Big Tech’s hyperscalers are racing to build gigawatt-scale AI data centers, with five facilities set to exceed 1GW by 2026, each requiring a nuclear reactor’s worth of power.
But infrastructure reality is catching up with AI ambitions: a projected 19GW gap between new data center demand and available grid capacity threatens to deflate the AI bubble.
The structural bottleneck isn’t computing, it’s electrons. Companies that solve the power equation will dominate the AI era; those that don’t will watch their chips sit idle.
In cutting-edge Microsoft data centers, racks of chips used to train AI models sit idle. “The biggest issue we are now having is not a compute glut, but it’s power,” said Microsoft’s chief executive Satya Nadella during a recent podcast interview.
The AI data center isn’t just bigger—it’s a different species. Traditional data centers operated in the 50-100MW range, running thousands of independent workloads. Today’s AI training clusters demand 1GW+ of synchronized power, with racks of GPUs working in concert on single model training runs. This isn’t incremental growth; it’s an architectural revolution.
The topic has been top of mind in a year when big tech “hyperscalers”—Amazon, Google, Meta and Microsoft—have set out plans to spend more than $400 billion in capital expenditure. That outlay, mainly on data centers, has triggered market fears of an AI-fuelled bubble. Big tech groups have been undeterred, arguing that demand outstrips supply.
The Numbers That Matter
The scale of what’s being attempted—and what’s missing—can be captured in five figures:
Current US data center capacity: ~51GW combined, representing 5% of peak national demand. This is the baseline from which the industry is attempting to scale.
New capacity required by 2028: 44GW additional, according to S&P Global Energy. This represents nearly doubling the existing infrastructure in just three years.
Grid capacity coming online: ~25GW—only 57% of projected demand. Given constraints to grid infrastructure, power capacity coming online in the next three years will fall dramatically short.
The gap: 19GW—just over 40% of the power needed. This is equivalent to 19 nuclear reactors worth of missing power, or roughly the entire generating capacity of a mid-sized European country.
Global electricity demand from data centers by 2030: ~945 TWh according to IEA projections, representing 3% of global consumption. AI-optimized servers alone will see electricity usage rise nearly fivefold, from 93 TWh in 2025 to 432 TWh by 2030.
The Five Giants: First Gigawatt Facilities (2026)
In 2026, five US data centers—each from a different company—are set to become the first globally to use more than 1GW at their peak, according to research group Epoch AI. These facilities mark a threshold crossing in AI infrastructure:
1. Anthropic–Amazon New Carlisle (January 2026) The facility has been operational since October 2025 with approximately 500,000 AWS Trainium 2 chips already installed. An additional 500,000 chips are slated for installation by year-end 2025, doubling capacity ahead of the power ramp-up. One million Trainium 2 chips represent 500 MW of chip power and computing equivalence of roughly 300,000 H100s.
2. xAI Colossus 2 (February 2026)Elon Musk’s facility will deliver equivalent computing power of 1.4 million H100 GPUs. The original Colossus became the world’s largest single AI training cluster just four months after construction began in 2024, running over 200,000 GPUs in coherent training. Colossus 2 represents a further massive scale-up.
3. Microsoft Fayetteville (March 2026) A borderline 1GW facility that’s part of Microsoft’s aggressive expansion. The company’s Fairwater facility in Wisconsin could exceed 3GW by 2028—one of several ambitious projects in the pipeline that would make it among the largest single computing installations ever constructed.
4. Meta Prometheus (May 2026) Located in Ohio, Meta’s cluster includes plans for 200MW of behind-the-meter generation—power produced onsite that bypasses the traditional grid entirely. This approach reflects the industry’s growing willingness to build private power infrastructure alongside computing facilities.
5. OpenAI Stargate Abilene (July 2026) The Texas facility is set to include 10 natural gas turbines with a capacity of 361MW for backup power. OpenAI alone has signed infrastructure deals totalling more than $1.4 trillion, amounting to an estimated 28GW in capacity over the next eight years. Chief executive Sam Altman has characterized the power crunch as existential: “A certain risk is if we don’t have the compute, we will not be able to generate the revenue or make the models at this kind of scale.”
