AI data centres can finance new solar farms while also making gas plants run more often. The outcome depends on timing, location, storage, grid queues and how clean energy is counted.
By Deep Field · Edited by Lachlan Brown
Published September 4, 2026 · How we edit
America’s AI boom is creating an energy story that resists a simple label. The same data centres that are increasing electricity demand and keeping fossil-fuel generators busier are also helping finance some of the country’s largest solar, wind and storage projects.
The International Energy Agency’s central outlook says data centres could account for nearly half of US electricity-demand growth through 2030. Globally, their electricity use was about 415 terawatt-hours in 2024 and could more than double to roughly 945 terawatt-hours by the end of the decade. AI is the largest driver of that increase, but the totals also include conventional cloud services, storage and other computing.
That scale helps explain why apparently contradictory claims can both be true. New demand can accelerate renewable construction because solar is quick to deploy and technology companies have clean-energy commitments. Yet a server campus needs power at night, during cloudy weather and whenever a grid is constrained. Gas plants—and in some regions coal—often meet the marginal load that variable renewables cannot cover at a particular hour.
The balance is not predetermined. It depends on where the facilities are built, how quickly transmission and generation can connect, whether clean-energy purchases add genuinely new supply and how flexibly computing demand can operate. The AI boom is pushing two parts of the US power system at once.
For much of the past two decades, US electricity demand was comparatively flat. Efficiency improvements offset growth in population, appliances and digital services. Utilities could plan around gradual change. Large AI campuses break that rhythm: a single proposed cluster can ask for hundreds of megawatts, and groups of campuses can rival the demand of a city.
Data centres are not new, and neither is machine learning. What changed is the scale of training and operating generative models, together with a rush by competing companies to secure computing capacity before they know exactly how much will be economically useful. More capable chips reduce the energy required for a given calculation, but they also make far more calculations affordable.
The IEA’s 945-terawatt-hour global estimate would put data-centre electricity consumption near Japan’s present total, a comparison SpaceDaily examined previously. In the United States, the agency expects data centres to account for nearly half of all electricity-demand growth between 2024 and 2030.
That does not mean half of US electricity will go to data centres. It means nearly half of the increase from today’s much larger base could come from them. The distinction is easy to lose in a headline, but it matters: the sector is becoming a dominant source of new load without becoming the dominant user of the whole grid.
No one knows exactly how many announced campuses will be built, how intensively their servers will run or how quickly chips and cooling systems will improve. Projects change size, move between counties or disappear. Forecasts based on utility connection requests can count speculative capacity several times; forecasts based on hardware shipments can miss facilities that are still being designed.
A Lawrence Berkeley National Laboratory report estimated that US data centres used 176 terawatt-hours in 2023, about 4.4 per cent of national electricity. Its forward range for 2028 was a remarkably broad 325 to 580 terawatt-hours, or 6.7 to 12 per cent of US consumption.
That range is not a flaw to be quietly averaged away. It is a description of real uncertainty. AI efficiency could improve rapidly; demand for AI services could rise even faster. Power constraints might delay construction, while behind-the-meter generation could allow some projects to bypass the queue. The IEA figure is a central scenario, not a meter reading from 2030.
The robust conclusion is more modest: expected growth is large enough to shape generation investment, transmission plans, electricity prices and emissions. Whether the top or bottom of the forecast range is closer will determine how severe those effects become.
Solar has several attractions for a technology company trying to secure power quickly. A project can often be built faster than a new nuclear station or a large thermal plant. Its fuel cost is zero and its operating costs are relatively predictable. A long-term power-purchase agreement with a creditworthy buyer can give a developer the revenue certainty needed to finance construction.
The company does not usually run a private cable from a distant solar farm to its servers. The project supplies a regional grid, while the data centre draws from that grid. Contracts and certificates establish the financial relationship and allow the buyer to claim the corresponding clean generation, subject to the accounting rules used.
