Discussion

This study provides the first consistent global and country-level ER estimate and explanation for geothermal power, based on a dataset of 514 plants and 51 expert interviews. Our analysis shows that geothermal power exhibits limited global learning, even though substantial local learning is possible under specific conditions. While global costs have not consistently declined with deployment, several countries, most notably the United States, feature positive ERs due to favorable geological conditions, policy environments, and knowledge transfer from the oil and gas industry. The large variance in local ERs underscores that geothermal learning is highly context-specific, shaped by the interaction of technological, geological, and institutional conditions. As such, geothermal power is unlikely to scale globally in the same way as solar PV or wind, and thus our results cast doubt on the IEA’s 80% cost reduction target by 2035.4 Nevertheless, geothermal could become the primary firm renewable power source in specific regional markets where key preconditions are met, such as the United States. There, a recent study about EGS and based on ER assumptions foresees cost reductions around 50%–75% by 2050, down to around 2 USD/W, even when considering the declining availability of resources.8 Given that there is sufficient buildout, our empirical results, which showed strong local learning in the United States, support these projections.

These results have implications for multiple audiences, most importantly policymakers, energy system modelers, and scholars of technological learning.

For policymakers, first, geothermal expansion strategies should be grounded in an assessment of regional capabilities and constraints. In areas with favorable hydrogeology—particularly, large and homogenous reservoirs—and an active oil and gas industry, geothermal can scale more easily and cost-effectively. Second, policy can actively improve learning conditions. Besides stable regulation, support for early-stage exploration (e.g., through risk guarantees or public funding), open access to geological and cost data, and incentives for knowledge sharing across projects reduce the need for customization. In addition, partnerships of geothermal actors with the oil and gas industry can strengthen local innovation systems. Examples include facilitating workforce transfer from oil and gas, co-developing modular components, or reusing existing infrastructure where feasible.

These insights extend to future geothermal technologies. While EGS and CLGS are often assumed to improve learning by decoupling performance from subsurface uncertainty,55,56 our results indicate that their potential still depends on managing customization needs and project design complexity. CLGS appears to offer lower customization due to its sealed-loop design and adaptability to user demands, but drilling technology still requires local adjustment. EGS may face continued challenges with fracture stimulation and regulatory approval, similar to those encountered in unconventional oil and gas extraction, which showed high local ERs57,58 and buildout in the US, but did not spread globally. The reasons why unconventional oil and gas extraction could not replicate its strong progress elsewhere parallel those that we find as challenges in EGS: Subsurface heterogeneity preventing fast spillovers, geological conditions unique to the US, especially in resource quality (shale thickness) and ease of accessing it (how easily rock can be fractured), weaker industrial structure outside the US, a lack of strong investors with a high risk appetite outside the US, and increased difficulty of logistics.59,60 Nuclear power offers a second instructive analogy, as it faces the tension between customized and generic design without geological causes. Where reactor programs committed to standardized designs and serial construction by stable supplier and constructor networks, such as France’s fleet buildout in the 1970s and 1980s or South Korea’s standardized reactor series, construction costs remained stable or declined.61,62 Where designs were customized to individual sites, regulatory regimes, and utilities, most prominently in the United States, costs escalated, and learning rates turned negative.62,63 The nuclear case shows that the balance between customization and standardization is not fixed by technology alone but shaped by industrial structure, regulatory stability, and procurement strategy. This suggests that the learning prospects of EGS and CLGS will also depend on institutional choices, not only on their technical characteristics. A third analogy comes from biomass power, where knowledge transfer to new regions is strongly limited by the need to redesign core components for a different, locally produced feedstock.46 Future work should investigate how emerging design innovations, such as plasma drilling,64 modular surface systems,30 or standardized project design,45 can help shift geothermal development toward more scalable and replicable trajectories.

From an energy system modeling perspective, our results challenge the default assumption of a single, global cost level and ER across all technologies, as it is currently the standard for integrated assessment models.65,66 For technologies with high customization needs, models ignoring geographic cost heterogeneity and locally confined learning are likely overestimating cost reductions. Accordingly, modelers should consider regionally differentiated experience curves or deploy context-specific cost projections. For geothermal, we advise modelers to consider the factors that drive geothermal learning dynamics from Table 1 for each region separately. Regions with heterogeneous hydrogeology, high regulatory density and variance, and no knowledge transfers from an existing oil and gas industry should be assigned higher ERs than those where the opposite applies. At minimum, this would mean separating the contiguous US from the rest of the world. If a global ER has to be applied, we recommend using a comparably low one for all geothermal technologies.

Finally, for scholars of technological learning, our findings highlight the importance of moving beyond global and paying more attention to local learning dynamics.67,68 Much of the literature has focused on estimating global ERs and decomposing cost reductions into learning-by-doing, R&D, and scale effects. However, the potentially substantive variance in learning across countries remains underexplored69—at least for technologies with a high need for customization.32 Our study contributes to addressing this gap by linking observed cost trajectories to underlying factors through a mixed-methods research design. We show that diminishing productivity of new sites, varying hydrogeology, regulatory density and variance, and tacit knowledge are not simply context-specific challenges. Instead, they systematically influence a technology’s capacity to accumulate spatially transferable experience. Recognizing the structural basis for these differences helps explain why learning is not globally uniform.

Future work with an even more detailed geothermal dataset could include more granular estimations of ERs, e.g., with a division into geological basins. Furthermore, an enhanced database could allow deriving multi-component ERs and multi-factor ERs, the latter of which could include spillover effects from the oil and gas industry. If carried out successfully, these experiments could even further inform modeling of highly customized technologies.

Limitations of the study

Our data only present a subsample of all geothermal plants, and the completeness of data varies strongly between countries and decreases for older projects. Despite this, our dataset is still substantially more complete than previous ones. For future technologies, there are more factors determining experience rates apart from complexity and the need for customization, even if these two factors are the most important ones. In addition, different forms of spillovers from other sectors are conceivable; however, these are hard to model, and our approach constitutes a more accurate way of forecasting technology cost than expert elicitation.18