The Thesis
Artificial intelligence has a power problem. Not a software problem, not a chip problem, not a talent problem. A power problem.
The infrastructure buildout required to satisfy AI's energy appetite has collided with a grid that was never designed for this moment. Interconnection queues stretch a decade into the future. Regulators are warning of reliability crises in the nation's most critical transmission corridors. Hyperscalers are restarting decommissioned nuclear reactors, negotiating with oil majors, and signing 20-year power contracts just to guarantee electrons will flow to their facilities.
The conclusion is not subtle: the centralized electrical grid, as currently constituted, cannot deliver the power AI requires at the speed AI requires it. The gap between demand and grid capacity is not a short-term disruption to manage through. It is a structural condition that will define the next decade of AI infrastructure.
The answer is distributed generation. Power produced at or near the point of consumption, independent of transmission infrastructure, built at modular scale, and deployable in months rather than years. This is not a fallback option or a niche workaround. It is the missing infrastructure layer that AI's future depends on.
AI's Power Problem by the Numbers
Start with the scale of what is being built.
Global data center electricity consumption reached approximately 415 TWh in 2024 1, roughly 1.5% of all electricity generated on earth. The International Energy Agency projects that figure will reach 945 TWh by 2030 2, a figure roughly equivalent to Japan's entire national electricity consumption, added to the global grid in six years.
In the United States alone, the picture is starker. U.S. data centers consumed approximately 177 to 192 TWh in 2024 3, representing 4 to 5% of all domestic electricity. The Electric Power Research Institute projects U.S. data center electricity demand will reach 380 to 790 TWh by 2030 4, potentially accounting for 9 to 17% of the entire U.S. grid. Goldman Sachs has revised its estimate of U.S. data center power demand growth upward to 220% above 2023 levels by 2030 5.
The AI component of this demand is growing fastest. Gartner estimates AI-optimized server electricity consumption will rise nearly fivefold, from 93 TWh in 2025 to 432 TWh by 2030 6. AI workloads today represent 15 to 25% of data center electricity use; that share is projected to reach 35 to 50% by 2030 7.
The capital mobilizing around this demand is correspondingly extraordinary. The four largest hyperscalers spent approximately $381 billion in CapEx in 2025 8 and are on track to spend a combined $635 to $665 billion in 2026 9, a 67 to 74% year-over-year increase. McKinsey projects $6.7 trillion in global data center investment through 2030 10.
None of this capital solves the fundamental constraint. You can build facilities. You cannot build a new grid in time.
Why the Grid Cannot Keep Up
The U.S. electrical grid is a 20th-century system being asked to absorb a 21st-century demand shock on a timeline measured in months. There are four compounding reasons why it cannot respond fast enough.
Interconnection Queues
As of the end of 2024, approximately 2,300 GW of generation and storage capacity 11 was actively seeking grid connection in the United States, roughly double the entire current installed U.S. generating capacity. Lawrence Berkeley National Laboratory data shows that only 14% of capacity that entered interconnection queues between 2000 and 2018 12 had actually been built by the end of that period. Projects that reached commercial operation in 2025 spent an average of eight years in the interconnection queue 13. In some markets, utilities have quoted timelines of 12 years simply to study a new interconnection request.
The arithmetic is unforgiving. Data centers must be operational in 18 to 36 months. Grid connections take 5 to 12 years. That is a structural gap, not a scheduling inconvenience.
NERC's Reliability Warnings
The North American Electric Reliability Corporation's 2025 Long-Term Reliability Assessment 14 found that summer peak demand could surge by 224 GW over the next decade, 69% more than the prior year's projection. Winter demand could surge by 245 GW. These are the highest compound annual growth rates in peak demand since NERC began tracking the metric in 1995. Five regions now carry a "HIGH RISK" designation by 2030: MISO, PJM, ERCOT, WECC Basin, and WECC Northwest. Thirteen of 23 assessment areas face resource adequacy challenges within five years.
NERC Director John Moura has stated plainly: "We see real load growth. The uncertainty and magnitude of load growth, and its impact on planning, is increasingly uncertain and has significant risk." 15
Market Saturation at Traditional Hubs
Northern Virginia, the largest data center market in the world, now has vacancy rates below 1% and interconnection queues extending 5 to 7 years. Colocation pricing has risen from approximately $120 per kW-month in 2021 to $184 per kW-month by late 2024 16, a 53% increase driven almost entirely by power scarcity. Dublin imposed a moratorium on new data center connections. Amsterdam and Singapore face similar constraints. The most established markets are functionally closed to new large-scale development.
AEP, the utility serving portions of Virginia and Texas, has customer commitments for 24 GW of new demand by 2030 17, including 18 GW from data centers alone, which would be five times the utility's current system size.
