The leading AI firms plan to spend more than $600b in 2026 as they rush to open datacentres.
Speed is everything. Price is no object. Many of the new projects will not be connected to the power grid but will rely on on-site electricity generation, and some 75% of the new facilities will use gas for this. The current forecasts for US gas demand do not account for this recent development. The surge in AI demand could push gas and perhaps diesel and jet fuel prices to new highs. Electric utilities and consumers will likely complain about the increases. US LNG exports will fall as supplies become uncompetitive with gas from countries such as Qatar and Australia, and liquefaction plants in the US will become stranded assets.
The AI boom may also drive diesel consumption higher. “The US is looking to tap the nation’s network of large industrial diesel generators used at datacentres, big box stores and elsewhere to help curb rising electricity costs and support the surge in power demand from AI”, according to Energy Secretary Chris Wright. The secretary needs to look again at the numbers. US diesel use in an already tight market could rise by 10% should his plan come to fruition. Farmers and truckers might see prices of $5–6/gal, and the increases could come just before the US midterm elections in November.
The Supreme Court's 20 February decision nullifying President Donald Trump’s authority to impose tariffs under the International Emergency Economic Powers Act technically freed nations coerced into purchasing energy exports from the US from their obligations. Many will renegotiate using their new leverage. US LNG exporters will likely also be squeezed by price increases driven by rising AI datacentre demand and by the availability of lower-cost LNG from other nations.
This is the new AI energy world
Few industries in modern history have expanded as rapidly, or as disruptively, as AI. What began as an arms race between a handful of technology firms has—in just two years—become the largest capital‑intensive industrial surge since the shale revolution. Unlike shale, however, the AI buildout is not merely reshaping a sector; it is on track to eclipse the global fossil fuel industry in spending and, increasingly, in influence over energy flows themselves.
By 2026, the four largest US AI firms are expected to spend sums that dwarfs entire national energy budgets. Their datacentre buildout is proceeding at breakneck speed, with cost, efficiency, and even regulatory predictability subordinated to a single imperative: getting the infrastructure online as fast as possible. Gas demand forecasts are broken and the US LNG’s role in global trade will need to be reassessed. The consequences for oil, gas, electricity prices and global energy geopolitics are just starting to surface.
The new energy superpowers
For decades, executives at oil and gas giants such as ExxonMobil, Saudi Aramco and Chevron have presided over the world’s largest capital allocation programmes. That hierarchy is being overturned. In 2026, ExxonMobil expects to invest $27–29b—a figure that would have commanded headlines in any previous era. Amazon alone, however, plans to spend $200b this year. The four largest US tech giants combined are set to deploy $650b, potentially eclipsing global upstream oil, gas and coal investment, which the IEA estimated at $1.15t in 2025.
This inversion is not a one‑year anomaly. In 2023, the International Energy Forum called for a 28% increase in annual upstream oil and gas investment to reach $640b by 2030. US tech firms will spend more than that in 2026 alone. Energy companies are understandably cautious: ExxonMobil’s executives continue to emphasise capital discipline even as they budget $27–29b of capex for 2026. Yet it is hard to ignore the scale of tech’s spending—and how quickly it is being translated into physical infrastructure.
AI companies, not oil majors or national energy ministries, could soon become the pivotal actors shaping hydrocarbon markets—first in North America and potentially worldwide.
Datacentres: A new industrial load
The magnitude of the buildout shows up starkly in the facility counts. According to America’s AI Surge, a report by advocacy group American Edge Project, the US hosts 4,149 operational datacentres, with an additional 2,788 either planned or under construction. The scale within individual states is staggering:
Virginia: 663 operational, 595 planned or under construction
Texas: 405 operational, 442 planned or under construction
Georgia: 162 operational, 285 planned or under construction
Pennsylvania: 98 operational, 184 planned or under construction
But the numbers most relevant to energy markets are on the demand side. US datacentres consumed 183TWh of electricity in 2024—more than 4% of national electricity demand, roughly equivalent to Pakistan’s annual electricity use. By 2030, IEA estimates point to 426TWh—a 133% increase.
