CS4: AI, Finance and Investment in Energy Transition

This study investigates how data--driven and AI methodologies have evolved to address the shifting paradigms of global energy security and transition risks over the past half-century. Specifically, it examines the co-evolutionary relationship between emerging energy crises-ranging from the oil shocks of the 1970s to the current complexities of renewable energy integration-and the technical advancements in data-driven decision-making.
Employing a systematic narrative review, we analyze five distinct stages of energy security concerns: (1) 1970s oil supply disruptions, (2) 1980s–1990s market liberalization, (3) post-1997 climate and environmental risks, (4) 2010s renewable integration challenges, and (5) the current AI-driven transformation. Within each stage, we compare the application, strengths, and limitations of various approaches, from early statistical methods to modern deep learning, Bayesian networks, and spatio-temporal hazard mapping.
The review reveals a clear shift from reactive, supply-centric risk modeling to proactive, system-wide resilience optimization. Findings indicate that while early methods focused on linear forecasting for fossil fuel markets, current AI-driven research primarily targets the intermittency of variable renewable energy (VRE) and grid stability. Specifically, AI-driven predictive maintenance and advanced load forecasting have significantly reduced operational risks. However, the study identifies a complexity-interpretability gap and highlights underdeveloped governance frameworks regarding data security and AI transparency as persistent bottlenecks in the transition toward a de-carbonized energy system.
The research provides a road-map for scholars and policymakers to leverage AI for mitigating transition-induced vulnerabilities. It proposes new research directions to address the systemic uncertainties towards the 2030 energy landscape. By bridging the gap between historical data-driven successes and future technological needs, this study informs strategies for achieving a resilient and secure global energy transition.

The rapid diffusion of large language models (LLMs) has intensified concerns about the environmental sustainability of artificial intelligence. While existing research has documented the rising energy demand of model training and inference, far less attention has been paid to how efficiently this energy is transformed into useful cognitive output. This paper introduces a novel metric-final-to-useful exergy efficiency-to quantify the proportion of electrical energy consumed during inference that produces valid and informative task-relevant results. By integrating thermodynamic reasoning with information-theoretic measures, the study moves beyond conventional energy-per-token accounting and evaluates the energy-to-utility conversion of AI systems.
Inference energy is estimated using an empirical decomposition approach that combines model computational complexity with generational improvements in GPU efficiency (GFLOPs/W), enabling consistent comparison across both open-weight and proprietary models. FLOPs per token for open models are drawn from published sources, while those of closed models are inferred through scaling relationships calibrated on comparable architectures. To capture the "useful" component of energy, the framework defines two complementary indicators: R(valid), the proportion of output tokens contributing to successful task completion, and I(gain), the normalized information increment relative to the input, computed using a reference language model. Their product, scaled by datacenter overhead (PUE), yields a unified efficiency index that can be compared across architectures, model sizes, and task types.
The analysis is expected to reveal systematic differences in energy productivity across model generations and architectural designs, highlighting the trade-offs between scale, performance, and sustainability. By linking energy consumption directly to informational value creation, the proposed framework offers regulators and industry stakeholders a principled basis for benchmarking "green AI" performance and designing carbon-aware deployment strategies.

The transition to sustainable energy systems depends on scaling innovative technologies in the electricity sector, yet bringing these innovations to market requires significant financial support and strategic guidance. This paper demonstrates how the implicit investment criteria of Breakthrough Energy Ventures (BEV), founded by Bill Gates and widely regarded as the benchmark venture capital fund in climate innovation, can be reverse-engineered and systematically applied to a broader dataset of electricity startups. Drawing on the concept of revealed preferences, we develop a three-pronged methodology: Qualitative Comparative Analysis (QCA), decomposition of startup attributes, and AI-assisted similarity modeling using large language models (LLMs). The framework is applied to a curated dataset of 743 electricity startups, extending earlier work (Fuentes, 2023), to assess whether BEV's portfolio choices can be reliably reproduced. Across all methods, convergence occurs on a very small subset of startups (1.3% of the database), underscoring the discriminative power of the approach. The study introduces a scalable and transparent methodology to emulate high-level investment logics. The findings suggest that AI-assisted tools can reduce information barriers and democratize access to strategic investment intelligence, with implications for policy, mission-driven capital, and the governance of AI in clean energy finance.

Access to clean and affordable energy remains a critical development challenge in Uganda, particularly in rural areas where electrification and clean cooking adoption rates remains low. While financial sector development is often considered a catalyst for investment and infrastructure expansion, there is limited empirical evidence in low-income economies. This study investigates the impact of financial development and private sector credit on access to clean energy in Uganda.
Using time series data obtained from the World Development Indicators and the Bank of Uganda, the study employs the Autoregressive Distributed Lag (ARDL) modelling framework to examine both short run and long run relationships. Clean Energy access is provided by access to electricity and renewable energy consumption, while financial development is measured by domestic credit to the private sector and financial inclusion indicators. Control variables include Gross Domestic Product per capita, Foreign Direct Investment, trade openness, inflation and urbanization.
The results from the study are expected to reveal whether financial deepening significantly enhance clean energy access. Preliminary results suggest that private sector credit plays a critical role in financing renewable energy infrastructure and household energy adoption. This study contributes to the literature on energy finance in Sub-Saharan Africa by providing country specific empirical evidence and recommendations for policy makers seeking to strengthen clean energy transition.

This research establishes an integrated techno-financial framework to evaluate the investment viability and risk profile of large-scale underground hydrogen storage in depleted gas reservoir. With the growing role of hydrogen in low-carbon energy systems, large-scale and long-duration storage infrastructure requires rigorous financial assessment to mobilize private capital and inform policy design. However, existing studies rarely link reservoir-scale technical design and performance with project-level financial valuation and risk analysis.
We combine high-resolution reservoir simulation with project finance modeling to translate physical storage behavior into cash flow projections, capital requirements, and risk-adjusted returns. A depleted sandstone reservoir in Colorado, USA, is used as a representative case. Simulation outputs are systematically converted into financial and modeling inputs for estimating capital and operating costs, capacity utilization, and operational constraints. The project is structured using limited-recourse project finance, with the cost of capital derived from CAPM-based equity valuation and weighted average cost of capital.
The framework evaluates net present value, internal and modified rates of return, and payback periods under market and policy uncertainty. Financial risk is quantified using sensitivity and Monte Carlo analyses to identify key value drivers, including hydrogen price spreads, cushion gas requirements, and fiscal incentives.
This study finds that, under representative market and policy conditions, large-scale underground hydrogen storage in depleted gas reservoir can achieve positive risk-adjusted returns, with strong sensitivity to policy incentives and key financial parameters. The proposed framework provides a transferable decision-support tool for investors, lenders, and policymakers to assess hydrogen storage projects under uncertainty. By explicitly integrating technical performance with financial optimization, this study contributes to the emerging literature on energy finance for low-carbon infrastructure.