Goldman Sachs AI Power Forecast Is Not 20% of Global Energy

Goldman Sachs Research’s December 2025 outlook forecasts data-center power consumption rising 175% from 2023 to 2030; it does not state that AI will consume 20% of global energy. The later forecast strengthens the case for rapid demand growth but does not support the much larger global-share claim.

The global evidence available since the original article points in the same direction: AI is accelerating electricity use in data centers, yet the sector remains a small share of worldwide electricity in the central scenario. The important change is therefore not an approach toward one-fifth of global energy, but a sharper regional challenge for grids where new facilities are concentrated.

What the Goldman Sachs forecast measures

The bank’s projection covers the power consumed by data centers as a whole, not AI computation in isolation. These facilities also run conventional servers, cloud workloads, storage, networking and cooling equipment, so their total electricity demand cannot automatically be attributed to AI.

The percentage is also a growth rate from a baseline, rather than a share of the world’s power supply. A rise of 175% means the modeled load reaches 2.75 times its earlier level. It does not mean data centers consume 175% of available electricity, and it provides no basis for assigning AI a one-fifth share of worldwide energy.

A separate figure in the bank’s outlook says AI’s portion of the overall data-center market could double to 30% over the following two years. That metric is market share inside the data-center industry; the page does not define it as a share of electricity consumption. Treating it as an energy percentage would mix two different measures.

The word energy creates another denominator problem. Electricity is only part of global energy use, which also includes fuels consumed directly in transport, industrial processes, heating and other activities. A claim about total global energy is consequently broader than a forecast about electricity used in one category of buildings.

The global benchmark remains far below 20%

The International Energy Agency’s global base case puts data-center consumption at about 945 terawatt-hours in 2030, just under 3% of worldwide electricity, with accelerated servers associated mainly with AI supplying almost half of the sector’s net demand increase.

That is a substantial expansion for infrastructure that must operate continuously and obtain dependable grid connections. It is nevertheless a forecast for all data centers, meaning AI alone would occupy only part of that global share. Cooling, networking, storage, conventional computing and power-support systems account for the rest.

The comparison resolves the apparent conflict between fast growth and a modest global percentage. Data centers begin from a relatively small base, so their consumption can more than double without approaching one-fifth of global electricity—much less one-fifth of total global energy.

The agency’s model also treats future demand as uncertain rather than fixed. Adoption of AI services, shipments of specialized hardware, server utilization, gains in hardware and software efficiency, cooling performance and delays in obtaining power can all change the eventual result. The central scenario is therefore a planning benchmark, not a measured outcome waiting to occur.

Why grids can still face severe pressure

A global average hides the geographic concentration of the load. Data centers tend to cluster where operators can secure fiber connectivity, land, customers and electrical infrastructure. Several large facilities seeking connections in the same area can strain local generation or transmission even when their worldwide share remains comparatively small.

The Lawrence Berkeley National Laboratory’s June 2026 update estimates 649 TWh for US data centers in its 2030 reference case, equal to 11.8% of national electricity use, while its compounded uncertainty scenarios span 521–843 TWh and shares of 9.5%–15.3%.

Those figures apply to the United States and to the entire data-center sector, not to AI alone or to the world. They nevertheless show why utilities and regulators may regard the expansion as a major planning issue: a geographically concentrated industry can reach a double-digit national share while remaining far below that level globally.

Timing compounds the problem. Computing facilities can be developed faster than major transmission lines and power plants, whose planning, permitting and construction often extend over longer periods. The resulting bottleneck is primarily about where and when adequate capacity becomes available, not evidence that AI is approaching 20% of global energy use.

The forecast’s real significance

The unsupported global-share claim should not obscure the scale of the underlying transition. AI is increasing demand for accelerated computing, encouraging construction of power-dense facilities and forcing grid operators to account for unusually large new loads. The consequences will be most visible in specific markets rather than as a uniform worldwide shortage.

Forecasts may move again as operators disclose more projects and chip efficiency changes. What remains supported is narrower but still consequential: Goldman Sachs expects steep growth in total data-center power consumption, the global base case remains around a few percent of electricity, and national exposure can be much higher where capacity is concentrated. None of the three findings establishes that AI will consume 20% of global energy by 2030.