The Infrastructure Bottleneck
The US power grid was built for a different era. After more than two decades of flat or anaemic growth, US power demand is now surging. Electricity usage is projected to rise by an average of 5.7% per year to 2030, based on forecasts from utility companies.
Though some of this demand is due to reshoring activity and a broader shift to electrifying buildings and transport, more than half of the expected increase stems from the rapid build-out of AI data centers, according to consultancy Grid Strategies.
But boosting the US power grid is an enormous and time-consuming task due to a complex web of regulatory, financial, and supply chain challenges.
The Chokepoints
Interconnection queues have become the primary bottleneck. These backlogs of projects waiting to plug into the network are slowing the rollout of new power capacity and leaving data centers facing lengthy delays. PJM, the largest grid operator in the US and home to “data center alley” in Virginia, is under particular strain. The average time from filing an interconnection request to achieving commercial operation now exceeds eight years, according to energy think-tank RMI.
Phantom data centers compound the problem. Nationally, developers approach multiple utilities in search of the lowest price, creating duplicate proposals that further bloat queues and make it harder for utilities to prepare for future demand. “There isn’t a significant barrier to entry [for speculative builders],” said Bobby Hollis, vice-president of energy at Microsoft. “So one of the biggest challenges is finding out how much demand is real. There are a lot of participants who don’t know what goes into building a data centre… [And] few opportunities to filter out the noise.”
Transmission lines represent another critical constraint. On average, federal permitting for a new US transmission line takes about four years, according to the Department of Energy. State processes add further delays. Last year, almost 900 miles of new high-voltage transmission lines were completed—the most since 2020, but still far short of the 5,000 miles per year that Americans for a Clean Energy Grid estimates is needed to support grid reliability and growth.
Transformer shortages are driving up prices and delaying projects. Anthony Allard, US managing director at Hitachi, a major supplier of grid equipment, said transformer lead times were “three to four times” longer than in 2020. The equipment that transfers electrical energy between circuits has become a critical bottleneck.
Gas turbine delivery times have more than doubled since 2023 to about four and a half years. As a result, building new gas capacity now costs about $2,400 per kilowatt—up 71% in four years, according to energy research group Wood Mackenzie. This matters because the International Energy Agency projects that most new US data center demand will be met by gas over the next decade.
The Geographic Concentration Problem
Virginia hosts “data center alley”—the densest cluster of operational data centers in the US. But concentration creates vulnerability. PJM’s grid is under particular strain, and the gap between new data center demand and spare grid capacity is growing across almost every region.
New gigawatt-scale projects are increasingly moving to southern states—Texas, Georgia, Tennessee—seeking power availability. The tradeoff: less established infrastructure, higher regulatory friction, and growing community opposition. Multiple gigawatt-scale projects are now planned in these southern states, representing a geographic shift in AI infrastructure development.
The satellite imagery tells the story. Tech groups have rapidly expanded key sites over the past 18 months: Amazon’s Project Rainier in Indiana, OpenAI’s Stargate in Texas, and xAI’s Colossus 2 in Tennessee have all transformed from open land to massive industrial facilities in the span of months, not years.
The Race for Power: Workarounds and Controversies
Grid setbacks have led hyperscalers and model builders to seek alternative power solutions, including building capacity outside the traditional grid. When the grid can’t keep pace, hyperscalers improvise—sometimes controversially.
Behind-the-Meter Generation
Companies are building their own power plants adjacent to data centers, bypassing grid interconnection entirely. This approach—sometimes called “behind-the-meter” generation—represents a fundamental shift in how AI infrastructure is being powered.
xAI Colossus (Memphis): The most controversial example. xAI operated its Colossus data center cluster in Memphis, Tennessee, with dozens of gas turbines for much of the past year without environmental permits, according to the Southern Environmental Law Center. The group has accused Elon Musk’s company of being the state’s largest “industrial source” of nitrogen oxide pollution and misleading residents over the number of turbines in operation at the site.
The company received a permit in July that enabled it to run 15 turbines as a form of back-up generation, though SELC said it has observed as many as 35 turbines on site. The facility sits near historically Black neighborhoods in Southwest Memphis, and community groups have raised concerns about air quality impacts. Aerial photography from the Southern Environmental Law Center shows rows of gas turbines installed adjacent to the data center building, with cooling infrastructure visible alongside.
