Power risks in AI infrastructure investment

For companies and local governments considering investment in AI infrastructure, the immediate question is not only “how quickly electricity can be supplied.” It is also “under what conditions additional electricity demand can be justified,” including decarbonized power sources, grid reinforcement. Efficiency.

This framing follows the direction presented by Kim Sung-hwan, Minister of Climate, Energy and Environment: increased power demand in the AI era should be addressed through a decarbonized power system, not by expanding coal and gas. If this direction becomes concrete policy, investment in data centers and AI semiconductors will depend not only on demand for AI services, but also on site selection, power contracts, cooling efficiency, and grid connection.

Minister Kim said that if electricity consumption from AI semiconductors, data centers, and robots overlaps with electrification in industry, transportation, and buildings, power demand could rise to nearly three times its current level. He also stated that this increase should not lead to greater operation of coal-fired power plants or a delayed response to the climate crisis. The practical constraint is: AI may expand, but the electricity used to support it should not become more carbon-intensive.

For AI companies, that constraint may become a permitting and schedule risk, not just a cost issue.

Demand forecasts have already begun to treat ‘additional AI demand’ separately

The electricity demand forecasting method in the 11th Basic Plan for Long-Term Electricity Supply and Demand has two broad layers. It starts with model-based demand that reflects macro variables such as economic growth and temperature, then adds additional demand from data centers and advanced industries.

AI data centers are reflected separately, based on large-scale data processing demand associated with the spread of AI. The AI semiconductor industry reflects investment demand from advanced industries such as semiconductor clusters. To avoid overlap with existing trends, only the incremental portion is calculated as additional demand.

This has policy significance. It means electricity demand from data centers and semiconductors is not treated only as part of general industrial demand. Once it appears as a separate item in power planning, the question later becomes not only “how much electricity is needed,” but also “where, when, and with what power sources it will be supplied.”

However, publicly available materials do not make all key assumptions verifiable. It is not confirmed from those materials how the forecasts reflected electricity coefficients by AI semiconductor chip type or production process, GPU utilization rates at individual AI data centers, cooling efficiency, or site-specific grid constraints. Companies should therefore not treat government demand forecasts as confirmation that power procurement is feasible for a specific project. A nationwide aggregate forecast and connection feasibility for an individual site are separate issues.

The common direction of overseas policy: moving toward ‘conditional acceptance’ rather than just electricity volume

Materials from the U.S. Department of Energy discuss rising AI and data center demand together with clean energy deployment, transmission expansion, data center efficiency improvements, and demand flexibility. This is not simply an approach of building more power plants. It manages demand by considering power sources, grids, efficiency, and demand response together.

The EU is moving in a more regulatory direction. It requires monitoring and reporting of environmental indicators such as data center energy performance and water footprint, and is pursuing a rating system and minimum performance standards based on reported data.

Singapore evaluates power and water efficiency indicators when allocating new capacity, while combining efficiency standards, subsidies, and the use of low-carbon energy. Ireland, on the premise of grid constraints and decarbonization needs, has a principle of favoring data center development that demonstrates the additionality of renewable energy use.

Taken together, these materials suggest that data centers are becoming harder to treat merely as real estate development or IT facility expansion. Policy is moving toward treating them as industrial facilities that occupy regional power grids and carbon budgets. The claim that “they use a lot of electricity but create economic benefits” is unlikely to be sufficient on its own. Projects may also need to explain whether their electricity use avoids crowding out existing demand and carbon reduction targets.

Investment decision rules: examine additionality and flexibility before power contracts

The following rules can be applied by practitioners reviewing sites related to AI data centers or AI semiconductors.

First, a power procurement plan should not be assessed only by capacity figures. Its linkage with decarbonized power sources should also be evaluated. If the policy direction is a decarbonized response rather than expansion of coal and gas, sites that rely on carbon-intensive electricity over the long term may face greater policy risk.

Second, grid connection feasibility should not be confused with nationwide electricity demand forecasts. The fact that the Basic Plan for Long-Term Electricity Supply and Demand reflects additional demand from data centers and advanced industries means that the state recognizes the increase in demand. It does not help ensure that electricity can be supplied to every candidate site on time.

Third, efficiency and demand flexibility should be treated as core conditions, not supplementary materials for an investment proposal. The U.S. DOE presents efficiency improvements and demand flexibility together, while the EU and Singapore manage performance and water use indicators. In that context, future data center competitiveness is unlikely to be judged by GPU procurement capability alone. The ability to explain and adjust electricity use may also become important.

Fourth, claims about renewable energy use should be supported by additionality. As reflected in Ireland’s principle, a project that is nominally allocated existing renewable energy supply may be evaluated differently from one that contributes to expanding new clean power sources.

For AI infrastructure investment proposals, a more realistic approval question is therefore not “whether the electricity consumption can be handled.” It is “whether that electricity consumption can be explained through grid reinforcement, procurement of decarbonized power sources, efficiency management. Demand flexibility.” Projects that do not incorporate these elements into the initial design may face schedule and cost volatility at the bottleneck of power policy, even if demand for the model exists.

Aionda

Power risks in AI infrastructure investment

Demand forecasts have already begun to treat ‘additional AI demand’ separately

The common direction of overseas policy: moving toward ‘conditional acceptance’ rather than just electricity volume

Investment decision rules: examine additionality and flexibility before power contracts

Further Reading

References

Get updates