The Hardware Bottleneck
Even where grid capacity exists, physical infrastructure limits the pace of expansion. Large transformer lead times now exceed 80 to 120 weeks 18, with some transmission-class units requiring 3 to 6 years. GE Vernova confirmed in late 2025 that its order backlog reached a record 80 GW 19 against annual output of 20 GW, effectively sold out through 2029. Every announced gigawatt of data center capacity requires physical infrastructure that the supply chain cannot produce fast enough.
The result: 30 to 50% of large data center projects 20 scheduled to come online in 2026 are expected to be delayed, with power cited as the primary cause.
The Case for Distributed Generation
If centralized generation and long-distance transmission are the bottleneck, the answer is to bypass both. Distributed generation, meaning power produced at or adjacent to the point of consumption, eliminates the interconnection queue problem at its source.
Behind-the-meter generation does not require transmission infrastructure. It does not compete for space in an interconnection queue. It is not subject to the hardware supply constraints affecting transformer procurement. It can be permitted and deployed in a fraction of the time required for grid-connected facilities.
The case rests on three structural advantages.
Speed to capacity. A modular, skid-mounted power generation and compute unit can be operational in months. Grid-dependent facilities cannot guarantee power availability within the planning horizon of most AI infrastructure investments. In an industry where competitive advantage is measured in training runs and inference latency, time-to-capacity is not a secondary consideration.
Stranded and underutilized energy assets. The United States has vast reserves of natural gas that are effectively stranded: too remote from pipeline infrastructure to reach commercial markets, too abundant to justify pipeline investment, and currently being burned off as waste through flaring. Over 140 billion cubic meters of natural gas are flared globally each year 21, representing more than $30 billion in wasted energy value annually. In the Williston Basin of North Dakota alone, flaring volumes run into the hundreds of millions of cubic feet per day. That stranded gas has a cost basis measured in fractions of a cent per kilowatt-hour. Capturing it for on-site power generation converts an environmental liability into productive computational infrastructure.
Reliability. Power failures account for 52% of all significant data center outages 22, the single largest cause. Grid-dependent facilities are exposed to weather events, transmission failures, and the systemic reliability risks NERC is now flagging. On-site generation with redundant fuel supply is structurally more reliable than a shared transmission system under increasing load.
Training vs. Inference: Why Distribution Matters More Than Ever
AI infrastructure has two distinct workload profiles, and they have different geographic requirements.
Model training is compute-intensive, latency-tolerant, and suited to centralized infrastructure. Training a large language model requires massive GPU clusters running continuously for weeks or months. The power density is extreme, the cooling requirements are demanding, and the proximity to end users is irrelevant. This is the workload that justifies billion-dollar campuses in West Texas and North Dakota.
Inference is different. When a model is deployed to serve users, each query must be answered in milliseconds. Latency becomes a competitive and technical constraint. According to CBRE's North America Data Center Trends analysis 23, "the shift from AI training to AI inference demand is creating a need for more regional and distributed data centers, not just hyperscale hubs. Inference workloads often need to be near end users, reshaping site strategy." As of late 2025, the edge data center market was valued at $14.7 billion and is projected to reach $71.9 billion by 2035 24 at a CAGR of 17.5%.
Sovereign AI is accelerating this trend further. Governments and regulated industries increasingly require data to be processed within national or regional boundaries. AI inference for financial services in Germany cannot route through Virginia. Healthcare AI in Japan cannot depend on a data center in Oregon. These requirements create demand for distributed compute nodes in locations that have never hosted large-scale infrastructure, locations that are often underserved by traditional grid infrastructure and well-suited to distributed generation.
The inference economy does not need one gigawatt in one place. It needs hundreds of megawatts distributed across dozens of locations. Distributed power generation is the only model that can serve it at that scale and that speed.
The Economics of Distributed Power
The economic case for distributed generation is not marginal. It is decisive.
Grid electricity for commercial and industrial users currently ranges from $0.07 to $0.13 per kWh 25 in most U.S. markets, with prices rising approximately 7% annually since 2019. In power-constrained markets like Northern Virginia, effective all-in costs are substantially higher when colocation premiums are factored in. Renewable PPAs, often cited as a cost-effective alternative, are themselves rising: North American solar PPA prices reached $61.67 per MWh in Q4 2025 26, a 9% year-over-year increase, with no price ceiling in sight.
Stranded natural gas at a wellhead, by contrast, has an acquisition cost of 10 to 30 cents per MCF 27, versus market prices that routinely exceed $3 per MCF. When that gas is converted to electricity on-site, the fully loaded power cost falls below one cent per kilowatt-hour 28. That is a 7x to 13x cost advantage over grid electricity, before accounting for the capital cost of grid interconnection, transmission charges, and the time-value of delayed deployment.
The modular capital structure amplifies this advantage. Traditional data center development requires large upfront commitments against uncertain power delivery timelines. Modular, skid-mounted units can be deployed incrementally, matching capital expenditure to contracted revenue. The first unit can be operational and generating cash while subsequent units are in transit. This is not how large utility-scale infrastructure works, and that difference in capital efficiency matters in a market where competing for power is simultaneously the largest cost and the longest lead time.
