Goldman Sachs warns: AI's energy demand is equivalent to "recreating Japan," with turbines, transformers, and political resistance being the biggest bottlenecks.

Goldman Sachs warns: AI's energy demand is equivalent to "recreating Japan," with turbines, transformers, and political resistance being the biggest bottlenecks.

The power demand for AI-driven data centers is expanding at a faster-than-expected pace, and the core contradiction on the supply side has evolved from "whether there is enough power" to "whether it can be built in time".

A recent Goldman Sachs study warns that from early 2024 to the end of 2030, the additional electricity consumption from AI will be equivalent to the entire electricity consumption of Japan, the world's fifth-largest electricity consumer. Shortages of turbines, transformers, transmission lines, and skilled workers, along with local political obstacles, are becoming the real bottlenecks restricting the implementation of supply.

Goldman Sachs has raised its forecast for U.S. electricity demand growth from a CAGR of 3.2% to 3.5%, and significantly increased its 2030 forecast for U.S. data center electricity demand from 83 gigawatts to 108 gigawatts. Global data center electricity demand growth from 2025 is projected to jump from 117% to 170%. Simultaneously, capital expenditure forecasts for hyperscale cloud service providers have also been revised upwards: from $1.2 trillion in 2027 to $1.7 trillion, and from $1.5 trillion to $2.1 trillion in 2029.

Goldman Sachs emphasizes that its baseline scenario does not predict large-scale blackouts across the United States, but regional power shortages, rising electricity prices, and public opposition will profoundly impact the actual construction progress. The PJM region is currently on the verge of collapse, and ERCOT is expected to face pressure around 2028 as load increases. Whether or not public and political support can be secured will largely determine whether the predicted power capacity can actually be built.

Demand continues to be revised upwards: Efficiency improvements are offset by larger-scale applications.

Goldman Sachs analyst Brian Singer points out that while improvements in chip and model efficiency are real, they haven't led to a contraction in overall computing spending. Lower unit costs of computing power have expanded the range of economically viable applications, and the increased token generation, broader agent workloads, and the continued expansion of budgets by hyperscale cloud service providers have collectively offset and outweighed the efficiency gains.

The vacancy rate in major U.S. data center markets has plummeted from 2% to 7% to 1% to 2%, and Goldman Sachs expects it to rise to around 3% by 2030. This tightening of supply and demand further supports the upward revision of demand forecasts.

Singer quantifies this trend with an intuitive frame of reference: "If we start counting from the beginning of 2024, the seven years until the end of this decade will see AI's additional electricity consumption equivalent to the entirety of Japan—the world's fifth-largest electricity consumer. This scale is extremely significant."

Supply path: Natural gas first, followed by renewable energy, and nuclear power last.

Goldman Sachs outlines a phased path for data center power supply: in the near term, simple cycle gas turbines and renewable energy plus energy storage will be the mainstays; in the medium term, a transition to combined cycle gas turbines will be implemented; and in the long term, nuclear power will be relied upon, with the restart of some existing nuclear power plants as a transitional solution. It is projected that by 2030, natural gas will account for approximately 60% of data center electricity consumption, while renewable energy (including energy storage) will account for approximately 40%.

With grid connection queues lasting 2 to 7 years, behind-the-meter (BTM) gas-fired power generation is becoming an important means of "jumping the queue." Goldman Sachs has raised its forecast for BTM gas-fired capacity to approximately 30 gigawatts by 2030 (with actual power generation exceeding 20 gigawatts), representing about 20% of total data center demand at that time. Goldman Sachs positions BTM as a transitional solution rather than a final state, believing that large customers will still prefer grid connection in the long term for lower costs and higher reliability.

In terms of regional distribution, the rise of MISOs (Midwest Independent System Operators) is structurally significant—regulated utility companies can provide a one-stop service for generation, transmission, distribution, and regulatory relationships, and electricity prices are determined by state-level regulators rather than wholesale markets, which helps to manage the political sensitivity of rising electricity prices.

Four "T"s and Seven "P"s: The bottleneck has shifted from megawatts to supply chains and politics.

Goldman Sachs analyst Allison Nathan summarizes the current core contradictions as four "Ts": turbines, transformers, transmission, and tradespeople. A major turbine manufacturer has stated that it expects to sell half of its 2031 capacity by the end of this year, and equipment delivery cycles are profoundly impacting the choice of power sources. Meanwhile, after a decade of stagnant power demand in the United States, the shortage of electricians and high-voltage welders is becoming increasingly prominent, and the structural contradiction of long training cycles is unlikely to be resolved in the short term.

Singer further proposed a seven-P framework, covering AI pervasiveness, chip and model productivity, electricity price, policy, parts, people, and physical environment. Among these, physical environment risks are particularly noteworthy: more than half of the data centers are located in high-physical-risk areas, facing challenges such as high temperatures, high humidity, and drought. The water-electricity trade-offs posed by cooling systems are most pronounced in West Texas—water resources are more difficult and expensive to obtain than new electricity.

Political resistance may become the ultimate constraint: More than 300 local bans have already taken shape.

Goldman Sachs analyst George Lee characterized interconnectivity as the "single most important issue" in the U.S. utility sector, noting that there are already over 300 local and regional suspension orders, while the number of state-level bans is relatively limited. He warned that if AI and the power industry fail to develop a coordinated public communication strategy, political factors will ultimately become the hard constraint.

Community opposition primarily stems from five concerns: the risk of power outages, rising electricity costs, water consumption, noise pollution, and waste heat emissions. Goldman Sachs' assessment is that projects will relocate to more regulatory-friendly jurisdictions, potentially further exacerbating geographical concentration. Meanwhile, competition among towns vying for tax bases, construction jobs, and infrastructure investment is also creating space for some projects to take root.

At the balance sheet level of utilities, Goldman Sachs expects capital expenditures of regulated utilities covered by the company to increase by about 60% over the next five years compared to the previous five years, with 30% to 50% of incremental financing expected to be completed through equity financing. The leverage ratio still has a buffer of about 100 basis points before the rating downgrade threshold.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.