Everything About AI
NextFin News - The artificial-intelligence boom has moved past the chip shortage into a harder constraint: electricity. Data center power demand is set to more than double by 2030, and the bottleneck is no longer just the grid - it is the neighbors, the ballot box, and the years-long queue for transformers and transmission lines.
The International Energy Agency projects data center electricity consumption will more than double to around 945 terawatt-hours by 2030, slightly more than Japan's total annual electricity use today. In the United States, data centers account for nearly half of electricity demand growth over the same period. Yet the North American Electric Reliability Corporation warns that nearly all of the country will face higher risk of energy shortfalls over the next five to ten years, and up to one-fifth of planned data center projects could face delays without major transmission investment.
This is not a cyclical squeeze that clears when supply chains normalize. It is a structural collision between a technology that scales exponentially and an energy system that scales linearly - and the fight over who gets to build, where, and with what power source has only just begun.
The Constraint Has Moved From Chips To Electrons
For two years, the AI narrative ran on semiconductors: which foundry could ship the most advanced GPUs, and which hyperscaler had the biggest allocation. That race is still on, but the binding constraint has shifted downstream to the substation. AI-optimized server racks draw 30 kilowatts to more than 100 kilowatts each, far above the 5 to 15 kilowatts of a traditional rack, overwhelming local distribution infrastructure never designed for that density.
The scale of the gap is now quantifiable. One industry analysis projects US data center demand reaching 75 gigawatts by 2030 against 58 gigawatts of committed supply - a 17 GW shortfall. A broader estimate, counting edge and enterprise AI workloads beyond hyperscale campuses, puts the gap at 35 GW. Interconnection queues hold a record backlog of proposed generation and load, and wait times for critical grid components such as transformers and cables have doubled over the past three years. Building new transmission in advanced economies takes four to eight years.
The consequence is already visible in the project pipeline. An analysis of roughly 140 US data center projects found at least 16 gigawatts of capacity slated to come online in 2026; only about 5 GW is physically under construction. The remainder sits in the announced stage with no visible construction progress, despite typical build timelines of 12 to 18 months. The bottleneck has migrated from the server rack to the substation, and from there to the voting booth.
The Political Front: Ballot Boxes Against Hyperscalers
Here is the second-order effect that the market is still pricing in: when the grid cannot say yes fast enough, the community gets a vote. In 2026, voters in at least 13 states faced 50 local ballot measures targeting data centers, with 41 of them decided on the November 3 ballot. The measures range from outright construction bans to voter-approval triggers for developments exceeding thresholds such as 20 megawatts of peak demand or 500,000 gallons of water per day. Ohio alone accounted for 21 of the measures across 18 jurisdictions, the single largest concentration of referendum risk.
The early results ran overwhelmingly against developers. A tally of the November votes found 44 communities restricting data center development against just one that allowed construction, with the average pro-restriction vote exceeding 70 percent. In Morgan County, Illinois, 88.3 percent of voters backed limits; in Ashville, Ohio, 81.65 percent voted to repeal a council deal behind a data center. The closest contest, in Sprigg Township, Ohio, still went 52.6 percent for restriction.
Local opposition has moved from nuisance complaints to a coordinated political force. A research group tracking the backlash reported that 45 data center projects worth $68 billion were blocked or delayed by local pushback between April and June alone. The grievances are concrete and hard to dismiss: data centers raise residential electricity bills, consume millions of gallons of water for cooling, run diesel backup generators for monthly maintenance, and produce a continuous low-frequency hum that residents describe as inescapable.
"In many communities, their massive electricity and water consumption has increased residential utility bills. In Michigan and elsewhere, they have already required more fossil fuel plants to be built or stay open, and threaten to derail the transition to clean energy."
The political geometry is awkward for both parties. Republicans, traditionally pro-development, are facing a grassroots backlash in rural and suburban districts where data centers are proposed; Democrats are caught between climate goals and local environmental concerns, since new data center load can keep fossil-fuel plants online and strain renewable-transition timelines. Neither party has a clean answer, which is precisely why the fight will drag on: there is no national consensus to cut through local vetoes.
