US power markets grapple with surging AI demand

As US power markets brace for an unprecedented surge in electricity demand, suppliers, traders and regulators are grappling with how to transform today’s creaking grid infrastructure and market design into a system fit to deliver. 

It is an enormous challenge. The primary driver of this demand growth, data centre expansion, is projected to account for 44% of new demand growth by 2028, according to the Nicholas Institute for Energy, Environment and Sustainability. However, forecasts for data centre build-out vary wildly, adding huge unpredictability to the task of future planning. It is clear though, that billions of dollars of investment is required in new capacity, much of which will need to be in renewables and battery storage if demand is to be met relatively quickly. While the futures market is not currently incentivising such investment, energy market stakeholders, including technology firms, are working on a range of initiatives intended to ensure future consumption requirements can be met. As well as building their own generation, tech giants could increasingly move into electricity trading, something that could transform power markets in the years to come.   

In the immediate future, the biggest impact of data centre demand will be rising prices, say analysts and market participants.

“Data centres are effectively bidding up the value of the marginal megawatt,” says James Preedy, vice-president of energy management at REV Renewables, an owner/operator of clean energy systems across the US. “And all load is effectively paying for that.”

Data centres are effectively bidding up the value of the marginal megawatt

James Preedy, REV Renewables

The challenge for data centre operators, regulators and industrial and retail consumers alike is that, as yet, there is not enough new generation in the pipeline to meet projected demand. “It’s not too out-of-pocket to say there’s a big supply crunch there,” says Patrick Finn, senior analyst, North American power markets, at energy data and analytics provider Wood Mackenzie in Boston.

Over the past decade, US power consumption has grown by just 0.3% a year, according to Wood Mackenzie, which now expects 2–3% growth each year to 2030. Hundreds of billions of dollars are being invested in data centres around the US in anticipation of the massive computing capacity that will be needed to meet expected demand for artificial intelligence.

However, there is no consensus as to how much of that demand will materialise and, therefore, how much new generating capacity will be needed (see figure 1.) At the high end of estimates, 451 Research, an arm of Standard & Poor’s, predicts that demand from US data centres will nearly triple by 2030, to 134 gigawatts (GW) from 50.5GW at the start of 2025.

And projections are changing fast. For example, in December 2025, BloombergNEF forecast data centre demand rising from around 40GW in 2025 to 106GW by 2035 – an increase of 36% on its April 2025 projection, which is nonetheless “more conservative than many others”, it notes.

“We’ve just seen a wave of new announcements,” says Helen Kou, head of US power research and analyst at BloombergNEF in San Francisco. Data centres are also getting larger. According to BloombergNEF, while only 10% of operating data centres currently exceed 50 megawatts (MW), the majority of data centres in development are above 100MW, with a handful of 1GW-plus projects set to come online in the next few years.

Forecasts from independent system operators (ISOs) and regional transmission operators (RTOs) that manage US wholesale power markets – and who are responsible for ensuring there is sufficient power supply to meet demand – tend to be particularly high (see figure 2). For example, in its latest estimates, published last April, Ercot has projected load on its grid in Texas of up to 231GW by 2031, compared with a peak demand record of 85.5GW during the 2023 heatwave. Data centres are projected to account for 78GW of that load. But, when it comes to forecasting data centre power demand, not all analysts are so bullish. Energy analytics provider Enverus said, as it released its January Energy Outlook Report, that “load growth in Ercot and PJM will fall short of ISO estimates, forcing them to downgrade their inflated long-term forecasts”.

The reason is simple, says Enervus’s Calgary-based senior power analyst Carson Kearl: “Power demand realised in the market for AI data centres can’t exceed the chip-manufacturing capacity of the large players. Our demand forecast in the near term is constrained to chip production by Nvidia, AMD, Google and Amazon.”

Why forecasting the future is hard

Every element of the AI boom is wreathed in uncertainty. There is little agreement over how long it will take to identify and develop profitable AI use cases and some feel  investment could dry up before many of these new data centres are built.

In addition, many data centre operators have put in more interconnection requests to ISOs than they expect they will need, either because they are putting in speculative requests for early-stage projects that may not go forward, or because they are submitting requests for the same project in different utility areas, according to think-tank the World Resources Institute.

There is also a real risk of a political backlash to data centre build-out. The potential impacts of unconstrained development on power prices, grid reliability and even water availability could see states legislate to slow down or even prevent data centre construction. In the first 11 months of last year, more than 40 bills were introduced to address the impact of data centre development on rate-payers’ bills.

“It’s not going to get easier for data centre developers to get things done,” predicts one US-based battery systems developer, “because we’re already seeing the impacts of higher prices … I would not dismiss the possibility that these high load-growth scenarios never materialise” due to political pushback, he says.

