The Grid Won't Scale—Long Utilities with Locked Power, Fade the Hyperscaler Narrative

The constraint no one is pricing

Only 12 GW out of 24 GW of U.S. data center capacity scheduled for 2026 is currently under construction, according to financial analyst Jefferies, and the situation is worse for 2027–2028: as much as 80% of planned capacity has not yet commenced construction. The reasons are familiar—zoning delays, permitting backlogs, labor shortages—but one constraint dominates: electricity access. Hyperscalers are guiding to $345 billion in combined 2025 capex, the majority earmarked for AI infrastructure, yet management teams at Microsoft, Alphabet, Amazon, and Meta now explicitly state that power availability and grid interconnection timelines, not chip supply, are the primary limit on how fast that capital translates into usable capacity.

The mispricing is structural. Equity markets still price hyperscaler capex as if power is abundant and fungible, when in reality the U.S. electric grid adds only 20 GW of net new generation per year—a pace set by transmission planning cycles and equipment supply chains that predate the AI boom—while interconnection queues across U.S. RTOs and ISOs now hold more than 2,200 GW of generation and storage projects with median wait times stretching to 4–6 years. Companies with secured power contracts and energized capacity are trading at valuations that do not reflect the scarcity premium they will command as the gap between planned AI deployment and available electricity widens.

Why every AI forecast assumes a grid that doesn't exist

Every bullish AI forecast—every hyperscaler earnings model projecting smooth exponential scaling through the late 2020s—assumes electricity will be there when needed. It won't. Training a single large language model consumes as much electricity as several thousand homes use in a year, and inference at scale requires always-on, high-utilization compute operating 24/7 at near-peak load. Hyperscalers are now planning individual AI campuses in the multi-hundred-megawatt to multi-gigawatt range, each equivalent to a mid-sized city's electricity demand.

The supply side cannot keep pace. Transformer manufacturing capacity is short by roughly 30%, with lead times stretching to 2.5 years. Permitting and rights-of-way approvals for new transmission and generation projects routinely take multiple years, and local opposition frequently delays or blocks projects even when demand and financing are in place. For data centers specifically, utilities' risk-averse approaches to serving large, high-utilization loads—combined with the need to ensure firm capacity without shifting costs to residential ratepayers—further extend timelines and raise the bar for contract approval.

The result: hyperscalers can write checks for chips and buildings in quarters, but securing and energizing multi-hundred-megawatt power connections takes years. Jefferies highlights that duplicative counting—hyperscalers making multiple requests to various energy utilities—has inflated planned capacity totals, and the firm does not expect the majority of extra load forecast for 2026 and 2027–2028 to materialize, suggesting 15–20 GW per year is more realistic than the 40+ GW some have forecast.

What hyperscalers are actually doing

Microsoft has committed to a multi-billion-dollar deal to restart Three Mile Island Unit 1 exclusively for its data centers, a 20-year power purchase agreement (PPA, a long-term contract locking in electricity supply and price) that underscores the company's view that securing firm, carbon-free baseload power is now a strategic imperative. Amazon's AWS has stated publicly that power is its "single biggest constraint," with CEO Andy Jassy explaining that AWS currently has more demand than capacity and that it will take "several quarters" to catch up, largely because of power limits. Google has signed a 20-year PPA with AES for co-located generation in Texas and a deal with Xcel Energy in Minnesota under which Xcel adds 1,900 MW of clean energy and Google pays all incremental service costs plus invests $50 million in battery storage. Meta announced a deal with Sage Geosystems to bring up to 150 MW of geothermal power online by 2027, explicitly targeting its data center growth.

These are not speculative bets on future technology—these are billion-dollar contracts locking in capacity years in advance because hyperscalers understand the grid cannot scale at the pace AI compute demand is scaling. AI's extreme electricity demands are driving a geothermal energy boom, with billions flowing into 24/7 clean energy sources, a signal that the industry is scrambling to solve a problem that is already limiting deployment.