OpenAI Stargate (Abilene): Planning documents indicate the Texas facility will include 10 natural gas turbines with a capacity of 361MW—substantial onsite generation that reduces dependence on grid interconnection timelines.
Meta Prometheus (Ohio): Planning documents outline Meta’s goal to add 200MW of behind-the-meter generation to its cluster, following the same pattern of private power infrastructure.
Analysts expect that this type of activity will persist while the network is unable to meet data center demand. The speed advantage is significant: building onsite generation can bypass years of interconnection queue delays, even if it introduces environmental and regulatory complications.
Nuclear Revival
There is also a drive to reopen some nuclear plants that were mothballed in the past five years. Tech companies are turning to nuclear as the only scalable zero-carbon baseload option capable of meeting gigawatt-scale demand.
Three Mile Island restart: Following a deal with Microsoft, Constellation is planning to restart the Three Mile Island nuclear plant in Pennsylvania from 2027 to address capacity shortages. The plant—site of America’s worst nuclear accident in 1979—would be recommissioned specifically to power AI infrastructure. The cooling towers, control room, and turbine building are being prepared for restart.
Duane Arnold (Iowa): Plans are advancing to revive this retired nuclear plant for data center power, following the same model of dedicated nuclear capacity for AI workloads.
SMR development: More than 80 commercial small modular reactor designs are currently being developed globally, according to the Institute for Progress. These miniaturized fission reactors could eventually provide dedicated power for data centers, though first commercial deployments are unlikely before 2030 and face significant regulatory and financing hurdles.
Demand Response: A Temporary Bridge
Many utility companies are pinning their short-term hopes on “demand response” solutions that require companies to curtail activity at peak times.
AI model builders typically run data centers at full capacity during “training runs”—where they feed large language models with vast amounts of data to improve accuracy. These rises in activity can clash with consumption from other customers—including households—during peak usage, increasing the risk of blackouts.
“We have to get smarter about using unused capacity on the grid,” said Daniel Eggers, executive vice-president at Constellation, a power company that supplies 2 million US homes and businesses.
Researchers at Duke University said earlier this year that if data center operators could restrict their consumption 0.25% of the time, the grid could accommodate about 76GW of additional demand. They cautioned that this would not replace the need to build new capacity.
Brandon Oyer, head of energy and water for the Americas at Amazon Web Services, said the company could tolerate some curtailment on a temporary basis, but did not consider it a “smart investment” to do so for a prolonged period of time. “Some customers might be able to tolerate that. Some customers might not. It’s going to be a very nuanced decision.”
The Hidden Resource: Water
Power isn’t the only constraint. Data centers consume vast quantities of water for cooling—and most new facilities are being built in water-stressed regions.
The Scale of Water Consumption
Direct US consumption (2023): Approximately 17.5 billion gallons, according to Lawrence Berkeley National Laboratory. This represents the water consumed onsite for cooling operations.
Projected growth: US data center direct water consumption could double or even quadruple the 2023 level by 2028, as AI workloads generate substantially more heat than traditional computing.
Location problem: About two-thirds of data centers built or in development since 2022 are in places already gripped by high levels of water stress, according to Bloomberg analysis. Five states alone account for 72% of new centers in high-stress areas.
Global AI water demand by 2027: 4.2-6.6 billion cubic meters—equivalent to 4-6 times Denmark’s annual water withdrawal. This represents just the AI-specific portion of data center water use.
Training impact: Research has shown that training the GPT-3 language model in Microsoft’s US data centers directly evaporated approximately 700,000 liters of clean freshwater. Larger models require proportionally more.
Per-facility scale: A 1 megawatt data center can use up to 25.5 million liters of water annually just for cooling—equivalent to the daily water consumption of approximately 300,000 people.
The Cooling Trade-Off
Just as human bodies cool themselves by sweating, data centers are often cooled by water evaporation—a process that dissipates heat and results in water being lost to the atmosphere. In many cases, the water is drawn from the same municipal systems that supply homes and businesses.
Operators face a fundamental choice: water-cooled systems are approximately 10% more energy-efficient but consume significant freshwater. Liquid cooling (immersion or direct-to-chip) reduces water use but increases power consumption. The tradeoff is unavoidable—optimizing for one constraint worsens the other.