In 2024, the power mix for datacentres depended heavily on fossil fuels:
40% gas
24% renewables
20% nuclear
15% coal
Earlier projections suggested gas would account for the largest share of new supply through 2030, adding roughly 130TWh, followed by renewables at 110TWh. But those projections are already obsolete. The grid’s interconnection queues are long; utilities face political pressure to hold consumer rates down; and the scale and concentration of AI loads are straining substation and transmission capacity in multiple states.
Grid constraints push datacentres off grid
The speed at which AI developers are building makes it impractical to rely solely on traditional grid connections. The result is a dramatic shift: behind‑the‑meter generation—large‑scale, on‑site electricity production purpose‑built for datacentres.
A recent Cleanview Energy Partners analysis highlights the pivot:
c.75% of planned datacentre generation will be powered by gas.
46 centres already plan to provide their own power.
Behind‑the‑meter capacity surged from 0% in early 2024 to c.55% of planned capacity by January 2026.
Developers are moving so quickly that turbine manufacturers report lead times of 5–7 years, forcing buyers into unconventional options: mobile gas generators strapped to semi‑trailers, aeroderivative turbines originally designed for aircraft and warships, fast‑ramping reciprocating engines that sacrifice efficiency, and refurbished industrial units. One hyperscaler, unable to source enough conventional turbines, reportedly placed a $1.25b order with Boom Supersonic—an aerospace firm with no prior history in power generation.
The economic logic is blunt. When each megawatt of AI capacity can generate $10–12m per year in revenue, the opportunity cost of delay overwhelms fuel and efficiency considerations. Power generation efficiency is out; speed to power is everything.
Estimating the datacentre impact on US gas markets
Data on the number and size of hyperscale facilities is sparse; data on their actual energy use is even scarcer. To frame plausible ranges, a practical approach is to translate planned capacity into annual electricity and then into fuel consumption.
Gas consumption for the Cleanview‑identified behind‑the‑meter capacity falls in the range of 3.2–3.8tcf/yr if all such capacity ultimately runs on gas. That implies 270–320bcf per month if all additions were to come online in 2026—an aggressive assumption that is unlikely to be fully realised but illustrates the scale of the upper bound.
To triangulate with a more conservative lens focused on plants actually in service (rather than planned capacity), financial services firm Jefferies’ equipment‑supply and commissioning analysis suggests a stepped increase of approximately 89bcf per month in 2027, followed by 77bcf per month in 2028, attributable to new gas‑fired generation supporting datacentres. Independent tallies from Measuring the Data Center Boom: Facts and Statistics, published by cybersecurity education site Programs.com, citing McKinsey, arrive at similar deltas based on observed project pipelines.
These increments do not appear to be reflected in the Energy Information Administration’s most recent short‑term forecast, which shows essentially no growth in US gas consumption over the relevant period. The mismatch between observed project activity and official demand projections raises the risk of an avoidable price shock if supply and infrastructure adjustments lag the true trajectory of demand.
The Lawrence Berkeley National Laboratory’s widely cited December 2024 study estimated datacentre electricity consumption at 176TWh in 2023 (4.4% of US load), with 2028 scenarios spanning 325–580TWh (6.7–12.0% of US consumption). But two caveats apply: the report is already dated relative to the 2025–26 AI capex surge, and it did not explicitly model the rapid shift to behind‑the‑meter generation—a development now central to the sector’s growth.
The strategic pivot is visible in corporate moves. Williams Companies—traditionally a pipeline operator—has ploughed $7b+ into “power innovation” projects since 2024, standing up nearly 2GW of behind‑the‑meter generation for grid‑constrained customers in two states. Its Aquila (Utah) and Apollo (Ohio) ventures have been upsized, and its Socrates project in Ohio is expanding into a second phase. Notably, Williams is exploring direct acquisition of gas production to vertically secure supply for datacentres—blurring the lines between upstream and midstream. Other firms are taking similar steps.
The message is clear : if the grid cannot deliver, datacentre owners will—through private generation largely fuelled by gas.
