AI is scaling faster than the energy systems built to power it. The latest wave of model training, inference, and always-on digital services is pushing data-center electricity demand into the spotlight, not as a distant sustainability issue, but as a near-term constraint that can slow deployments, raise costs, and reshape where compute gets built.
Across the U.S. and globally, the risk is no longer just “do we have enough generation?” It is whether grids, interconnection queues, equipment supply chains, permitting processes, and local communities can absorb sudden, concentrated load growth. In other words: a power crunch could throttle AI-driven growth.
1) Why electricity is becoming AI’s limiting factor
Data centers already consume meaningful electricity at the global scale. The International Energy Agency (IEA) estimates global data-center electricity use was about 415 TWh in 2024, roughly 1.5% of world demand, and projects it could more than double to around 945 TWh by 2030 (about 3%). Those percentages sound manageable until you remember that data centers cluster in specific regions, so local grids can hit capacity limits long before national totals look stressed.
In the U.S., the concentration problem is especially acute. The IEA notes that data centers are projected to drive nearly half of U.S. electricity-demand growth to 2030, and that U.S. per-capita data-center electricity consumption could rise from about 540 kWh in 2024 to more than 1,200 kWh by 2030. That kind of jump turns “load growth” from an annual planning assumption into a structural shift for utilities and regulators.
Domestic datasets tell a similar story. The U.S. Department of Energy (DOE), summarizing an LBNL report, said data centers were about 4.4% of U.S. electricity in 2023, with projections ranging to 6.7% and 12% by 2028; electricity use rose from 58 TWh (2014) to 176 TWh (2023), and could reach 325, 580 TWh by 2028. When growth happens this quickly, the constraint often becomes “how fast can we connect and deliver power,” not “is there any power somewhere in the system.”
2) Grid bottlenecks, interconnection queues, and long lead times
Even when generation exists, data centers still need wires, substations, transformers, and interconnection approvals. Goldman Sachs has warned that incremental supply is constrained by permitting delays, supply chain bottlenecks, and costly upgrades, and has estimated grid investment needs of roughly $720 billion through 2030, projects that often take years from planning to energization.
Real-world timelines illustrate the issue. In Northern Virginia, a major data-center hub, reporting in 2024 cited Dominion indicating large data centers could face as long as seven years to connect to the grid. For AI developers, a multi-year wait is effectively a capacity cap: model roadmaps and product launches cannot depend on interconnections that arrive after the market moves on.
Texas offers another view of the same problem, queues that outgrow practical build rates. In late 2025, reporting described an ERCOT surge to roughly 226 GW of large-load interconnection requests, with 70%+ tied to data centers. ERCOT also made organizational changes to address rapid demand growth and interconnection modernization, including a group explicitly focused on enterprise data and artificial intelligence, an institutional signal that the planning burden itself is escalating.
3) Equipment supply chains are now part of the AI story
AI growth is not only constrained by permitting and transmission corridors; it is also constrained by physical hardware that sits between power plants and server racks. Transformers, switchgear, and gas-turbine components have become gating items in many grid-upgrade projects, and long lead times can delay both utility and customer-side buildouts.
On Feb. 4, 2026, Siemens Energy announced a $1 billion U.S. manufacturing expansion aimed at relieving grid-equipment bottlenecks. The company said it will expand domestic production of transformers and switchgear as well as gas-turbine equipment, explicitly tying the move to rapidly rising electricity demand connected in part to AI data centers. That is a telling inversion: rather than power equipment quietly supporting the digital economy, the digital economy is now forcing a reindustrialization of grid hardware.
Some reporting around the Siemens Energy expansion also cited expectations that U.S. electricity demand could rise by about 25% by 2030, underscoring why the “power crunch” narrative has traction. If demand rises that quickly, the supply chain for grid equipment becomes a strategic lever, because even well-funded plans can stall if utilities cannot procure the components needed to deliver power.
4) When the grid can’t move fast enough, on-site power moves in
As grid connections lag, developers are revisiting “behind-the-meter” strategies: on-site generation and storage that reduce dependence on utility interconnections. Reporting in Feb. 2026 described accelerating interest in on-site fuel-cell generation for AI data centers, driven by the need for fast, reliable power and frustration with slow grid buildouts and aging infrastructure.
Goldman Sachs estimates cited in that reporting suggest fuel cells could supply up to 15% of the AI sector’s power by 2030. Whether or not that exact share materializes, the direction is important: investment is shifting toward solutions that deliver predictable timelines, even if they add complexity in operations, fuel sourcing, and emissions accounting.
Texas also exemplifies a more extreme bypass: “private grid” or dedicated generation campuses. Feb. 2026 reporting described a 7.65 GW gas-powered campus permitted to supply data centers, with first power projected in early 2027 and described as the largest U.S. air permit for gas generation; the project includes batteries and solar. The willingness to build huge, dedicated generation, rather than wait for public-grid upgrades, signals how strongly interconnection constraints can shape AI infrastructure decisions.