The IEA expects renewables to remain the fastest-growing source of electricity for data centres. In its base case, renewable generation serving the sector rises at an average 22 per cent a year from 2024 to 2030 and meets nearly half of the growth in data-centre electricity demand.
Corporate purchases have already become large enough to influence the market. The Corporate Energy Buyers Association’s tracker, cited by Rhodium Group, recorded about 130 gigawatts of clean electricity procured by US corporate buyers across all sectors from 2014 through 2025. Not all of that belongs to data centres, but hyperscale computing companies are among the major buyers.
There is therefore a defensible sense in which AI demand accelerates solar deployment. A large customer signs a long contract; a developer uses it to finance a plant that might not otherwise be built as soon. The complication begins when an annual renewable purchase is mistaken for continuous physical supply.
A company can buy as much renewable electricity as its facilities consume over a year and still rely on fossil generation during many individual hours. A solar farm may produce more than the data centre needs at noon and nothing after sunset. The accounting balances annually; the grid has to balance every second.
This is why claims that a campus is “powered by 100 per cent renewable energy” require context. They commonly describe matched purchases over a reporting period, not a sealed electrical system in which every electron comes from a clean generator at the moment it is consumed. Grid electricity mixes together, while generators respond to demand according to location, price and availability.
Batteries can shift midday solar into the evening. Wind may complement solar at some sites. Wider transmission can connect regions with different weather. Hydroelectric, geothermal and nuclear plants can provide low-carbon power without depending on daylight. Flexible computing jobs can also move towards hours when clean supply is abundant.
Each option has limits. As a previous SpaceDaily article on data-centre batteries noted, storage is becoming part of the technology sector’s energy strategy. But a battery designed for four hours cannot cover several calm, cloudy days, and building enough long-duration storage remains expensive and site-dependent.
Natural-gas turbines can produce electricity when requested, and an existing plant can often raise its output sooner than a new transmission line or generator can be completed. Gas plants also fit a utility’s obligation to serve a customer continuously, including during peak evenings, heat waves and periods of weak renewable production.
A 2026 Rhodium Group analysis explored what happens if data-centre demand grows 30 per cent faster through 2030 than in its baseline. Supply-chain limits, gas-turbine backlogs, permitting delays and crowded interconnection queues meant that very little additional generating capacity could arrive before 2030 in response to the higher load.
In those scenarios, existing gas and coal plants met 55 to 85 per cent of the extra demand above the baseline data-centre forecast through 2030. The number does not describe all data-centre electricity, and it does not say fossil fuels supply 85 per cent of the sector. It isolates the response to demand growing faster than Rhodium had already assumed.
That distinction explains how solar additions and fossil generation can rise together. New renewable plants may cover a large part of annual growth, while gas plants run more often during the hours or in the regions where that new clean supply is unavailable. A data centre can help cause both investments without either description being false.
Another apparent contradiction comes from the baseline used. “Data centres increase emissions” can mean the power sector emits more than it would have without the additional load. It does not necessarily mean national power emissions are higher than they were a decade earlier.
Rhodium’s baseline, which already included strong data-centre growth, projected US power-sector emissions in 2030 at 17 to 21 per cent below 2024 levels. In the higher-demand cases, emissions in 2035 were 6 to 13 per cent higher than in the corresponding baseline. Yet they still landed 6 to 22 per cent below 2024 because the overall grid became less carbon-intensive.
Both comparisons matter. Falling total emissions would represent progress. Emissions remaining above an otherwise plausible lower path would represent an opportunity cost. The additional computing load can slow decarbonisation even if it does not reverse it nationally.
Corporate renewable procurement can shrink that gap, but its effect depends on additionality and timing. A contract that brings a new solar farm onto the same constrained grid is more consequential than buying certificates from a project that would have operated anyway. Hourly matching is more demanding—and more informative—than balancing consumption and generation once a year.
Rhodium’s modelling becomes especially revealing after 2030, when the grid has had more time to respond. Under favourable conditions for clean energy—falling technology costs and relatively expensive fossil fuel—higher data-centre demand more than doubled new solar additions relative to the baseline. Solar and wind supplied the largest share of added capacity.
