Energy for AI

Kim Sung-woo

Korea is in the midst of a stock market frenzy. Stock prices are influenced by a company's future value, and Korea's market is heavily weighted toward semiconductor firms whose share prices are shaped by global artificial intelligence (AI) investment. This is why investors in the Korean market are fixated on the future of AI. Their focus is on how much AI will reshape past industries and cultures, and whether it can generate returns commensurate with expectations.

Recently, a new element has entered this mix: energy. Concerns have emerged that the energy essential for AI training and inference may not be supplied as smoothly as anticipated, potentially delaying AI adoption in certain regions.

According to data released last August by global consulting firm McKinsey, worldwide data center capacity is projected to reach a cumulative 220 gigawatts by 2030, six times the 2020 level. This growth is driven primarily by the expansion of AI data centers. Relatedly, earlier this month, research firm Gartner projected this year's data center power consumption to rise 26 percent from last year to 565 terawatt-hours, and estimated next year's figure at 702 terawatt-hours, illustrating the explosive growth in power demand.

Compared with Korea's 2025 power consumption of 625 terawatt-hours, the pace of this increase becomes tangible. By 2030, data center power consumption is estimated to exceed 1,200 terawatt-hours, surpassing not only Japan but even Russia based on 2025 figures. The energy industry inherently requires long lead times to build large-scale supply, while AI demands massive power quickly. This mismatch has made securing energy one of the top priorities for AI development companies.

This phenomenon is already materializing in the market. Last March, U.S. tech outlet TechCrunch warned that up to half of currently announced data center construction could be delayed due to power supply issues, and observed that Big Tech companies are building their own power plants or signing separate power purchase agreements to overcome such crises. In addition, AI analytics firm SynMax reported that, as of last April, 40 percent of data centers targeted for completion in the United States this year risk being delayed, while 60 percent of those slated for completion by next year have yet to break ground.

Data centers have various options for securing electricity, but no readily available alternative is truly optimal. Drawing from the grid is difficult. So many new users have joined that grid connection waiting times have doubled, and key expansion equipment such as transformers is in short supply. Turning to relatively fast and stable gas-fired generation is also problematic because generator orders are backlogged, requiring waits of more than five years. Small modular reactors, which supply carbon-free power, are widely viewed as unlikely to see meaningful deployment before at least 2030, given the need to prove their commercial viability. Meanwhile, fuel cells, which have recently emerged as an alternative, have their own drawbacks: For now they must run on fossil fuels, and both system costs and clean hydrogen prices remain burdensome. Finally, the most environmentally friendly options — solar and wind — are intermittent, generating power only when the sun shines or the wind blows, which does not align with AI's need for round-the-clock supply. Even with energy storage systems, achieving a full 24-hour supply is realistically difficult, so support from other energy sources is inevitably required.

An additional and important point when choosing among these options is the burden associated with fossil fuel use. AI development companies have already established and disclosed carbon-neutral or carbon-reduction targets. Because these are voluntary, softening them is possible, but persuading key stakeholders — such as future-generation consumers and long-term investors — of the reasons for change is far from easy. Their past carbon emissions are minimal compared with other industries, and they would want to avoid being unfairly cast as principal culprits of today's climate crisis — caused by past emissions — simply because their future emissions are expected to be large. Accordingly, if the energy essential to the AI business is supplied by fossil fuels on a large rather than small scale, this could become a significant burden on a company's sustainability.

Taken together, there is no single best alternative to satisfy the explosively growing large-scale power demand in the short term; the only path is to choose second-best alternatives and supplement their weaknesses. For example, one approach is to meet primary power demand through renewable energy and energy storage, while covering only auxiliary power from low-efficiency gas generation (such as single-cycle turbines) or the grid. Another is to meet primary demand with fuel cells, gradually blending clean hydrogen alongside natural gas. The U.S. Energy Information Administration's observation that 93 percent of planned new generation capacity in 2026 will be solar, wind and energy storage, together with the forecast of energy market analytics firm Rystad Energy that roughly 10 gigawatts of fuel cells will be installed at U.S. data centers over the next five years, suggests that this approach is already being adopted.

Clearly, energy has come to play a critical role in the future of AI, one of the dominant themes of the stock market. Supplying energy in a stable and quick manner to meet explosive short-term demand growth is important, but so too is the recognition that, for AI to remain sustainable over the long term, the energy powering it must itself be sustainable.

Kim Sung-woo, head of Environment & Energy Research Institute at Kim & Chang, is a board member of the Korea Institute of Energy Technology Evaluation and Planning.