This is the mechanism that turns a physical shortage into a political one. A transformer can be ordered; a substation can be permitted. But a county electorate that votes down a project does not get overruled by a capital budget. The fight is just beginning because the institutions that must approve each site - planning boards, city councils, utility commissions, state legislatures - were never built for load growth of this speed and concentration.
The Nuclear Bet: Hyperscalers Become Utilities
The market's answer to local opposition and grid queues is to bypass both. Every major US hyperscaler has now signed at least one nuclear deal, and tracking services counted roughly 13 announced projects committing more than 9.8 gigawatts of nuclear capacity to AI data center infrastructure as of 2026. Tech companies are no longer just buying power; they are underwriting the next generation of nuclear development.
The deals split into two waves. The first secures electrons quickly by restarting or contracting existing reactors. Microsoft's agreement to restart Three Mile Island Unit 1 - rebranded the Crane Clean Energy Center - commits 835 megawatts under a 20-year power purchase agreement, with Constellation Energy investing $1.6 billion to bring the reactor back into service, possibly as early as 2027 and with commercial operation now targeted around 2028. Amazon draws power from the Susquehanna plant through Talen Energy. These are the near-term fixes.
The second wave finances technology that will not deliver commercial power until the 2030s. Google has committed to up to 500 megawatts of small modular reactors with Kairos Power. Amazon invested in SMR developer X-energy. Meta has made multi-gigawatt commitments across developers including TerraPower and Oklo. Oklo reports a customer pipeline exceeding 14 gigawatts, including a 1.2 GW nuclear campus in southern Ohio to support Meta's regional data centers - with the first phase targeted as early as 2030 and full deployment by 2034 - plus a master power agreement with data center developer Switch for up to 12 GW and a 500 MW pact with Equinix.
This is a structural change in how new nuclear gets financed. Before 2025, powering data centers with nuclear energy was largely theoretical; now technology giants are signing long-term offtake agreements that give reactor developers the revenue certainty needed to raise capital. The hyperscaler has become the anchor customer the nuclear industry has waited decades for. But the timeline mismatch is the catch: the first nuclear electrons reach an AI data center around 2027, while the SMR fleet arrives in the 2030s. The gap between now and then must be bridged.
The Gas Bridge And The Efficiency Trap
What fills the gap in the meantime is not clean. Because renewables and transmission move slower than AI load, developers are turning to natural gas - often behind the meter, wired directly to the data center rather than the public grid. Analysts estimate data center developers have announced roughly 101 gigawatts of behind-the-meter gas capacity in the United States; more than 57 GW of that has publicly disclosed equipment orders, and about 7 GW is already under construction. Planned gas capacity queued for interconnection between 2026 and 2030 expanded by roughly 24 gigawatts between the start of 2025 and January 2026, reaching a 64 GW pipeline.
The environmental trade-off is direct. One study identified at least 74 gas-fired power plants planned to fuel the data-center industry, emitting 662 million tons of greenhouse gases annually - equal to adding 140 million cars and trucks to the road, or roughly three times California's total power generation. The political backlash and the emissions problem are therefore the same problem: a community asked to host a data center is also being asked to host the fossil generation that keeps it running.
The counterargument from the industry is efficiency. Energy use per AI task has been falling by at least a factor of ten every year, an unprecedented rate in energy history; Google reported its median Gemini text prompt uses 0.24 watt-hours, a 33-fold reduction over a year. But efficiency has not reduced total demand - it has lowered the cost per task and invited more of it. Google's systems processed more than 3.2 quadrillion tokens a month in 2026, up from 480 trillion a year earlier and 9.7 trillion the year before that. This is the Jevons paradox in real time: the cheaper each token becomes, the more tokens the world consumes, and the power bill keeps rising.
Cyclical Squeeze Or Structural Regime Shift
The central question for investors is whether this is a cyclical bottleneck - tight for a few years until transmission and generation catch up - or a structural regime shift in the cost and availability of compute. The evidence points to structural, for three reasons.
