The explosive growth of AI is creating an unprecedented energy crisis for hyperscalers. Deloitte estimated in June 2025 that US data center power demand could rise from 33 GW in 2024 to 176 GW by 2035, and grid infrastructure is not keeping pace. The biggest strategic question in tech is no longer about chips or models — it is about megawatts. Microsoft, Google, and Amazon must decide whether to build their own power infrastructure, including nuclear, or bet that the grid catches up in time.
The strategic fork
176 GW
Projected US Data Center Demand by 2035
Deloitte estimate (June 2025), up from 33 GW in 2024; AI data centers account for 123 GW of the 2035 total
$300B+
Hyperscaler Capex Plans (2025)
Combined 2025 capital expenditure guidance from Microsoft, Google, Amazon, and Meta, much of it for AI infrastructure (as of February 2025)
Up to 7 years
Grid Connection Wait
Dominion Energy's expected wait, as of August 2024, to connect new Northern Virginia data centers needing more than 100 MW
$1.6B
Nuclear Restart Cost
Constellation's estimate (Q3 2024) of the capital spending needed to restart Three Mile Island Unit 1, whose output Microsoft will buy under a 20-year PPA
The Fork
Go All-In on Bespoke Power
- ●Build behind-the-meter generation (natural gas, solar, battery storage) directly at data center campuses
- ●Invest in small modular nuclear reactors (SMRs) for carbon-free baseload power with 20+ year horizons
- ●Sign long-term PPAs with energy companies and restart mothballed nuclear plants
- ●Acquire land positions near power generation assets and transmission infrastructure
- ●Partner with or acquire energy startups developing next-generation power technologies
Risk
Massive capital lock-in, which Stratrix estimates could exceed $100B across the industry. Potential for stranded assets if AI demand growth plateaus or efficiency gains reduce power needs. Regulatory and construction delays on nuclear projects could push timelines out by years. Public opposition to nuclear siting near population centers.
Wait for Grid Modernization
- ●Lobby aggressively for faster permitting of transmission lines and generation facilities
- ●Distribute AI workloads across multiple geographies to arbitrage available grid capacity
- ●Invest in chip-level and cooling efficiency to reduce power per unit of compute
- ●Sign shorter-term PPAs and maintain capital flexibility for future options
- ●Bet that state incentives and the remaining federal clean-energy credits (curtailed for wind and solar by the July 2025 budget law) will accelerate renewable and grid buildout
Risk
Compute bottlenecks at the critical moment when AI capabilities are scaling fastest. Competitors who solved power first capture enterprise AI market share. Geographic distribution adds latency and operational complexity. Grid modernization timelines are historically unreliable — major transmission projects routinely take many years.
The Power Race: Key Events and Projections
2024
The Nuclear Renaissance Begins
Microsoft signs a 20-year PPA with Constellation Energy to restart Three Mile Island Unit 1. Google agrees to buy up to 500 MW from Kairos Power's planned small modular reactors. Amazon buys a data center campus next to Talen Energy's Susquehanna nuclear plant and invests in SMR developer X-energy.
Q1 2025
Capex Arms Race Escalates
Microsoft, Google, Amazon, and Meta guide to more than $300 billion of combined 2025 capital expenditure, much of it for AI infrastructure. Power sourcing becomes a recurring theme for investors.
Q3 2025
Regulatory and Community Pushback
Local opposition mounts against data center construction in Virginia, Georgia, and Wisconsin. Utilities warn of grid reliability concerns. FERC begins reviewing data center interconnection policy.
2026-2027
First SMR Decisions
Projected timeline for final investment decisions on first commercial SMR deployments at data center campuses. NRC licensing reviews reach critical milestones. Construction timelines become clearer.
2028-2030
The Capacity Reckoning
Projected period when AI demand growth collides with grid reality. Hyperscalers with secured power gain decisive competitive advantage. Those without face deployment delays and potential loss of enterprise AI contracts.
2030-2035
Long-Term Power Landscape
SMRs, if successfully deployed, begin generating baseload power. Grid modernization projects initiated in 2024-2025 start coming online. The energy landscape for AI becomes clearer — winners and losers are evident.
