Conference Agenda

Recent innovations in artificial intelligence (AI) present new opportunities to study how firms choose their innovation paths. Innovations in AI have been occurring at least since the 1950s, but the advent of foundation models (FMs) represents a structural break in this evolution. A diversity of firms—from large technology incumbents such as Meta to new entrants like DeepSeek—have entered the race to develop FMs. One feature of this competition is a firm choosing its openness features, i.e., the degree to which the model’s components, such as weights, training data, code, and documentation, are accessible, modifiable, and transparent for others, allowing downstream developers to build generative AI applications at minimal cost.

That open-source FMs differ in their choices of degrees of openness implies differences in their beliefs about costs and potential revenue. It is generally accepted that FMs incur substantial sunk costs in training their large-scale models. FMs in general receive revenue from Application Programming Interfaces (API) usage (Per-Token Fees), subscriptions to consumer-facing services, such as OpenAI’s ChatGPT Plus, and fees charged for allowing companies to customize and fine-tune the FM model using proprietary data. All other things being equal, choosing to close an FM allows the owner to generate more short-run revenue by charging API fees to downstream developers.

But openness involves tradeoffs for FMs. Some downstream developers who value openness will adopt open-source FMs for customization and product-specific fine-tuning. The FM owner can then learn from such developers’ use and leverage that knowledge to subsequently improve the next-generation FM, consequently increasing demand. Xu et al. (2025) describe this as a “data flywheel effect.” Openness also comes at a cost for an FM: openness could benefit rivals, i.e., the process by which the generalized knowledge embedded within the FM spreads, leaks, or is repurposed by downstream developers, researchers, and other firms that may compete with the FM owner’s offerings. More specifically, rivals may benefit from the knowledge spillovers, leading to less investment and lower welfare.

Despite the central role of openness in shaping competition among foundation models, the trade-off between its short-run costs and long-run payoffs remains largely underexplored in the empirical literature. This paper fills this gap by developing and estimating a two-stage dynamic structural model of FM competition. In the first stage, FM owners choose the degree of openness of their models. In the second stage, downstream AI developers choose which FM to adopt for application development. Openness enters both stages. On the demand side, it affects developers’ adoption choices by changing accessibility, customization costs, and deployment flexibility. On the dynamic side, openness shapes the future state of competition by influencing the magnitude of knowledge spillovers and the evolution of subsequent model quality. The model therefore allows the private value of openness to depend not only on contemporaneous payoffs, but also on the discounted value of future ecosystem rents.

The empirical analysis combines data from Hugging Face with information collected from published model documentation. These sources are used to construct measures of openness, model quality, and downstream adoption. The identification strategy addresses two challenges. First, because FM markets are nascent, the time dimension of the data is relatively short. However, the rapid pace of innovation and frequent turnover across model generations provide substantial variation in relevant state variables over a compressed time horizon. Second, openness choices may be endogenous if firms respond to unobserved demand shocks. The paper addresses this concern by exploiting the production cycle of FMs: because pre-training, safety alignment, and release decisions require meaningful lead times, openness is plausibly predetermined relative to contemporaneous short-run shocks. The empirical specification also includes owner and time fixed effects to absorb persistent heterogeneity across firms and aggregate market trends.

We find that short-run returns alone are insufficient to rationalize observed openness choices. Preliminary estimates suggest that the immediate benefits of openness account for less than half of its contemporaneous costs, implying that a static profit framework cannot explain why firms open their models to the extent observed in the data. Instead, firms appear to place substantial value on the long-run benefits of openness. The main mechanism is an intertemporal feedback loop: greater openness expands downstream adoption, downstream activity generates knowledge spillovers, those spillovers improve future model quality, and higher future quality attracts additional downstream adoption in later periods. This feedback mechanism explains why firms may rationally sacrifice short-run profits in exchange for future market expansion and suggests that firms choose open-source strategies as a long-run pathway to innovation.