Solar, wind and storage are all needed to meet the base demand growth scenario. Without a significant amount of deployment of these technologies, the US power grid will not be able to support new data centres

Wood Mackenzie

Power demand is on the rise

However, while the magnitude of increased demand is in question, the direction is not, necessitating massive investment in new capacity. The Federal government is weighing in. In January, the White House National Energy Dominance Council, alongside 13 governors, published a document to urge PJM to enable the construction of $15 billion of baseload power generation.

In the near and medium term, the capacity build-out is most likely to come from solar/batteries and off-grid gas, says William Demas, head of Macquarie Asset Management (MAM) Green Investments for the Americas. “In the long term, we expect that nuclear will be relevant, as well. However, we expect solar/battery energy storage systems will become even more cost-competitive in the near term and will continue to increase prominence and scale in the data centre and energy mix.”

Wood Mackenzie forecasts 90GW in demand growth by 2030 (of which 59GW is from data centres). A combination of new gas, increasing output from existing thermal generation, and postponing thermal retirements will only deliver 37GW.

However, a bottleneck in the supply of gas turbines – with some analysts reporting delivery times for new orders in the early 2030s – is constraining gas-fired power generation. New nuclear is on the table, but the timelines for developing new large-scale projects mean little is likely to come online before later next decade.

Significant investment in renewables and storage is, therefore, urgently needed. “Solar, wind and storage are all needed to meet the base demand growth scenario,” according to Wood Mackenzie. “Without a significant amount of deployment of these technologies, the US power grid will not be able to support new data centres and other sources of load growth.”

However, thus far, Finn says supply has been slow to react to demand. Mechanisms in de-regulated markets such as PJM are failing to incentivise new capacity additions. Prices in its capacity market – where generators bid to provide firm generating capacity – have “gone nuts”, he says, leading to regulators introducing a price cap, of $325 megawatts (MW) a day. Although the last two auctions closed at this cap, the revenues provided by capacity payments and from selling power forward are insufficient to incentivise new combined-cycle natural gas plants, Finn adds (see figure 3).

Inevitably, power prices will have to rise if the likely shortfall in capacity is going to be closed. One mystery is why futures prices have not risen in anticipation of higher demand. Preedy at REV Renewables believes that data centre operators are signing power purchase agreements for 2027 and beyond that are struck around $100/MWh  – while annual peak strip futures in PJM are currently trading at around $70/MWh. “The question is, why don’t you buy the power in the forward curve at $70?” he asks. Part of the reason, he believes, is that operators are paying a premium for ‘cleanness’ from renewables, and some firming of the power, so the supply meets the operator’s load. There’s also some value in “additionality”, he adds, whereby the buyer can show it is bringing some new capacity onto the grid.

Another reason, he suggests, is that the market is anticipating that demand-side response (DSR) could help close the gap. But he is sceptical. “Historically, we haven’t used demand response in PJM very much,” he says. This has meant providers of demand response capacity – often manufacturers or other large consumers with some flexibility in their demand profile – were getting paid without needing to curtail their electricity use.

However, this has now changed – and the experience of having to halt production to relieve pressure on the grid has led some providers of DSR to rethink its value, he says. “For now, data centres may need to provide their own behind-the-meter response during scarcity events, either through load reduction or by turning on behind-the-meter generation,” he adds.

Finding solutions

Faced with these challenges, system operators, utilities, data centre operators and regulators are working on a range of possible solutions.

For example, the PJM RTO launched a ‘Critical Issue Fast Path’ consultation last year to speed decision-making around new large loads. In January, it announced the results of this process, which include a voluntary ‘Bring Your Own New Generation’ process. This proposes providing expedited grid interconnection to new loads (such as data centres) that can demonstrate they are simultaneously enabling new generating capacity to offset the additional demand they will place on the grid. This echoes a proposal last year from the Data Center Coalition, which represents operators, utilities PPL and Exelon, and the governors of Maryland, New Jersey, Pennsylvania and Virginia.

Meanwhile, some of the hyperscalers are responding to concerns about pricing. In January, Microsoft announced an initiative it calls Community-First AI Infrastructure, that includes a commitment to “pay our way to ensure our data centres don’t increase … electricity prices”. It will do that by, among other things, asking utilities and public utility commissions to set the rates they charge Microsoft high enough to cover the costs of adding the necessary generation, as well as by working with utilities to plan for new generation and contracting in advance for its data centres’ projected power demand.

Another initiative is seeking to ensure data centres are not solely a source of stress to grids across the US but can also play a role in supporting them. The Electric Power Research Institute (EPRI) has launched its DCFlex – Data Centre Flexible Load – initiative to explore how data centres can support and stabilise electric grids. The initiative aims to demonstrate how AI data centres can “transcend their traditional roles as passive electricity consumers, transforming into dynamic assets that enhance grid reliability, resilience and affordability for all electricity consumers.”