Utilities are writing bespoke, multi-gigawatt contracts with take-or-pay structures

Utilities are responding by writing increasingly bespoke, long-term power contracts with data center customers that include take-or-pay minimums (the customer pays for capacity whether or not they use it), upfront infrastructure payments, and large renewable PPAs. Duke Energy has 7.6 GW of executed energy service agreements with data centers, nearly two-thirds already under construction, with contracts structured to ensure data centers pay for capacity and grid upgrades whether or not they fully utilize them. Management explicitly ties 5–7% EPS growth through 2030 to contracted data center load and associated capital deployment. Southern Company reports 7 GW of contracted large-load agreements through 2029, ramping to 8 GW in the 2030s, with data centers driving projected 8% annual electric sales growth—well above the sector's typical 1–2% organic growth. AES has over 12 GW of signed long-term PPAs, with roughly 10–12 GW specifically serving data center customers including Meta and Google, and recent quarters show 1.6 GW of new PPAs entirely with data centers.

These are not pipelines—these are executed contracts with investment-grade counterparties, structured with take-or-pay minimums and upfront payments to de-risk the utility's capital deployment. Duke's 7.6 GW of executed agreements, if fully built out, would represent roughly $190 billion in full-stack capex at Microsoft's internal estimate of $25 million per MW, implying tens of billions in utility-side infrastructure investment (substations, transmission, distribution upgrades) over the next 5–7 years. Southern's 7–8 GW of contracted large-load through 2029–2030s implies a similar scale. AES' 12 GW of signed PPAs translates into a multi-billion-dollar contracted revenue backlog that underpins forward earnings visibility.

The market is pricing capex as if it translates into capacity immediately

Equity markets are still pricing hyperscaler capex as if power is abundant. The consensus narrative remains focused on chips, models, and compute—NVIDIA earnings, GPU supply, and software breakthroughs dominate sell-side research and investor attention. Power infrastructure, by contrast, is treated as a boring, solved problem: utilities are assumed to deliver electricity on demand, and data center operators are assumed to face no material friction in securing multi-hundred-megawatt connections.

The gap exists because the power constraint is not yet visible in hyperscaler earnings. Microsoft, Alphabet, Amazon, and Meta are all growing revenue, expanding margins, and guiding to higher AI-related capex, which reinforces the narrative that the buildout is on track. What is not yet visible in quarterly results is the lag between capex commitment and usable capacity: a hyperscaler can book $10 billion in data center construction spending in a single quarter, but if the utility interconnection study takes three years and the transformer delivery takes another 18 months, that capacity does not generate revenue until 2028 or later. The market is pricing the capex as if it translates into capacity immediately, when in reality it translates into a multi-year queue position.

Informational asymmetry also plays a role. Utility executives and grid planners understand the interconnection bottleneck intimately—they live it every day—but this knowledge is siloed in regulatory filings, technical working groups, and utility earnings calls that most tech investors do not follow. Hyperscaler management teams are now starting to flag power as a constraint on earnings calls, but the language is still hedged and the quantification is vague, so the message has not yet crystallized into a clear, market-moving narrative. The result: the people who understand the power constraint (utilities, grid operators, data center developers) are not the people setting hyperscaler valuations (tech-focused equity analysts and portfolio managers), and the people setting hyperscaler valuations are still anchored to a world where power is a footnote, not the critical path.

Duke Energy — 7.6 GW of executed contracts, two-thirds under construction

Duke Energy (DUK) has 7.6 GW of executed data center energy service agreements, nearly two-thirds already under construction, with take-or-pay and minimum-take contract structures that protect earnings and ratepayers. Management explicitly ties 5–7% EPS growth through 2030 to contracted data center load and associated capital deployment. Trading at 18.7x earnings and 1.77x price-to-book with a $96.5 billion market cap, DUK is priced in line with the utility sector median but not yet for the step-change in capex deployment and earnings visibility from the data center backlog. The catalyst is regulatory approval of additional data center rate classes and capex recovery mechanisms in North Carolina, South Carolina, and Florida—Duke's three largest service territories. The risk is if contracted load fails to materialize due to hyperscaler project cancellations, or if regulators balk at cost recovery, forcing Duke to absorb infrastructure costs without corresponding revenue.

