Artificial intelligence (AI) is often discussed as if it were weightless: code, models, and interfaces floating in the cloud. In reality, AI is becoming one of the most electricity-intensive growth vectors in the digital economy. That does not make AI inherently incompatible with sustainability. It does mean, however, that investors can no longer treat compute as an abstract software input. Electricity availability, grid carbon intensity, temporal matching, cooling efficiency, and network constraints are becoming financially material determinants of which AI business models are durable and which are merely fast-growing. For sustainable investors, the core question is whether the energy system supporting AI is being built in a way that is economically resilient, climatically credible, and socially defensible.
The scale shift is already visible. The International Energy Agency (IEA, 2025) estimates that data centers consumed about 415 terawatt-hours (TWh) of electricity in 2024, or roughly 1.5% of global electricity use, and that data-center electricity demand is set to more than double to around 945 TWh by 2030. The IEA also notes that global investment in data centers nearly doubled since 2022, reaching about half a trillion dollars in 2024. In the US, the IEA expects data centers to account for nearly half of electricity demand growth through 2030. In other words, AI is not just a semiconductor story or a software story; it is increasingly a power-sector story (IEA, 2025).
That framing matters because “the energy cost of AI” is often misunderstood. The headline number, total TWh consumed, is important but incomplete. For investors, the more relevant variables are where demand shows up, when it shows up, how flexible it is, what infrastructure must be built to serve it, and what emissions profile attaches to that supply. The IEA underscores that AI-focused data centers are much more geographically concentrated than traditional industrial loads: nearly half of US data-center capacity sits in five regional clusters, and half of US projects under development are in pre-existing large clusters. This concentration makes local bottlenecks, rather than aggregate national averages, the real locus of risk. A sustainable investment lens therefore has to be spatial and temporal, not just volumetric.
The US illustrates both the opportunity and the strain. Lawrence Berkeley National Laboratory’s 2024 update estimates that US data centers used 4.4% of total US electricity and projects a rise to roughly 325-580 TWh by 2028, depending on the scenario. EPRI’s more recent 2026 update puts 2024 US data-center electricity use at roughly 177-192 TWh and projects 380-790 TWh by 2030; its scenarios imply data centers could consume 9-17% of US electricity by 2030, up from roughly 4-5% today. The gap between Berkeley Lab’s and EPRI’s scenarios is itself instructive: even reputable estimates vary materially, which means investors should price uncertainty, not just point forecasts (EPRI, 2026; Shehabi et al., 2024).
What turns this from a general ESG issue into a portfolio-construction issue is the speed mismatch between compute growth and power-system buildout. The IEA estimates that around 20% of planned data-center projects could face delays if electricity-sector bottlenecks are not resolved. It also notes that building new transmission in advanced economies can take four to eight years, while wait times for transformers and cables have doubled over the last three years. FERC’s recent market reporting shows that active interconnection queues in the US still totaled 2,289 GW at the end of 2024, even after a modest decline from 2023, underscoring how contested grid access remains. For sustainable investors, this means the marginal economics of AI may increasingly depend on regulated utilities, transmission planners, interconnection reform, and equipment supply chains rather than on compute demand alone (FERC, 2025; IEA, 2025).
The local nature of the issue becomes even clearer outside the US. Ireland’s Central Statistics Office reported that data centers accounted for 22% of metered electricity consumption in 2024, up from just 5% in 2015. That is a remarkable statistic because it shows how quickly concentrated data-center load can transform a regional power system and change the politics of energy infrastructure. Similar localized pressure appears in EPRI’s state-level scenarios, where Virginia already exceeds a 20% data-center share of electricity use and could move materially higher by 2030. Sustainable investors who rely only on firm-level climate targets without looking at grid region, resource adequacy, and community energy effects are likely to miss where the real transition risks sit.
This is where conventional ESG shorthand starts to fail. It is tempting to frame AI’s energy demand as evidence that AI is “bad for climate.” The IEA’s analysis points to a more nuanced conclusion. In its base case, emissions from electricity use by data centers rise from about 180 million tonnes today to roughly 300 million tonnes by 2035, and to as much as 500 million tonnes in a higher-growth case. Yet even then, these emissions remain below 1.5% of total energy-sector emissions. At the same time, the IEA estimates that wider adoption of existing AI applications across energy, industry, transport, and buildings could generate emissions reductions equivalent to around 5% of energy-related emissions in 2035. AI, then, is neither climate savior nor climate villain. It is an accelerator whose net effect depends on the carbon intensity and flexibility of the systems around it.
