altered firm-level infrastructure investment, or whether private capital and market structure remained the operative forces. We examine South Korea as a critical case. Prior to Korea's 2025 national consortium, which bundled GPUs, data, and talent into a coordinated program, public support had been dispersed across grants, data provision, and fragmented initiatives. We draw on the 2023 AI Industry Survey, a census-scale dataset covering the full firm population two years before the consortium took effect. The analysis proceeds in three stages. First, latent class analysis identifies strategic firm types based on capability, resource, and market-position indicators. Second, the BCH (Bolck-Croon-Hagenaars) approach, a bias-adjusted three-step method correcting for classification uncertainty, characterizes infrastructure patterns across classes. Third, multinomial logit models estimate the effects of public and private resources on infrastructure choice, with class membership as a covariate.

Four firm types emerge from the analysis. Hardware Producers account for 1 percent of the population, Application-Centric Mid-Tier Firms 51 percent, Public-Sector Service Integrators 30 percent, and Capital-Independent Large Firms 18 percent. A consistent pattern follows across these classes. Neither government funding nor public data provision shows meaningful association with infrastructure choice. Instead, private data partnerships and next-generation GPU scarcity emerge as the operative forces.

Pre-consortium policy announced national intent without changing market outcomes, and AI capacity remained concentrated in foreign-operated cloud platforms. This outcome sits uneasily with the sovereignty rationale behind public investment. The subsequent move toward large-scale coordinated investment, exemplified by Korea's 2025 consortium and parallel initiatives in the U.S. and EU, represents a defensible direction. Our findings suggest, however, that such investments will shift firm behavior only when paired with two conditions. One is sustained attention to upstream semiconductor supply chains shaping GPU availability. The other is policy instruments that engage rather than bypass the private data partnerships through which mid-tier firms actually build capacity. The three-stage framework offers a replicable tool for jurisdictions assessing whether their AI investments meet these conditions.

Authors

• keunho Jang, Sungkyunkwan University
  Presenting Author

• Rosa Minhyo MINHYO Cho
  Non-Presenting Co-Author