The regions with the most available renewable energy resources to support data centers—particularly solar—are sometimes the ones with the least water. Conversely, data centers set up to use less water in hotter regions end up requiring more power to run. This creates impossible optimization problems for facility planners.
Microsoft has developed closed-loop cooling designs that recirculate water without evaporation—potentially resolving the trade-off at higher capital cost. These designs will be deployed first in facilities in Wisconsin and Arizona, planned for 2026. Crusoe Energy Systems, a developer behind OpenAI’s Stargate site in Abilene, also plans to use closed-loop cooling systems.
Community Backlash
The scale of investment required to meet rising demand is stoking fears of higher utility bills and spurring locals already concerned about water usage and the impact data centers could have on their health. The backlash has forced Amazon, Google, and Microsoft to scrap plans for facilities this year in states including Minnesota and Wisconsin.
Data centers’ freshwater consumption in Loudoun County, Virginia, has increased by 250% since 2019, resulting in drought-like conditions for residents and leading to public calls for more industry oversight. In Newton County, Georgia, some proposed data centers have reportedly requested more water per day than the entire county uses daily.
The Geopolitical Dimension
AI infrastructure is now a matter of national competitiveness. Silicon Valley has cast its wager on AI as both a source of future economic gains and a matter of national importance. Tech companies say they are racing against China to build artificial general intelligence—systems that surpass human abilities.
The US-China Compute Race
The power disparity between US and Chinese infrastructure investment is stark:
China’s 2024 power additions: 429GW—more than one-third of the entire US grid, and more than half of all global electricity growth. In October, OpenAI wrote an open letter to the US government noting this figure and urging it to set an ambitious target of building 100GW per year of new capacity.
US 2024 additions: 51GW, representing just 12% of global growth. The gap in infrastructure investment capacity represents a structural disadvantage in the AI race that software innovation cannot overcome.
OpenAI’s ask: 100GW per year of new US capacity—nearly double what was added in 2024 and a pace the current regulatory and supply chain environment cannot support.
While constrained by US curbs on accessing Nvidia’s most advanced AI chips, Chinese developers have made strides. DeepSeek released its powerful R1 model in January, comparable with US rivals despite being developed at a fraction of the cost and computing power. Chinese companies also receive subsidies to offset the impact of trade barriers.
The Policy Response
The White House has responded with a sense of urgency. US President Donald Trump set out an action plan in July to “win the AI race,” adding that “from this day forward it’ll be a policy of the United States to do whatever it takes to lead the world in artificial intelligence.”
The Trump administration has sought to ease delays by fast-tracking environmental assessments and other permits. It has also invoked emergency powers to delay the planned retirement of older, typically coal-fired plants.
The president’s moves against renewable energy projects have been challenged in court by environmental groups, who warn that they undermine efforts to deploy some of the cheapest and quickest ways to add capacity. “Our biggest concern is ending the attacks on solar and batteries… if you’re forfeiting the energy race to China, then you’re forfeiting the AI race to China,” said Jesse Lee, a senior adviser at Climate Power, an advocacy group. “You can’t win that race if you’re trying to restrict wind, solar and batteries. There’s a five- to seven-year wait time for natural gas turbines right now. That’s just not an option.”
The Stargate Project
The Stargate project—a $500 billion, 10GW joint venture between OpenAI, SoftBank, Oracle, and MGX—represents the clearest articulation of AI as national infrastructure. President Trump announced the venture at a White House press conference in January 2025, calling it “the largest AI infrastructure project in history.”
Five sites have been announced, bringing Stargate to nearly 7 gigawatts of planned capacity and over $400 billion in investment over the next three years:
Shackelford County, Texas
Doña Ana County, New Mexico
Lordstown, Ohio
Milam County, Texas
A midwestern site not yet publicly disclosed
Whether this “Manhattan Project for AI” delivers depends entirely on solving the power equation. The Information reported that the project had not started and no funds were raised to meet the project’s initial $500 billion budget as of August, with market uncertainty, American trade policy, and AI hardware valuations causing delays.
The Key Insight
The concern for hyperscalers is that this patchwork of measures will not be enough to power data centers coming online over the next few years. In this scenario, a raft of projects will no longer be viable because they cannot meet contractual commitments.