Signal
- ●About 2,061 GW of generation and storage was waiting in US interconnection queues at the end of 2025, down from a peak of nearly 2,600 GW at the end of 2023 (LBNL, 2026)
- ●Dominion Energy said in August 2024 that new Northern Virginia data centers needing more than 100 MW could wait up to seven years for a connection
- ●The IEA estimated in January 2024 that a ChatGPT request used about ten times the electricity of a Google search, though Google's August 2025 measurement of a median Gemini text prompt (0.24 Wh) was far lower
- ●Three major hyperscalers have independently converged on nuclear as part of their power strategy
- ●EPRI projected in May 2024 that US data centers could use 4.6% to 9.1% of US electricity generation by 2030, up from about 4% in 2023
Noise
- ●AI efficiency gains will eliminate the power problem entirely — models always get more efficient
- ●Fusion power is just around the corner and will solve everything
- ●Data center demand projections are wildly overstated — they always are
- ●Hyperscalers can simply move all workloads to countries with cheaper, more available power
- ●On-chip efficiency improvements will reduce total power needs faster than demand grows
The hyperscaler energy gamble qualifies as a genuine strategic fork because it is irreversible, high-stakes, and time-bound. Building nuclear reactors and dedicated power infrastructure requires commitments measured in decades and tens of billions of dollars — capital that cannot easily be redirected if the bet goes wrong. Conversely, choosing to wait is also a commitment: every quarter spent hoping for grid modernization is a quarter where competitors with secured power capacity can deploy AI infrastructure that you cannot. The asymmetry is striking. If AI demand continues on its current trajectory, the companies that secured their own power will have built an almost unassailable competitive moat. If demand plateaus — due to an AI winter, dramatic efficiency gains, or regulatory intervention — those same companies will be saddled with expensive, purpose-built power assets with limited alternative uses. This is not a decision that can be hedged easily. The timelines for nuclear construction, transmission line permitting, and grid interconnection mean that the choices made in 2025 will determine competitive positions well into the 2030s.
Nuclear Regulatory Uncertainty
NRC licensing for small modular reactors remains untested at scale. No commercial SMR has been built in the US. NuScale's cost overruns and the cancellation of its Idaho project in 2023 demonstrate the execution risk.
Community Opposition to Data Centers
Growing NIMBY resistance in Virginia, Georgia, and other data center hubs is slowing permitting. Concerns about water usage, noise, aesthetic impact, and strain on local power grids are fueling organized opposition.
Transmission Line Permitting Gridlock
High-voltage transmission lines routinely take years to permit and build in the US. Even with bipartisan support for grid modernization, permitting reform moves slowly through Congress and state legislatures.
Capital Allocation Pressure from Investors
Spending tens of billions of dollars a year each on AI infrastructure requires sustained investor confidence. Any sign that AI revenue growth is decelerating could trigger pressure to cut capex, stranding partially built power projects.
Geopolitical Energy Competition
China, the Middle East, and other regions are aggressively building AI data center capacity with fewer regulatory constraints. If power bottlenecks slow US deployment, AI leadership could shift geographically.
Inside the War Room
Microsoft's Three Mile Island Bet
In September 2024, Microsoft signed a 20-year power purchase agreement with Constellation Energy to restart the Three Mile Island Unit 1 reactor — a facility that had been shut down since 2019. Constellation estimated the restart would require about $1.6 billion of capital spending; the value of the PPA itself was not disclosed. It was a high-profile case of a tech company's power purchase underpinning the restart of a shuttered nuclear plant. The symbolic weight of the Three Mile Island name added to the boldness of the move.
Amazon's Talen Energy Nuclear Play
In March 2024, Amazon bought a data center campus directly adjacent to Talen Energy's Susquehanna nuclear power plant in Pennsylvania, planning to draw nuclear power directly. The deal drew scrutiny from neighboring utilities concerned about power diversion from the broader grid, and in November 2024 FERC rejected an amended agreement that would have expanded the co-located supply. In June 2025 the companies restructured the deal into a grid-delivered power purchase agreement for 1,920 MW through 2042.