Going it alone

But faced with the challenges of sourcing power directly from the grid – and drawn-out regulatory and permitting processes that utilities are subject to in building out new generating and transmission capacity – many data centre operators are instead turning to behind-the-meter solutions to their power demand problems. “The speed at which the hyperscalers want to be able to energise these facilities is vastly faster than the speed at which utilities and ISOs can support that infrastructure,” Preedy says. “The logical conclusion – which has been a data centre catch-all approach for the last several years – is that they are going to going to source power for themselves in whatever way they can.”

These options include installing gas turbines, such as the 422MW of capacity installed by xAI at its Colossus data centres in Memphis, Tennessee.

This approach is not without its challenges, notes Drew Boyce, head of the physical natural gas desk at technology and advisory firm Mobius Risk Group in Houston, Texas. “It’s not as simple as putting a genset close to a pipeline,” he says, noting that operators have to navigate a complex process of arranging transportation, connection and contracting – and then they have to manage price and delivery risk.

In addition, the long waits for gas turbines make them a limited option. “There is a multi-year backlog for the delivery of new gas turbines and, frankly, there is not a significant amount of new gas development that we see,” says Demas at MAM. “Since manufacturing of those turbines is not being scaled drastically, more demand can’t be met beyond what is already contracted to be delivered in the medium term.”

The speed at which the hyperscalers want to be able to energise these facilities is faster than the speed at which utilities and ISOs can support that infrastructure

James Preedy, REV Renewables

Some are turning to batteries to enable them to begin operation more quickly. Demas points to a project that Calibrant Energy, a developer, operator and owner of on-site power solutions backed by MAM, is working on with Aligned Data Centers. The project will deliver a 31MW/62MWh battery energy storage system at Aligned’s data centre campus in the Pacific Northwest. “The on-site system, planned to be operational in 2026, will enable the facility to come online and scale operations years earlier than would be possible with traditional utility upgrades,” Demas says.

Operators are also turning to novel solutions, such as reciprocating engines and even aeroderivative gas turbines – effectively commercial aircraft engines repurposed to generate electricity rather than lift.

The sector is also providing a boost to fuel cell manufacturers. In October, Bloom Energy announced a $5 billion AI infrastructure partnership with investment management giant Brookfield, to deploy its advanced fuel cells behind-the-meter at AI data centres. “AI factories demand massive power, rapid deployment and real-time load responsiveness that legacy grids cannot support,” KR Sridhar, chairman and chief executive of Bloom Energy, said in the announcement. The company says it has already deployed “hundreds of megawatts” of fuel cell technology to data centres, working with American Electric Power (AEP), among others.

However, says Kou at BloombergNEF, these solutions are likely to defer rather than avoid increased load coming on to grids around the US. “We see data centres using on-site generation as a temporary solution until they can get connected,” she says. This is, primarily, about reliability. “The grid is a larger power system, and larger power systems provide higher reliability … So, at some point, this demand is going to show up on the grid.”

An evolution

Regardless of how the near-term circle of supply and demand is squared, participants believe the emergence of huge AI data centre operators is likely to bring a new force to US power markets. “Some of these data centre operators are going to be as big in terms of installed capacity in generation and load in the next few years as the largest independent power producers,” says Preedy.  “They’re going to have several gigawatts of load to manage … they are going to become some of the biggest individual power players in the country.”

He notes that some tech giants are already hiring power traders and could become significant sources of short-term supply-and-demand response as they become more sophisticated in managing their power demand in relation to prices. 

Amazon, Google and Microsoft already have approval from the Federal Energy Regulatory Commission to trade electricity, which means they can sell their own surplus supply back to the grid, while Meta filed a Ferc application to trade electricity in November 2025.

However, many of the technology companies building or renting data centres are often unfamiliar with the commodity price risk they are assuming, says Mohit Arora, head of strategy at Mobius. “What they often assume is that if they have bought indexed power or gas, they’re hedged, when all they’re getting is the electron or the molecule.”

“When a hyperscaler gets a power bill from gas generation that had to go and buy $50 gas for a three-day strip, that’s going to be a shock for them,” adds his colleague, Boyce.

Data centre operators face challenges beyond understanding price dynamics. “There’s going to be a lot of things for data centres to overcome before they are able to put any sort of fully fledged commodity platform in place,” Preedy adds, such as handling the accounting implications of marking derivatives books to market. “It’s potentially a big management distraction, so they will have to see this as a core part of their business and invest in the underlying systems and people to manage it.”

This is a development that’s widely expected though, meaning that in the coming years, technology firms may be transformative not only for electricity demand, but for the traded market itself.