Innovation in AI accelerator platforms: patent-based analysis of innovation dynamics to inform goverance

• In Event: Infrastructure, Industrial Policy, and the Political Economy of AI

Abstract

Artificial intelligence (AI) accelerators, including graphics processing units (GPUs), tensor processing units (TPUs), and related domain-specific architectures, are increasingly becoming core infrastructure for modern economies and national security. The geographic concentration of production and excludability of these chips has motivated industrial policy, including the U.S. CHIPS and Science Act, the EU Chips Act, and China’s National Integrated Circuit Industry Investment Fund. While these interventions primarily target fabrication, developing accelerators requires not only manufacturing but also complex design capabilities shaped by the interactions between hardware architecture and low-level software stacks.  

To inform policy interventions targeting chip design capabilities, this paper conceptualizes AI accelerator design as a platform where innovation emerges through tight coupling between hardware architecture and low-level software systems. While existing platform innovation literature emphasizes modular interfaces and complementor ecosystems, innovation in AI accelerators occurs through cross-layer co-design, rather than modular development across a stable boundary. Due to this coupling, shocks in one layer can fundamentally alter the direction of innovation in another layer. This distinction has critical implications for industrial policy and the political economy of technological leadership.  

This paper asks whether discrete software-layer shocks in AI accelerator platforms are associated with systematic changes in hardware innovation within and across platforms. We characterize software shocks as within-platform (e.g., new capabilities in CUDA) or ecosystem-wide (e.g., the public release of TensorFlow). Although innovation across layers is inherently bidirectional, we treat software-layer changes as observational shocks to examine corresponding shifts in hardware design trajectories. To measure these trajectories, we construct a dataset of AI accelerator hardware designs using U.S. patent data via keyword and classification code search, followed by a large-language-model-based classification pipeline. To measure the direction of innovation, we apply Latent Dirichlet Allocation topic modeling to patent claims, allowing documents to be assigned to multiple topics. The analysis tracks changes in topic composition, entry, and similarity across firms and over time.  

Preliminary results indicate cross-layer influence. Software shocks originating within a platform are associated with the strongest shifts in hardware innovation within that platform, with weaker but occasionally detectable spillovers to competitors. Higher-level software shocks are associated with broader changes in hardware innovation across platforms, suggesting a broader reorientation of hardware design. We also document heterogeneity in topic composition across platforms, indicating differentiated technological trajectories. These findings provide empirical evidence of the tight coupling between low-level software and hardware architecture, supporting the view of AI accelerator design as platforms.  

For policymakers, the results highlight tradeoffs among performance gains from tightly coupled innovation, cross-platform portability for competition, and the preservation of innovation diversity. Heterogeneity in technological trajectories suggests that coupling not only improves performance but also supports differentiated innovation paths. Enabling cross-platform portability, rather than enforcing cross-layer compatibility, may strengthen competition without forcing convergence towards a single design. However, policies promoting openness may accelerate the diffusion of capabilities across borders. Overall, this study provides a more granular foundation for industrial policy targeting AI accelerator design.

Author

• Anurag Panda, ETH Zurich
  Presenting Author