The coming gas squeeze
The combined effect of these projects is significant. The Cleanview‑based range suggests fully operational behind‑the‑meter capacity could lift US gas demand by up to 10% on a sustained basis. There is little evidence that domestic production can increase at equivalent speed without commensurate price signals, infrastructure expansions or both.
In the near term, higher prices are the only mechanism with enough speed and force to rebalance the market. Short‑run demand elasticities are low:
Residential demand is generally price‑insensitive in the short run (often estimated at elasticities around –0.1 or less). Households may be vocal about rising bills but will not rapidly reduce winter heating demand.
Utility gas demand is likewise relatively price‑insensitive, given limited near‑term fuel‑switching options and regulatory pathways to pass through higher fuel costs.
Industrial demand is somewhat more elastic—some chemical, metals and process industries can reduce operating rates or shut marginal capacity—but this segment represents only c.25% of US gas consumption, limiting the aggregate relief it can provide.
To clear the market, a 20–30% increase in wellhead prices may be required on current assumptions. That level, in turn, would most quickly manifest in the export channel:
US LNG is the marginal outlet for domestic gas. As US prices rise, American cargoes become less competitive against lower‑cost producers.
Qatar, already the lowest‑cost LNG supplier, is doubling export capacity. It has pushed term contracts linked to c.12–13% of Brent, which—at prevailing oil prices—often undercut the spot‑parity price implied by rising US wellhead gas and liquefaction costs.
If domestic prices rise further due to AI demand, US LNG exports could be cut by one‑half to two‑thirds as traders and offtakers pivot to cheaper long‑term supplies from Qatar and others. American LNG terminals—built on assumptions of long‑run competitiveness—risk periods of underutilisation, potentially stranding portions of the asset base.
The Supreme Court’s 20 February decision limiting presidential tariff authority under the International Emergency Economic Powers Act adds a geopolitical twist: countries previously pressured to purchase US energy exports can now renegotiate from a position of greater leverage. Combined with a domestic price‑driven squeeze and aggressive Qatari marketing, this could accelerate a rebalancing of LNG trade flows away from the US.
Why renewables will not fill the gap fast enough
Under normal circumstances, a spike in electricity demand would catalyse a surge in renewables and storage. But the timing and policy environment cut the other way:
The federal government has blocked or delayed dozens of large wind and solar projects on federal land over the past year. Projects on private land that require federal approvals (e.g., Fish & Wildlife Service) have also slowed.
Even where projects are permitted, interconnection queues and transmission bottlenecks impose multi‑year delays ill‑suited to developers the economics of which hinge on being first to market.
To be clear, renewables will continue to grow and improve their share of the energy mix. But the next 24–36 months—the period that matters most for this wave of AI buildout—are likely to be dominated by gas‑fired solutions, especially behind the meter. Some capacity will be hybridised with on‑site solar and batteries, but the dominant firm‑power solution remains natural gas.
The diesel backstop—and its costs
If gas is tight or interconnection delays persist, datacentres can and will turn to diesel—if only as a bridge or backup. Many facilities already maintain extensive diesel generator farms for reliability.
The US Department of Energy has floated tapping the nation’s network of large industrial diesel generators—at datacentres, big‑box retailers and other sites—to curb rising electricity costs and support AI demand. The theoretical scale is striking, with reports that this could unlock electricity equivalent to c.35 conventional nuclear plants without building new power stations. However, the fuel math is unforgiving:
A large diesel generator consumes roughly 3bl/hr to produce 2,000kWh (per General Power’s technical specifications).
It would take c.57 such units operating continuously to produce 1TWh/yr.
That equates to c.4,600b/d day per TWh at best‑in‑class efficiencies; more realistic averages suggest c.5,000b/d per TWh.
Scaling to 35GW of continuous diesel generation implies at least 175,000b/d of incremental diesel demand—and potentially c.200,000b/d if smaller, less efficient units are used. In an already tight global diesel market, that volume could push prices 20–50% higher. US retail diesel could rise towards $5/gal (from a baseline around $3.71/gal) even if only a portion of this capacity runs for extended periods.