First, the demand driver does not mean-revert. AI training and inference loads grow with model capability and adoption; there is no inventory cycle that clears them. Historical data center build-outs followed enterprise IT budgets that expanded and contracted. AI load is different: it is tied to a technology curve that compounds, not to a business cycle that oscillates.
Second, the supply response is institutionally constrained, not just capital-constrained. Capital is abundant - hyperscalers announced hundreds of billions in AI infrastructure spending. What is scarce is permission: interconnection queue positions, transmission rights-of-way, local zoning approval, and reactor licensing. A line of credit cannot buy a four-to-eight-year transmission permitting timeline or reverse a local ballot measure.
Third, the cost structure has changed permanently. Power is moving from an operating expense priced at the marginal grid rate to a strategic input secured through long-term contracts, equity stakes, and outright project ownership. Companies that locked in low-cost, firm, carbon-free power early - through nuclear PPAs or advantaged grid positions - hold a durable cost advantage over rivals that arrive late and pay the scarcity price.
The counter-thesis deserves weight. Skeptics argue that demand will fall faster than expected if AI monetization disappoints, that efficiency gains will cut power per inference, and that load can migrate to regions with surplus capacity or overseas. There is truth in each: not every announced project will be built, and some demand will relocate. But the counter-thesis underestimates the asymmetry. Even if demand growth halves from current projections, the US power system is adding capacity at a pace that leaves little headroom; reliability assessments already assume conservative load growth. And migration has limits - latency-sensitive AI inference must sit near population centers, which are exactly where the grid is most constrained and opposition is strongest.
The falsifying signal is specific: if US data center power demand growth slows to below 5% annually for two consecutive years while grid additions accelerate, the structural-scarcity thesis weakens materially. A more immediate tell is the project pipeline - if the share of planned builds actually under construction rises back above two-thirds within 18 months, the bottleneck is easing rather than tightening.
Who Wins, Who Loses, And What To Watch
Cash in the mechanism and the beneficiaries are clear. Companies positioned in energy infrastructure - regulated utilities with data-center-heavy service territories, independent power producers with nuclear fleets, uranium miners, electrical equipment manufacturers, and cooling-technology providers - face a multi-year demand wave that is largely independent of the AI application cycle. Established nuclear producers and power generators with firm capacity command a scarcity premium; equipment makers with transformer and switchgear backlogs have visibility into the mid-decade.
The exposed are the hyperscalers and AI developers that arrive late to the power race. A company without secured, firm, low-cost electricity faces three bad options: pay a premium for scarce capacity, accept delayed model rollouts, or cede ground to rivals that locked power down first. Compute without electrons is stranded capacity. The second-order implication is that AI competition is becoming an energy-portfolio competition, and the market will increasingly underwrite the two together rather than separately.
Looking ahead, split the timeline in three. In the short term - through 2027 - expect volatility driven by local ballot outcomes, interconnection rulings, and permitting decisions; each veto or approval will move individual names. In the medium term - 2027 to 2030 - the gap between demand and committed supply widens, rewarding firms that secured firm capacity and penalizing those that did not. In the long term, if the current wave of nuclear and transmission commitments materializes, the scarcity premium compresses and the winners are those who owned the assets through the bottleneck.
Scenarios: the base case is a slower, more expensive build-out concentrated in power-rich regions, with nuclear and gas bridging the gap until transmission catches up. The upside case is a policy breakthrough - federal permitting reform or state-level preemption of local vetoes - that accelerates transmission and unlocks queued projects. The downside case is a cascade of local defeats and prolonged licensing delays that strands announced AI capacity and forces hyperscalers to ration compute growth.
The data center fight is just beginning because the constraint has changed nature: it is no longer a supply chain that money can fix, but a permission structure that money alone cannot buy. The companies that win the AI decade may be the ones that won the power fight first.
Explore more exclusive insights at nextfin.ai.