5) The political backlash: bills, rates, and “opaque deals”
Power constraints are not only engineering problems; they become political problems once they show up in monthly bills. Feb. 2026 coverage linked rising U.S. electricity bills and “opaque deals” to the strain from AI data-center buildouts interacting with outdated grid infrastructure. When voters feel the cost, permitting and expansion can slow as public officials demand new conditions or moratoria.
Washington State’s proposed HB 2515, reported Feb. 3, 2026, reflects this pressure. The bill would require utilities to adopt policies for servicing large data centers amid concerns that data-center growth could raise other customers’ bills and complicate decarbonization goals. Reporting quoted the motivation directly: the move comes in response to concerns the growth could lead to higher utility bills for other consumers.
For AI operators, these politics matter operationally. If regulators shift grid-upgrade costs to data centers, require new rate structures, or tighten service requirements, electricity becomes less like a commodity input and more like a negotiated constraint, potentially changing where companies build, how quickly they can expand, and whether smaller AI players can compete with hyperscalers.
6) The emissions dilemma: gas buildout as the “fast” answer
When regions experience a power crunch, the fastest capacity additions often come from familiar thermal technologies. But that speed can collide with climate targets and long-lived infrastructure lock-in. Reporting on Global Energy Monitor findings (Jan. 29, 2026) said a large share of planned gas-fired capacity is being justified by data-center demand, raising concerns that AI growth could indirectly drive decades of additional emissions.
The dedicated-generation example in Texas demonstrates the tradeoff in concrete form: rapid, dispatchable capacity that can serve data centers on predictable schedules, but with significant air-permit and emissions implications, even if partially hybridized with batteries and solar. Once built, these assets shape marginal costs and policy debates for years, influencing not only AI’s footprint but the broader power mix.
Policy leaders are trying to thread the needle. DOE Secretary Jennifer Granholm has acknowledged that data-center growth has created greater demand on domestic energy supply while arguing the U.S. can meet this growth with clean energy. The challenge is timing: clean generation plus transmission and interconnection upgrades can take longer than the AI market’s appetite for new compute.
7) What “throttling” looks like in practice for AI and cloud growth
Power constraints rarely appear as a single line saying “AI canceled.” More often, they show up as slower ramp schedules, delayed availability zones, higher prices for colocated capacity, and geographic reshuffling as developers chase faster interconnections. Gartner predicted in 2024 that 40% of existing AI data centers could be constrained by power availability by 2027, arguing that insatiable demand for power will exceed utilities’ ability to expand fast enough. As a forecast, it is not destiny, but it captures the shape of the risk.
Financial forecasts also imply near-term gaps. Goldman Sachs projected (Feb. 2025) global data-center power demand could rise 50% by 2027 and 165% by 2030 versus 2023, and described a baseline where demand reaches about 84 GW by 2027 versus about 59 GW of “current” capacity, highlighting why transmission, permitting, and equipment lead times can become hard limits.
Even when national generation is adequate, local saturation can force tough choices. Pew Research synthesis (Oct. 2025), drawing on IEA/EPRI/LBNL, noted U.S. data centers used about 183 TWh in 2024 and could reach around 426 TWh by 2030 (+133%). It also pointed to places like Virginia, where data centers used about 26% of the state’s electricity supply in 2023, exactly the type of concentration that triggers long queues, higher rates, and political scrutiny.
8) A second constraint: water, cooling, and local acceptance
Electricity is only part of the physical footprint. Cooling requirements can amplify siting constraints, especially in regions with water stress or heightened scrutiny of industrial consumption. Pew/LBNL noted U.S. data centers directly used around 17 billion gallons of water in 2023, with hyperscale and colocation facilities accounting for about 84%.
These impacts can slow projects through permitting, mitigation requirements, or public opposition, especially when communities perceive that the benefits (jobs, tax revenue) do not match the burdens (water use, land use, higher bills). The IEA has emphasized that the largest announced data center could use as much electricity as five million households, a scale comparison that can sharpen local resistance and complicate approvals.
When water and power constraints combine, developers face a narrower set of viable sites. That can lead to “compute islands” in regions with favorable utility relationships and infrastructure room, intensifying concentration, and ironically making the next wave of constraints even more severe.
AI’s next phase will be shaped not only by chips and algorithms, but by substations, transformers, and the politics of who pays for grid upgrades. With global data-center electricity demand projected by the IEA to more than double by 2030, and U.S. growth expected to be heavily driven by data centers, the risk of a power crunch throttling AI-driven growth is no longer theoretical.
The upside is that the constraint is also a roadmap. Siemens Energy’s $1 billion manufacturing push, ERCOT’s interconnection modernization efforts, the rise of behind-the-meter fuel cells, and even contentious state-level bills all point to the same conclusion: energy and AI are now co-dependent industries. The winners will be the organizations that treat power procurement, grid strategy, and community impact as core product inputs, because in the AI era, electricity is a competitive advantage.