That distinction has major implications for sustainable investing. The relevant issue is whether a company’s growth model is aligned with an increasingly constrained and decarbonizing electricity system. Two firms can both be “AI leaders” and still have very different sustainability profiles depending on siting decisions, power procurement strategies, cooling technologies, and their willingness to invest in additional carbon-free generation and grid-supporting assets. Sustainable investing in the AI era therefore requires a shift from category labels to system diagnostics.
The first diagnostic is the difference between annual procurement claims and actual hourly grid impact. This is where corporate climate accounting can be directionally useful but still insufficient. Microsoft’s FY2024 environmental data show location-based Scope 2 emissions of 9.96 million metric tons CO2e, while its market-based Scope 2 emissions were only 259,090 metric tons. That gap is not evidence of manipulation; it reflects legitimate contractual renewable procurement under greenhouse-gas accounting rules. But it does demonstrate why annual matching alone cannot answer the sustainable-investing question. A firm can sharply reduce market-based emissions while still operating in regions whose marginal hourly power remains fossil-intensive or infrastructure-constrained. This is exactly why the next frontier for investor analysis is not merely renewable energy purchasing, but temporal and locational quality of supply (Microsoft, 2025).
Google’s approach makes this distinction explicit. The company continues to pursue a 24/7 carbon-free energy goal by 2030, not just annual renewable matching. Its sustainability materials explain that annual purchases of wind and solar can offset electricity consumption on a global annual basis, but that this is an imperfect proxy for truly decarbonized operations because hourly supply and demand still diverge by region. In 2024, Google reported signing contracts for approximately 8 GW of clean-energy generation capacity, reducing its data-center energy emissions by 12% year over year even as data-center electricity consumption rose 27%, and reporting a fleet-wide average PUE of 1.09 versus an industry average of 1.56. That is a sophisticated case study in why sustainable investors should reward not only absolute procurement volume, but also efforts to improve hourly matching, efficiency, and grid impact (Google, 2025).
Amazon disclosed support for nuclear-energy development, including SMRs, as part of its carbon-free energy strategy. For investors, this underscores an important point: there is no single decarbonization pathway for AI infrastructure. The credible operators are increasingly assembling portfolios of renewables, storage, advanced cooling, power electronics, and firm carbon-free power rather than relying on one technology or one accounting mechanism. (Amazon, 2025).
Efficiency, however, should not be romanticized. It is necessary, but it is not sufficient. The recent literature stresses both the progress and the uncertainty. Masanet et al. (2024) argue that analysts still need much better data on AI workloads, hardware deployment, utilization rates, and system design before they can estimate AI electricity use with confidence. Xiao et al. (2025), in Nature Sustainability, show that the environmental burden of AI servers is highly sensitive to siting, grid mix, and operational best practices; even where best practices reduce impacts meaningfully, the industry is unlikely to meet near-term net-zero aspirations without relying substantially on offsetting and restoration mechanisms. In practical terms, this means sustainable investors should resist both techno-optimism and techno-pessimism. Better cooling, more efficient chips, and lower PUEs matter. But demand growth can still outrun efficiency gains.
That combination of demand growth and imperfect efficiency is why the real investment opportunity set sits one layer beneath the AI application stack. The most obvious beneficiaries are renewable developers, storage providers, grid-equipment manufacturers, and transmission enablers. The IEA expects roughly half of global growth in data-center electricity demand through 2035 to be met by renewables, with storage and the broader grid playing a critical supporting role, while natural gas and nuclear also expand. DOE has likewise highlighted clean-energy resources, transmission upgrades, and flexible demand strategies as essential to meeting data-center load without undermining affordability or reliability. Sustainable investors should therefore view AI-driven power demand not only as a risk to manage, but as a demand pull for the infrastructure of the energy transition.