Google's Kairos Power Agreement
In October 2024, Google agreed to buy up to 500 MW from small modular reactors that nuclear startup Kairos Power plans to build, with the first unit targeted for 2030 and the rest by 2035; financial terms were not disclosed. In Stratrix's reading, the deal signaled that all three major cloud providers had independently concluded that the grid alone could not meet their AI power needs.
The Northern Virginia Bottleneck
Dominion Energy in Virginia — home to the world's largest concentration of data centers — said in August 2024 that new data centers needing more than 100 MW could wait up to seven years for a connection. The bottleneck has forced hyperscalers to explore alternative geographies and self-generation options.
Projected Outcomes
If Path A Wins
Hyperscalers with dedicated power achieve 2-3 year deployment advantages over grid-dependent competitors
Nuclear-powered data centers provide carbon-free baseload power, strengthening ESG narratives and regulatory positioning
Capital commitments that Stratrix estimates at $100B+ create high barriers to entry, consolidating the cloud market around 3-4 players
If AI demand plateaus, companies face decades of obligations on underutilized power assets — potential write-downs in the tens of billions
Successful SMR deployments catalyze a broader nuclear renaissance, reshaping the US energy landscape beyond tech
If Path B Wins
Preserved capital flexibility allows rapid reallocation if AI economics shift or new technologies emerge
Compute bottlenecks in 2027-2030 cede AI deployment leadership to competitors who secured power early
Geographic distribution of workloads increases operational complexity and latency for enterprise customers
Grid modernization, if it materializes on time, provides cheaper power than bespoke generation — but historical precedent suggests it will not
Remaining federal and state incentives may accelerate renewable buildout, partially closing the gap, though the July 2025 budget law requires wind and solar projects that start construction after July 4, 2026 to be in service by the end of 2027 to claim credits
“The hyperscaler energy gamble is the defining infrastructure decision of the AI era. The companies making the boldest power commitments today are placing a bet that AI demand will not only persist but accelerate. If they are right, power becomes the ultimate moat — more durable than chip supply, model quality, or developer ecosystems. If they are wrong, they will have built the most expensive stranded assets in corporate history. The signal pattern favors Path A: three independent hyperscalers converging on nuclear, grid constraints worsening, and AI demand projections being revised upward. But the execution risk is extreme, and the timeline for nuclear is measured in decades, not quarters.”
Open Strategic Decision
The 'Power as Moat' Pattern
The hyperscaler energy gamble reveals a recurring pattern in technology history: when a resource constraint becomes binding, the companies that vertically integrate to secure that resource gain a durable competitive advantage. Just as Standard Oil controlled refining, and TSMC controls advanced chip fabrication, the hyperscaler that controls its own power supply may control the pace of AI deployment. The strategic lesson is that competitive advantage often migrates to whoever solves the binding constraint — and in 2025, the binding constraint for AI is not algorithms, not data, and not chips. It is electricity.
The decisive moment
The AI revolution has a power problem. As hyperscalers race to deploy ever-larger language models and inference infrastructure, they are slamming into a constraint that no amount of software optimization can solve: electricity. US data center power demand, about 33 GW in 2024, could reach 176 GW by 2035, according to Deloitte estimates published in June 2025, with AI data centers accounting for 123 GW of that total. The existing grid, built for a slower era of incremental demand growth, simply cannot deliver power at this pace.
This is not a hypothetical future problem. In Northern Virginia — home to the densest cluster of data centers on Earth — utilities are already imposing multi-year wait times for new grid connections. In August 2024, Dominion Energy said new data centers needing more than 100 MW could wait up to seven years for a connection. Similar bottlenecks are emerging in Texas, Ohio, and the Pacific Northwest. The hyperscalers are finding that the hardest part of building an AI data center is no longer sourcing GPUs from Nvidia — it is sourcing reliable electricity.