The political optics are not trivial. Independent truckers shut down freight movement over price spikes in 1979; the ‘Tractorcade’ protests the same year saw thousands of farmers rally over fuel prices and crop economics. A strategy that leans on diesel to modulate electricity prices could backfire quickly if it ignites fuel inflation for agriculture and freight—especially heading into the 2026 midterm elections. There will be winners and losers in this scenario.
Among the winners will be pipeline and midstream firms that pivot to private power (e.g., Williams), which are well positioned. The ability to bundle gas supply, small‑to‑mid‑scale generation, and reliability services behind the meter is a compelling offer to hyperscalers. Gas turbine and engine OEMs—including suppliers of aeroderivative turbines, reciprocating engines and even refurbished equipment—will face a multi‑year demand surge. And flexible gas producers with access to gathering, takeaway and offtake options aligned to industrial loads (as opposed to LNG) could enjoy premium realisations.
The losers will include US LNG exporters, which are most exposed if domestic prices rise, export netbacks compress and long‑term Qatari (or Australian) contracts become more attractive to buyers. Elsewhere, industrial gas consumers with higher price sensitivity—certain chemical and metals segments—may be forced to reduce runs or pause capacity as marginal economics deteriorate. Utilities and ratepayers, meanwhile, face a complex mix of outcomes: higher wholesale power prices where gas sets the marginal megawatt; local rate pressure where grid upgrades are required; and, paradoxically, potential load erosion if large campuses defect from the grid to private power.
Policy and market implications
Modernise interconnection and transmission, fast: The single most effective way to reduce behind‑the‑meter reliance on gas and diesel is to expand transmission and streamline interconnection. Multi‑year queues that push critical loads to private generation are a policy choice as much as an engineering inevitability.
Targeted capacity markets for reliability: Well‑designed capacity mechanisms (or equivalent reliability products) can incentivise cleaner firming resources—long‑duration storage, advanced nuclear, CCS‑equipped turbines—without defaulting to diesel.
Permitting reform with guardrails: Rapid growth creates externalities; the answer is better siting and expedited but rigorous reviews, not blanket slowdowns that push hyperscalers towards dirtier stopgaps.
Efficiency standards and transparency: Server efficiency, liquid cooling and workload optimisation remain crucial and underexploited. Requiring standardised reporting on energy use and fuel sourcing for large datacentres would improve planning and grid coordination.
LNG strategy recalibration: If domestic AI demand lifts US gas prices, export competitiveness will swing. Policymakers and developers should revisit capacity additions, contract structures and exposure to oil‑linked benchmarks favoured by Qatar.
The energy future no one planned for
The AI industry’s capital spending is no longer just large—it is system‑defining. It is reshaping gas demand curves, tightening diesel markets, challenging LNG export economics, straining electrical grids and accelerating a shift towards private power generation on a scale unseen since early industrialisation. The sector’s operating philosophy—cash is not an issue; speed to power is everything—means traditional energy price dynamics may no longer apply. Developers will buy whatever fuel is available, at whatever price, because the opportunity cost of delay dwarfs the operational cost of energy.
This creates a new world in which AI hyperscalers, not oil majors or utilities, set the marginal price and flow of hydrocarbons in US markets—and, if current trends hold, increasingly influence global markets as well. The direct consequence is a likely gas squeeze, spilling into LNG trade flows and forcing hard choices about which loads get served and at what price. Diesel as a backstop is neither scalable nor politically palatable; it is a bridge that becomes costly the moment you step onto it.
None of this is inevitable. Better policy on interconnections and transmission, smarter capacity incentives, thoughtful permitting and aggressive efficiency could bend the curve. But the next two years are already largely scripted by the project pipelines and procurement orders in place. The AI boom has made data the newest energy‑intensive commodity—and transformed the companies that compute it into the most powerful energy buyers on earth.
Dr. Philip Verleger received his PhD in Economics from MIT in 1971. He has studied energy and financial markets since. In 2023, Dr. Verleger was named 'Energy Writer of the Year' along with his editor Kim Pederson. Previous winners of the award include Daniel Yergin (2000) and Vaclav Smil (2019).