Still, the opportunity is not indiscriminate. There is a crucial difference between financing assets that help reconcile AI load with deep decarbonization and financing assets that merely extend fossil dependence under the cover of “meeting urgent demand.” The IEA explicitly notes that renewables and natural gas lead near-term supply additions for data centers, while nuclear and geothermal are poised to contribute more over time. The review by Lal and You (2025) similarly warns that without deliberate planning, AI infrastructure expansion can reinforce fossil-fuel dependency, especially under disruptive growth scenarios. In sustainable-investing terms, this argues for a transition lens rather than a purity lens: investors should differentiate between flexible, bridge-like solutions embedded in a decarbonization pathway and capital that locks in long-lived, high-carbon dependence without credible transition logic.
This has immediate consequences for equity selection. In utilities, for example, the winners are not simply those with exposure to data-center load growth. The more attractive sustainable-investing profile belongs to utilities that can add load while preserving affordability, deploying low-carbon generation, expanding storage, modernizing networks, and structuring tariffs so that new large loads pay fairly for system upgrades. Microsoft’s recent public messaging that it will “pay its way” so that its data centers do not raise community electricity prices is a reminder that social license is becoming part of the business model, not a side issue. Investors should be skeptical of load growth stories built on regulatory arbitrage or implicit cost shifting to residential customers.
In real assets and infrastructure, the differentiation logic is even sharper. A data center with access to unconstrained transmission, low-carbon firm power, efficient cooling, and flexible backup assets is not equivalent to one dependent on congested grids, weak interconnection visibility, and carbon-intensive marginal supply. Likewise, a renewable portfolio tied to long-duration offtake in data-center regions is not equivalent to a generic merchant portfolio with weak additionality or poor deliverability. As the market matures, investors are likely to place a premium on power-secured sites, colocated storage, grid-enhancing technologies, flexible interconnection structures, and regional power-market sophistication. In that sense, sustainable value in AI may increasingly accrue to the physical and regulatory architecture beneath the data center, not to the shell alone.
The credit market will also need to adapt. Sustainability-linked instruments tied to generic emissions intensity metrics are becoming less convincing for AI-heavy issuers because they can obscure the real drivers of system impact. More decision-useful KPIs would include hourly or 24/7 carbon-free energy progress, location-based and market-based Scope 2 reporting side by side, PUE trends, water use in stressed basins, percentage of load backed by additional carbon-free capacity, and evidence of grid-supporting flexibility such as storage, demand shifting, or backup generation that can participate in system balancing. The ISSB’s IFRS S2 standard, effective from 2024 and further amended in 2025, gives investors a stronger base for climate-related disclosure, while 2025 guidance on transition-plan disclosures pushes companies toward more decision-useful articulation of strategy and implementation. For AI infrastructure, the next step is to push those disclosures from generic climate commitments toward power-system specifics.
A genuinely useful engagement agenda for sustainable investors should therefore ask a harder set of questions than standard ESG questionnaires usually do. First, where are the company’s incremental AI loads being added, and what is the hourly carbon intensity of those grids? Second, how much of the load is backed by additional, local, carbon-free generation versus portfolio-level annual matching? Third, what is management’s plan for transmission and interconnection risk? Fourth, are cooling and chip-efficiency gains keeping pace with power-density growth? Fifth, how is the company accounting for embodied carbon in buildings, ICT hardware, and supply chains? Sixth, what community-benefit, tariff, or cost-sharing mechanisms are in place to preserve affordability and social license? These are not abstract questions. They go directly to capex timing, margin durability, regulatory risk, and the credibility of transition claims.
The embodied-carbon issue deserves special emphasis because it is often overshadowed by operational electricity. Microsoft notes that datacenter ICT hardware supply chains are among the largest contributors to its Scope 3 emissions, while Amazon reports broad efforts to reduce embodied carbon in building materials and construction across its facilities, including data centers. Once AI infrastructure scales rapidly, the sustainability profile depends not just on electricity consumed during operation, but on steel, concrete, semiconductors, cooling equipment, and networking gear embedded upfront. A sustainable investor who looks only at operational renewable matching may therefore understate lifecycle emissions and overstate climate alignment. As AI capex accelerates, hardware and construction supply chains become part of the investment thesis.