Others will have to simply wait for upgrades to the electricity grid and the construction of new generation capacity to be completed.
In a race between global superpowers, AI could be slowed down by decades-old grid infrastructure and a failure to provide adequate capacity. For some, the power crunch eases concerns of overbuild. For tech companies and the Trump administration, it may undermine billions of dollars in investment.
“We may not get all this done in the timeframe that hyperscalers would like… and they won’t be able to interconnect until we’ve got the resources to meet them,” said Jim Robb, chief executive of the North American Electric Reliability Corporation. “It’s going to be a white-knuckle ride.”
The AI revolution’s limiting factor isn’t algorithms or chips—it’s infrastructure. The 19GW gap between demand and available power represents the largest constraint on AI scaling today. Companies racing to train frontier models are discovering that the bottleneck isn’t compute—it’s electrons.
For strategic operators, this means the AI winners of 2030 are being determined now—not by who has the best models, but by who has the best power supply. For policymakers, it means AI leadership is fundamentally an energy policy question. And for everyone else, it means the white-knuckle ride is just beginning.
The question isn’t whether AI will transform the economy. It’s whether the grid will let it.
Recap: In This Issue!
The AI boom is hitting a physical ceiling. A 19GW power shortfall by 2028 threatens to strand trillions in AI infrastructure.
Gigawatt-scale data centers are redefining hyperscaler strategy. Power access, not compute access, is becoming the dominant competitive moat.
AI leadership is collapsing into energy policy. Nations and companies that can deliver electrons at scale will control the next decade of AI.
The Power Constraint That Will Define AI
AI progress is no longer limited by chips. It is limited by electricity. Hyperscalers are planning gigawatt-scale campuses that require the output of full nuclear reactors, but the US grid cannot supply the required power fast enough. A 19GW gap between demand and available capacity is forming the hard ceiling on AI scale.
The Numbers That Matter Most
US data center capacity today: 51GW
New demand by 2028: 44GW
New supply coming online: 25GW
Power gap: 19GW
Global data center consumption by 2030: 945 TWh
AI server electricity usage growth: from 93 TWh in 2025 to 432 TWh in 2030
This is the structural constraint behind every AI roadmap.
The First Gigawatt Data Centers (2026)
Five sites will cross 1GW in 2026, signalling a point of no return in AI infrastructure design.
Two thirds of new facilities are in water-stressed regions
A 1MW data center can consume 25.5 million liters of water annually
Water efficiency and power efficiency now move in opposite directions, forcing structural tradeoffs.
The Geopolitical Dimension
AI competitiveness is now energy competitiveness.
China added 429GW of new power capacity in 2024.
The US added 51GW.
OpenAI is calling for 100GW per year of new US capacity just to stay competitive.
China is scaling infrastructure faster than US permitting allows.
Software innovation cannot compensate for power imbalance. The AI race is becoming an infrastructure race.
Stargate: The Manhattan Project for AI
The OpenAI–SoftBank–Oracle–MGX venture targets ten gigawatts of capacity. Five sites are public, representing more than four hundred billion dollars. None can proceed without guaranteed baseload power. The entire venture depends on solving the power bottleneck first.
Closing Takeaway
The constraint on AI is no longer model quality or chip supply. It is power. Without solving the 19GW gap, the AI supercycle slows down. With it, AI accelerates into a new infrastructural era defined by energy abundance, not algorithmic breakthroughs.
The winners of the 2030s will be the players with guaranteed electrons. The question is not whether AI can scale. It is whether the grid will allow it.
References
Financial Times “The power crunch threatening America’s AI ambitions” (December 2025)
margin: 0 0 12px;">Data Center Buildouts as Capital-Intensive Moat Construction
margin: 0 0 16px;">Through the BIA lens, the AI data center race is fundamentally about capital expenditure as a competitive barrier to entry. The mental model of scale economics reveals that hyperscalers investing tens of billions in purpose-built AI infrastructure are creating physical moats that no startup can replicate — the same dynamic that made telecom networks and railroad systems natural monopolies. Layer 5 strategic mapping shows the critical insight: the companies controlling data center capacity today are positioning themselves as the ‘picks and shovels’ layer of the AI gold rush, capturing value regardless of which models or applications ultimately win.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
