The New Bottleneck
In Q4 2025, developers added just 25 gigawatts of electricity capacity to the U.S. data center pipeline according to Wood Mackenzie—half of what was added the previous quarter. This is not a slowdown in demand. It is infrastructure hitting physical limits.
The chip shortage that dominated headlines through 2024 has resolved. Foundries ramped production, lead times normalized, and hyperscalers stockpiled GPUs. But those chips now sit idle in racks because the facilities housing them cannot draw enough power to run them at scale. According to Ben Hertz-Shargel, Wood Mackenzie analyst, utilities lack both grid capacity and generating capacity to build fast enough for these new large energy demand centers.
The constraint has shifted from silicon supply to electron supply. And electrons move slower than semiconductors.
What Is Happening
By the end of 2025, data centers requiring 241 gigawatts of electricity sat in development pipelines according to Wood Mackenzie, an increase of 159% from the beginning of the year. Only one-third of those projects are under active development. The rest face indefinite delays or will never be built.
The numbers tell the story. According to the Lawrence Berkeley National Laboratory, U.S. data centers consumed 176 terawatt-hours in 2023, representing 4.4% of total national electricity demand. By 2028, an estimated 44 gigawatts of additional capacity will be required according to S&P Global Energy. Grid infrastructure coming online in the next three years will provide only 25 gigawatts for data centers according to the Financial Times analysis. That leaves a gap of 19 gigawatts—more than 40% of the power needed.
Morgan Stanley put the shortfall even higher. The bank projects a U.S. power deficit of 44 gigawatts through 2028 before considering alternative time-to-power solutions, amounting to roughly a 20% shortage for data center demand.
Alphabet, Amazon, Meta, Microsoft, and Oracle—the five major hyperscalers—have committed $969 billion to data center infrastructure according to Moody's analysis from February 2026. More than two-thirds of that figure, $662 billion, represents future lease commitments for facilities yet to start construction. The capital is allocated. The sites are selected. The chips are ordered. But the power is not available.
Capital expenditure growth from the largest data center developers will decelerate for the first time since 2023 and match only 58% of last year's growth rate according to Wood Mackenzie. This deceleration is partly driven by Google and Meta choosing to connect through the grid rather than building independent power plants—a decision that trades speed for regulatory simplicity and unknowingly crystallizes the constraint.
Oracle represents the exception. The company has taken on debt to fund Stargate data center campuses powered by behind-the-meter natural gas. On-site generation allows Oracle to bypass grid connection queues and avoid driving up energy prices for surrounding communities. The approach works at the margin but does not scale across the industry.
The Speed-to-Power Era
A new metric now governs data center strategy: speed-to-power. The concept, articulated in a March 2025 report by the Center for Strategic and International Studies, measures how fast a potential data center site can access the electricity needed to power its stock of chips.
Speed-to-power displaces traditional site selection criteria. Proximity to fiber, real estate costs, tax incentives—all of these matter less than the answer to one question: How many gigawatts can this location deliver, and when?
The shift has second-order effects. Data from SemiAnalysis shows over 80 gigawatts of data center capacity under various stages of development could be brought online in the United States by 2030. These facilities could consume over 800 terawatt-hours per year. But the conditional verb matters. Could, not will.
The U.S. has not needed to rapidly expand electricity generation capacity in decades. Utilities operate on planning horizons measured in years, sometimes a decade. AI companies operate on product cycles measured in quarters. The mismatch creates friction that capital alone cannot resolve.
Equipment lead times compound the problem. High-voltage transformers require 18 to 60 months for delivery depending on specifications. Substation construction adds another 12 to 24 months. Interconnection queue backlogs at regional transmission organizations stretch years. A data center developer securing site control today faces a realistic timeline of three to five years before energization.
Meanwhile, AI model training runs happen now. Inference workloads scale now. Competitive advantage accrues to whoever deploys first. The temporal mismatch between infrastructure build cycles and product market dynamics forces hard choices.
Some hyperscalers are financing new power plants directly. Microsoft has discussed nuclear restarts. Amazon acquired a data center campus co-located with a nuclear facility. Meta and Google are exploring partnerships with utilities for dedicated generation. These arrangements can work but require multi-billion dollar commitments, regulatory approvals, and years of construction time.
Bitcoin miners are converting existing facilities into high-performance computing centers. IREN signed a five-year lease with Microsoft. APLD executed a 15-year powered-shell lease with an unnamed hyperscaler. These conversions provide immediate access to energized sites with existing grid connections and utility relationships. The model offers compelling value creation according to Morgan Stanley, especially as power constraints become a defining challenge for AI expansion.
Natural gas turbine transactions could add 15 to 20 gigawatts of supply according to Morgan Stanley. Bloom Energy fuel cells might contribute 5 to 8 gigawatts. Nuclear-powered data center deals could deliver 5 to 15 gigawatts. Combined, these time-to-power alternatives could close part of the gap. But deployment at scale faces permitting hurdles, supply chain constraints, and financing challenges.
What It Means
The power constraint reframes the AI infrastructure investment thesis. Capital flowing into semiconductor equipment, chip design, and server manufacturing assumes those components will be deployed. If deployment stalls due to power unavailability, return on invested capital compresses.
For operators, speed-to-power becomes a competitive moat. Companies with existing energized facilities, utility relationships, and interconnection agreements hold asymmetric advantage. Site control matters less than power control.
For founders building AI-native products, infrastructure availability directly impacts go-to-market strategy. Model training costs may not decline as rapidly as expected if compute scarcity persists. Inference costs face similar pressure. The economics of AI deployment depend on abundant, cheap compute. Power constraints threaten both.
For policymakers, the gap between demand and supply creates risk. If the United States cannot energize data centers at the pace required, AI workloads will migrate to jurisdictions that can. CSIS frames this as a national competitiveness issue. Electricity supply is the most acutely binding constraint on U.S. computational capacity and therefore U.S. AI dominance.
The constraint also creates opportunity. Utilities face the first meaningful load growth in decades. Regulated returns on rate base make power infrastructure attractive for long-term capital. Equipment manufacturers see sustained demand. The entire supply chain from generation to transmission to distribution requires expansion.
But opportunity requires execution. The U.S. power sector has not scaled rapidly in modern memory. Workforce constraints, permitting delays, and equipment lead times create real limits on how fast the system can respond. Five years from now is tomorrow in the power sector, as the CSIS report notes.
Bottom Line
The AI buildout is not chip-constrained anymore. It is electron-constrained. Hyperscalers have committed nearly $1 trillion to data center infrastructure, but only one-third of projects in the pipeline are under active development because the power is not there. The gap between electricity demand and supply will reach 44 gigawatts by 2028 according to Morgan Stanley, leaving roughly 20% of planned capacity dark. Speed-to-power is now the defining competitive metric, and the race for AI supremacy runs through utilities, substations, and generation capacity—not foundries or chip designers. The companies and countries that solve the power constraint first will capture the value creation from AI. The rest will watch their silicon sit idle.