Under the high-emissions assumptions, cheap gas and more expensive clean technologies produced the opposite result. Gas dominated new capacity through 2035, and emissions increased further above the baseline. The same data-centre demand acted as an accelerator; market prices, policy and construction constraints decided what it accelerated.
This is why confident claims that AI will either rescue renewables or lock America into gas go beyond the evidence. Demand creates a market for electricity. It does not by itself choose the generators. Tax policy, fuel prices, interconnection rules, local permitting, equipment supply and corporate contracts do that.
The scarce resource is often time. A planned generator may wait years for network studies, transformers, permits and transmission upgrades. A large customer can announce a 2028 opening date without the clean projects intended to serve it having a guaranteed connection by then.
In parts of Britain and North America, connection timelines have become long enough to inspire proposals for orbital computing, a bottleneck SpaceDaily examined in earlier coverage. The space concepts remain speculative, but their sales pitch identifies a terrestrial problem: a project developer can sometimes obtain servers faster than a utility can obtain power infrastructure.
If a data centre connects before its associated solar, storage and transmission, existing plants fill the gap. If all of them arrive together, the emissions result changes. A clean-energy contract scheduled for a later year does not remove carbon released while the facility waits.
Utilities also have to plan for reliability under uncertainty. They cannot assume every solar proposal will be completed, every battery will remain charged or every computing job will be flexible. That encourages firm capacity, which today often means natural gas even when the utility also commissions substantial renewable generation.
More transmission and faster interconnection would allow completed renewable projects to reach customers sooner. Batteries and other flexible resources can cover short gaps. Firm low-carbon technologies—including geothermal, hydroelectric and nuclear power—can address longer ones where geography, cost and construction schedules permit.
Data centres themselves can help. Training runs, data processing and some maintenance jobs can be shifted towards cleaner hours or regions. Latency-sensitive services cannot simply pause when a cloud passes, but not every calculation is equally urgent. Demand-response contracts can reward operators for reducing non-essential load when a grid is tight.
Efficiency remains important even if it cannot settle the total on its own. Better chips, cooling and server utilisation reduce electricity per unit of computation. The rebound effect works in the other direction: cheaper computation encourages more models, queries and applications. Efficiency slows demand growth only when its gains outrun that expansion.
Finally, rate design can require large new customers to pay for the generators, substations and network upgrades built for them, protecting households from some costs. Clear hourly emissions reporting would also make it harder to hide the physical grid behind annual renewable claims.
The AI buildout is not clean or fossil by definition. It is an unusually large new electricity customer arriving faster than much of the infrastructure needed to serve it. That customer can underwrite solar farms and batteries while also extending the hours that gas plants run.
The honest measure is not the number of renewable contracts or gas turbines considered separately. It is what new generation actually gets built, when it connects, where it operates and which plants respond during every hour of demand.
Under favourable conditions, AI could become a powerful customer for a larger clean grid. Under constrained or gas-friendly conditions, it can slow the decline in emissions even as solar capacity rises. For the next several years, both processes are likely to be visible at once.
About this article
This article is for general information and reflection. It is not professional advice. For your specific situation, consult a qualified professional.
Twelve figures that put your place in the cosmos into staggering perspective — with real imagery from NASA, Hubble & Webb. Pop in your email and we’ll send you the free PDF.
Standards
Space Daily articles are edited and fact-checked before publication. We use AI tools in the newsroom. See our editorial standards and masthead.
About this column
Deep Field covers astronomy and cosmology. Telescopes and what they find, planets near and far, and the strange physics of the deep universe, traced back to the journal papers and observatory data behind the headlines.
Space, science, and the human mind. Since 1995.
A Brown Brothers Media Pte. Ltd. publication. All rights reserved © 2026.
Highlights of the week in space, science, and the human mind — delivered to your inbox.
Also available via RSS.

source