Water is another issue that should be integrated, though not conflated, with energy analysis. Google openly notes that water cooling can reduce electricity use and associated carbon emissions relative to air-based cooling in some geographies, but that this can increase water footprints and create tradeoffs. Xiao et al. (2025) show that AI server deployment creates compound energy-water-climate effects that vary materially by location. For sustainable investors, this means the best AI infrastructure is not just low-carbon in aggregate; it is appropriately designed for local environmental conditions, including water stress.
There is also a broader strategic point. Sustainable investing in the AI era should not default to a defensive posture. AI is already being used to improve transmission utilization, fault detection, renewable forecasting, industrial efficiency, and building energy management. The IEA estimates that AI-enabled tools could unlock up to 175 GW of transmission capacity without building new lines and could produce material efficiency gains across industry and buildings. If those applications scale, they can improve the economics of decarbonization itself. The investment implication is that “AI exposure” should be decomposed into at least two buckets: AI as a source of electricity demand, and AI as an enabler of lower-carbon energy systems. The highest-conviction sustainable strategies will often seek both sides of that equation simultaneously.
The central mistake investors can make is to treat AI’s energy appetite as a temporary side effect that efficiency will eventually solve. The better view is that electricity is becoming a strategic bottleneck and competitive differentiator for AI. Companies that secure clean, reliable, flexible power and disclose credibly on how they do so will likely earn a lower risk premium over time. Companies that rely on opaque procurement claims, congested grids, weak transition logic, or socially contentious load growth may face higher capital costs, permitting friction, and reputational discounting. In this sense, sustainable investing is not peripheral to AI valuation. It is becoming part of AI fundamentals.
The most productive conclusion, then, is neither anti-AI nor naïvely pro-AI. The energy cost of AI is real, material, and rising. But it is not a reason for sustainable investors to retreat from the sector wholesale. It is a reason to invest with a far more granular understanding of power systems, infrastructure bottlenecks, emissions accounting, and transition quality. The relevant opportunity is to fund the companies, assets, and networks that make AI compatible with a lower-carbon, more resilient electricity system—and to withhold capital from models that externalize grid stress or hide behind weak climate proxies. The sustainable-investing question is not whether AI will consume more energy. It will. The real question is whether investors will help shape the kind of energy system that AI grows into.
References
Amazon. (2025). 2024 Amazon Sustainability Report. Amazon.
Electric Power Research Institute. (2026, March 27). Powering intelligence: Updated U.S. data center scenarios. EPRI.
Federal Energy Regulatory Commission. (2024). Explainer on the interconnection final rule.
Federal Energy Regulatory Commission. (2025, March 20). FERC State of the Markets Report for 2024 released.
Google. (2025). 2025 Environmental Report. Google.
Google. (2025). Frequently asked questions: Google data centers. Google Data Centers.
Google. (2025). Sustainable & efficient operations. Google Sustainability.
International Energy Agency. (2025). Energy and AI. IEA.
International Financial Reporting Standards Foundation. (2025, June 23). Guidance on disclosures about transition plans under IFRS S2.
International Financial Reporting Standards Foundation. (2025, December 11). ISSB issues targeted amendments to IFRS S2 to support application.
Lal, A., & You, F. (2025). Advances and challenges in energy and climate alignment of AI infrastructure expansion. Advances in Applied Energy, 20, 100243. https://doi.org/10.1016/j.adapen.2025.100243
Masanet, E., Lei, N., & Koomey, J. (2024). To better understand AI’s growing energy use, analysts need a data revolution. Joule, 8(9), 2427-2436. https://doi.org/10.1016/j.joule.2024.07.018
Microsoft Corporation. (2025). 2025 Environmental Data Fact Sheet. Microsoft.
Microsoft. (2026, February 18). A milestone achievement in our journey to carbon negative.
Shehabi, A., Smith, S. J., Hubbard, A., Newkirk, A., Lei, N., Siddik, M. A. B., Holecek, B., Koomey, J. G., Masanet, E. R., & Sartor, D. (2024). 2024 United States data center energy usage report. Lawrence Berkeley National Laboratory. https://doi.org/10.71468/P1WC7Q
Xiao, T., Nerini, F. F., Matthews, H. D., et al. (2025). Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA. Nature Sustainability, 8, 1541-1553. https://doi.org/10.1038/s41893-025-01681-y
Central Statistics Office. (2025, June 10). Data centres metered electricity consumption 2024. Ireland CSO.