Duke's contracts are structured to de-risk the utility's position: data centers pay upfront for dedicated substations and transmission upgrades, and take-or-pay minimums ensure Duke recovers capital even if the customer underutilizes capacity. This is critical because state regulators are sensitive to cross-subsidization—residential ratepayers will not tolerate paying for infrastructure that exclusively benefits a handful of large tech companies. Duke's approach insulates residential customers from data center costs while locking in a multi-year earnings growth driver. The 7.6 GW pipeline, if fully energized, implies roughly $15–20 billion in utility-side capex over the next five years, supporting Duke's guided 5–7% EPS growth through 2030. Trading at 18.7x earnings, the stock prices steady regulated utility growth but not the structural re-rating that comes from being one of the few utilities with a locked-in, multi-gigawatt data center backlog already under construction.

Southern Company — 8% annual electric sales growth from data centers

Southern Company (SO) has 7 GW of contracted large-load agreements through 2029, ramping to 8 GW in the 2030s, with data centers driving projected 8% annual electric sales growth—well above the sector's typical 1–2% organic growth. Strong credit provisions in contracts ensure large customers pay for dedicated infrastructure, supporting capex deployment without cross-subsidizing residential ratepayers. Trading at 23.8x earnings and 2.80x price-to-book with a $104.3 billion market cap, SO is priced at a premium to Duke, reflecting market recognition of its large-load pipeline. The catalyst is conversion of late-stage pipeline projects into executed contracts, particularly in Georgia and Alabama where Southern operates Georgia Power and Alabama Power. The risk is shareholder and regulatory scrutiny of cost allocation if data center demand underperforms, or if Southern's nuclear construction projects (Vogtle Units 3 and 4, now operational but over budget and behind schedule) create political headwinds for additional large-capex approvals.

Southern's 8% guided electric sales growth is a step-change for a sector that typically grows at 1–2% annually, driven by population growth and modest electrification. The data center contracts are the difference: 7–8 GW of contracted load translates into billions of dollars in annual revenue once energized, and Southern's take-or-pay structures ensure revenue stability even if utilization fluctuates. The premium valuation—23.8x earnings versus Duke's 18.7x—reflects this visibility, but the stock does not yet price the structural growth embedded in those agreements. If Southern executes on its late-stage pipeline and converts additional prospects into signed contracts, the earnings growth trajectory through 2030 will significantly exceed current sell-side estimates, justifying a further re-rating.

Constellation Energy — nuclear restarts and the Microsoft PPA

Constellation Energy (CEG) owns the largest U.S. nuclear fleet and has signed a 20-year PPA with Microsoft to restart Three Mile Island Unit 1 (835 MW of carbon-free baseload power), proving the model works. Nuclear is the only technology delivering 24/7, carbon-free, weather-independent baseload power at the scale hyperscalers need, and Constellation has additional idle or underutilized reactors that can be brought online faster than any new build. Trading at 24.9x earnings and 2.82x price-to-book with a $96.5 billion market cap, CEG is priced for nuclear's structural advantage in serving AI workloads. The catalyst is additional hyperscaler nuclear PPAs—if Constellation signs two more deals of comparable scale in the next 18 months, the stock re-rates sharply. The risk is nuclear operating risk (unplanned outages, cost overruns on restarts) and political or regulatory opposition to reactor life extensions or restarts, particularly in states with anti-nuclear constituencies.

The Three Mile Island restart is a proof of concept: a hyperscaler willing to commit billions of dollars over 20 years to secure firm, carbon-free power from a reactor that has been offline since 2019. The economics work because nuclear provides the exact load profile AI data centers require—constant, high-capacity-factor baseload power with no weather dependence and no carbon emissions. Constellation's existing fleet positions it to replicate this model: the company operates multiple reactors that could be restarted or have their output redirected to data center customers under long-term PPAs. If Constellation signs two more hyperscaler deals of similar scale—say, 1,500–2,000 MW combined—the contracted revenue backlog would support 30–40% upside from current levels, as the market re-rates CEG from "utility with nuclear exposure" to "scarce provider of the only power source that meets hyperscaler requirements."