The strategic fork is stark. Path A calls for massive, vertically integrated power investments: behind-the-meter natural gas plants, long-term power purchase agreements, and — most ambitiously — small modular nuclear reactors that could deliver carbon-free baseload power directly to data center campuses. Microsoft has already signed a deal to restart a unit at Three Mile Island. Google has agreed to buy power from Kairos Power's planned reactors, and Amazon has invested in reactor developer X-energy. The capital commitments are staggering; Stratrix estimates they could exceed $100 billion across the industry over the next decade.
Path B takes a more measured approach: lobby for faster grid permitting, distribute workloads across geographies to arbitrage available capacity, invest in efficiency gains, and bet that grid modernization and renewable buildout will eventually close the gap. This path preserves capital flexibility but risks compute bottlenecks at the worst possible moment — just as AI capabilities are scaling and enterprise adoption is accelerating.
The stakes extend beyond any single company. Whichever path the hyperscalers choose will reshape energy markets, influence climate policy, and potentially determine which nation leads in AI deployment. Power is the new moat.
Apply the lessons
A framework for assessing whether to build dedicated power generation or rely on grid modernization for AI data center deployment.
Audit your power exposure
Map your organization's current and projected compute needs against available grid capacity in each geography. Identify where power constraints will bind first and estimate the timeline to bottleneck.
Evaluate vertical integration options
Assess the feasibility and cost of behind-the-meter generation, long-term PPAs, and nuclear partnerships. Compare the total cost of ownership for bespoke power versus grid dependence over 10-20 year horizons.
Stress-test against demand scenarios
Model your power strategy against three scenarios: AI demand accelerates beyond projections, demand follows consensus estimates, and demand plateaus due to efficiency gains or market saturation. Identify the strategy that performs acceptably across all three.
Build optionality into commitments
Structure power investments with embedded options where possible: modular deployments, interruptible PPAs, dual-use infrastructure. Avoid irreversible commitments until key uncertainties (SMR licensing, grid reform, AI demand trajectory) resolve.
Frequently asked questions
Why are AI data centers consuming so much more power than traditional data centers?+
Why are hyperscalers considering nuclear power for data centers?+
What happens if AI demand growth slows significantly?+
How does the US grid constraint compare to other countries?+
AI data centers are slamming into a wall that software can't fix: electricity. The grid can't keep up.
Do you build your own dedicated power?
NVIDIA's AI pivot
The compute side of the same AI-scale story.
More Strategic Forks
Other consequential decisions worth studying.
The Sovereign AI Infrastructure Decision
Europe's regulatory convergence — NIS 2, the Cyber Resilience Act, and the AI Act, alongside the Commission's Cloud Sovereignty Framework for cloud procurement — is forcing the world's largest cloud providers into a binary infrastructure decision. Build sovereign-stack variants that satisfy every requirement but fragment your platform and erode margins, or maintain global unity with compliance add-ons and risk being locked out of government and regulated-industry contracts. The clock is ticking: member states had to transpose NIS 2 by 17 October 2024, and under Regulation (EU) 2026/1744 the AI Act's high-risk obligations now apply from 2 December 2027 (2 August 2028 for AI in regulated products).
Current ForksEurope's Rearmament vs. Industrial Policy Trade-off
Europe faces a major industrial policy fork. Defense spending commitments are surging past 2% of GDP, and at the June 2025 Hague summit NATO allies agreed a target of 5% of GDP by 2035 (3.5% for core defense plus 1.5% for related spending). But every euro allocated to tank production is a euro not spent on battery gigafactories or hydrogen infrastructure. For companies like Rheinmetall, BAE Systems, and Airbus Defence, the strategic question is pivotal: which bet defines the next decade?
Current ForksThe Global South AI Adoption Fast Lane
While Europe builds regulatory frameworks and the US debates safety guardrails, India and Southeast Asia are pursuing aggressive AI deployment through fast lanes for innovation. Microsoft pledged in February 2024 to give 2 million people in India AI skills by 2025, and in January 2025 raised the goal to 10 million people by 2030. Abu Dhabi's Government Digital Strategy 2025–2027, announced in January 2025, aims to make it the world's first fully AI-native government by 2027. Organizations focused exclusively on Western markets may find themselves outpaced by competitors building capabilities in the world's fastest-growing regions.