The financial market has absorbed this logic. Data center investment deals reached a record $61 billion in 2025 29, with data center debt issuance nearly doubling to $182 billion. Applied Digital, which builds purpose-built AI data centers in secondary markets including North Dakota, saw its stock appreciate 237% in 2025 30, the top-performing internet services stock of the year, after signing a $5 billion lease 31 with an investment-grade hyperscaler for its second North Dakota campus.
What the Market Is Telling Us
The clearest signal that the grid cannot meet AI's power needs is not found in infrastructure reports or NERC assessments. It is found in the increasingly unusual lengths to which hyperscalers are going to secure reliable power outside traditional channels.
Microsoft signed a 20-year, 835 MW Power Purchase Agreement 32 with Constellation Energy to restart Three Mile Island Unit 1, a reactor decommissioned in 2019. Constellation invested $1.6 billion 33 in the restart, with the U.S. Department of Energy providing an additional $1 billion loan. Microsoft committed to taking 100% of the plant's output for two decades.
Amazon invested $20 billion 34 to convert the Susquehanna nuclear site into a dedicated AI-ready campus, secured a $1.8 billion PPA with Talen Energy 35 for 1.9 GW through 2042, and committed to backing 5 GW of small modular reactor projects through X-energy. Google signed a deal with Kairos Power for 500 MW of small modular reactors 36, the first corporate SMR fleet deal in U.S. history. Meta issued an RFP for 1 to 4 GW of new nuclear generation 37.
In aggregate, major technology companies signed contracts for more than 10 GW 38 of possible new nuclear capacity in the United States in a single year. Goldman Sachs forecasts 85 to 90 GW of new nuclear capacity 39 needed globally by 2030 to meet data center demand alone, less than 10% of which is currently available.
Chevron, an oil and gas company with no prior experience in data center infrastructure, announced plans to build a 2.5 GW off-grid natural gas facility 40 in West Texas, the company's first-ever power plant, expressly designed to serve hyperscale AI clients. Pacifico Energy received the largest air permit ever granted in the United States, for 7.65 GW of gas-fired generation 41 at a single off-grid site in Pecos County, Texas.
One-third of hyperscalers and colocation providers 42 now plan to bring power production entirely on-site by 2030, according to Bloom Energy, a 22% increase from a survey conducted just six months earlier. On-site and hybrid approaches represent less than 10% of announced data center projects by count but account for approximately 50% of announced capacity in megawatts 43, driven by a small number of large, grid-independent campuses.
These are not the decisions of companies that believe the grid problem is temporary. They are the decisions of companies that have concluded, through hard experience, contract negotiations, and capital allocation, that reliable, low-cost power outside traditional grid infrastructure is a structural necessity.
The pioneer of flare-gas-to-compute, Crusoe Energy, sold its entire Digital Flare Mitigation business to NYDIG in March 2025 44 and pivoted to centralized hyperscale AI development. The technology it built, modular containerized data centers powered by wellhead natural gas, was not shut down. It was acquired, at scale, because the model works. Crusoe's pivot to centralized AI did not invalidate the distributed model; it vacated it, leaving an operational gap in the precise market where demand is accelerating.
The Path Forward
Distributed power generation is not a temporary workaround for an industry in transition. It is a structural solution for a constraint that will persist for at least the next decade.
The math is straightforward. Goldman Sachs estimates 47 GW of extra U.S. generation capacity 45 is required within five years and $790 billion of grid capital expenditure 46 through 2030. BCG warns of a potential shortfall of up to 80 GW of firm power 47 by 2030. The IEA estimates 20% of planned data centers 48 could face delays simply getting connected to the grid. Sightline Climate projected that 30 to 50% of data center capacity 49 scheduled for 2026 would be delayed.
None of those gaps close by building more centralized infrastructure connected to a constrained transmission system. They close by building generation at the point of demand.
The geography of stranded energy makes this more than a theoretical proposition. The Williston Basin in North Dakota contains some of the most productive oil-producing formations in the United States, along with natural gas reserves that far exceed pipeline takeaway capacity. Flaring, the practice of burning off that gas as waste, is a regulatory problem, an environmental problem, and an economic inefficiency. Deploying modular power generation and GPU compute infrastructure directly to those sites converts stranded energy into productive capacity, without waiting for a grid connection, without competing for a spot in an interconnection queue, and without depending on transmission infrastructure that regulators are warning will be under acute stress for years to come.
The infrastructure layer between stranded energy and AI compute demand does not yet exist at scale. The grid was never going to provide it. Distributed generation is the architecture that can.
AI's power constraint is not a problem that will be solved by the same centralized infrastructure model that created it. The operators who recognize that earliest will have the capacity to serve the next decade of AI growth. Those who wait for the grid will wait for years.