Current ForksThe Global South Critical Minerals Bargain
The scramble for critical minerals has handed resource-rich Global South countries their strongest negotiating position in decades. Indonesia's nickel ore export ban, Chile's push for state control of its strategic lithium salt flats, and the DRC's 2025 cobalt export ban and quotas all point to the same question: should these nations use their geological leverage to force industrialization, or accept pragmatic partnerships that keep the minerals flowing?
Current ForksThe Autonomous Vehicle Liability Threshold
Autonomous vehicle companies are racing toward mass deployment with liability frameworks still fragmented and insurance models untested. In August 2025, a Miami jury held Tesla partly liable for a fatal crash involving its Autopilot driver-assistance system. The industry's response to its first landmark fatality case against a fully driverless service will set precedent for decades — and the choice between aggressive legal defense and collaborative regulatory engagement could determine whether autonomous driving reaches mainstream adoption or stalls in legal limbo.
Current ForksThe Continuous Planning Transformation
The annual, static budget is obsolete before the ink dries. Forward-thinking finance organizations are moving to continuous planning — shifting from describing the past to prescribing the future. Yet in FP&A Trends' 2024 survey, 57% of organizations still took one to three months to prepare a budget, and 53% needed more than five days to produce a forecast. The question isn't whether continuous planning is better. It's whether organizations can survive the cultural upheaval required to get there.
More from Stratrix
Test a decision, free
Set a live decision against the companies that made it before you.
AcquisitionsThe Distance TestDivestituresThe Divestiture TestDiversificationThe Diversification TestMake or buyThe Make-or-Buy TestAll Access
Every Casebook, three Strategy Packs a seat each year and the Analyst, for teams that face these decisions all year.
$1,999 a year · $499 a seatWhat is included →New analyses, by email
The defining moves in business, checked against the record, as they publish.
Sources & further reading
- International Energy Agency (2024). Electricity 2024: Analysis and Forecast to 2026. IEA Publications.
- Electric Power Research Institute (2024). Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption. EPRI.
- Deloitte (2025). Deloitte: AI Data Center Power Demand Could Surge 30x by 2035, Amid Power and Grid Capacity Constraints. PR Newswire (republished by BOE Report).
- Lawrence Berkeley National Laboratory (2026). Queued Up: 2026 Edition, Characteristics of Power Plants Seeking Transmission Interconnection As of the End of 2025. Lawrence Berkeley National Laboratory.
- Constellation Energy Corporation (2025). Annual Report on Form 10-K for the fiscal year ended December 31, 2024. U.S. Securities and Exchange Commission.
- Talen Energy Corporation (2025). Talen Energy Expands Nuclear Energy Relationship with Amazon (Form 8-K, Exhibit 99.1). U.S. Securities and Exchange Commission.
- Bloomberg News (2024). Data Centers Face Seven-Year Wait for Power Hookups in Virginia. Data Center Knowledge.
- Kairos Power (2024). Google and Kairos Power Partner to Deploy 500 MW of Clean Electricity Generation. Kairos Power.
- X-energy (2024). Amazon Invests in X-energy to Support Advanced Small Modular Nuclear Reactors and Expand Carbon-Free Power. X-energy.
- Casey Crownhart (2025). In a first, Google has released data on how much energy an AI prompt uses. MIT Technology Review.
- Pierce Atwood (2025). Congress Phases Out Energy Tax Credits. Pierce Atwood.
- Gulf Daily News (2025). Tech megacaps plan to spend more than $300bn in 2025 as AI race intensifies. Gulf Daily News.
Cite this analysis
Stratrix. (2026). The Hyperscaler Energy Gamble. Strategic Forks. Retrieved from https://www.stratrix.com/strategic-forks/hyperscaler-energy-gamble
From the fork to the next read.
Study the strategic fork, understand the decision, then follow the thread across the companies and lenses it connects to.