Vistra — scarcity asset with firm dispatchable capacity in ERCOT and PJM

Vistra Corp (VST) operates 41 GW of firm, dispatchable capacity—gas peakers and nuclear baseload—in ERCOT and PJM, the two markets where AI data center construction is concentrated. Vistra can supply firm, around-the-clock power to data centers without relying on new grid build-outs, capturing scarcity rents as hyperscalers compete for existing capacity. The company has already signed multi-year PPAs with hyperscalers, locking in margin at rates well above historical merchant averages. Trading at 24.1x earnings and 9.64x price-to-book with a $53.6 billion market cap, VST has re-rated sharply as the market begins to recognize the value of in-place, energized generation capacity. The catalyst is additional long-term power contracts with hyperscalers in ERCOT or PJM, where interconnection queues are longest and new supply is years away. The risk is if grid capacity expands faster than expected or if AI demand moderates, compressing scarcity rents and reducing the premium Vistra can command on contracted PPAs.

Vistra is the thesis in equity form: scarcity asset, contracted cash flows, structural moat. The company's 41 GW of firm capacity is already built, already interconnected, and already delivering power to the grid. When hyperscalers need multi-hundred-megawatt connections and the utility tells them the interconnection study will take four years, Vistra can offer a behind-the-meter solution: co-locate your data center next to our power plant, and we'll deliver firm capacity from day one. This bypasses the grid bottleneck entirely, and Vistra can charge a premium for it. The multi-year PPAs the company has signed with hyperscalers lock in margin well above the $30–40/MWh that merchant generators typically earn in ERCOT and PJM, because hyperscalers are willing to pay $60–80/MWh for firm, guaranteed capacity that doesn't depend on grid availability. If Vistra signs another 2–3 GW of hyperscaler contracts over the next 12 months, the stock trades to $200.

AES Corporation — co-located generation bypasses transmission queues

AES Corporation (AES) has 12 GW of signed long-term PPAs, with 10–12 GW serving hyperscaler data center customers including Meta and Google. Recent quarters show 1.6 GW of new PPAs entirely with data centers. Co-located generation projects bypass transmission queues and interconnection delays—the exact constraint the thesis identifies. Trading at 7.8x earnings and 2.36x price-to-book with a $10.4 billion market cap, AES is dramatically cheaper on a valuation basis than regulated utilities, reflecting its merchant/IPP structure and higher perceived risk. The catalyst is additional hyperscaler PPA announcements and successful execution of co-located projects, which would prove the model and support a re-rating toward peer IPPs. The risk is that PPA economics depend on build costs and market power prices; if renewables overbuild or demand disappoints, contracted revenue may not translate into strong margins.

AES' co-located generation model is the solution to the interconnection problem: instead of waiting 4–6 years for a new solar or wind farm to get through the queue, AES builds generation directly adjacent to the data center and delivers power without touching the grid. This reduces transmission losses, eliminates interconnection risk, and accelerates project timelines from years to months. The 12 GW of signed PPAs with hyperscalers translates into a multi-billion-dollar contracted revenue backlog, and the take-or-pay structures in those contracts de-risk the cash flows. AES trades at 7.8x earnings—roughly half the valuation of regulated utilities with comparable data center exposure—because the market still prices AES as a merchant generator with commodity exposure, when in reality the hyperscaler PPAs have locked in 10–12 GW of capacity at fixed rates for 10–20 years. If AES executes on its co-located pipeline and signs another 2–3 GW of hyperscaler contracts, the stock re-rates to $18–20.

NextEra Energy — Florida exposure and the largest renewable portfolio

NextEra Energy (NEE) operates Florida Power & Light and the country's largest renewable generation portfolio. FPL serves Florida, a top-five U.S. data center market, and the renewable arm can offer long-term clean PPAs—exactly what hyperscalers need to hit net-zero commitments while locking in firm capacity. Trading at 21.9x earnings and 3.24x price-to-book with a $178.8 billion market cap, NEE is priced for steady regulated utility growth but not yet for a step-change in data center-driven capex deployment. The catalyst is additional large-load contract announcements in Florida or Texas service territories, where FPL and NEE's renewable arm can bundle regulated utility service with clean PPAs. The risk is Florida's regulatory environment—state regulators have historically been conservative on cost recovery for large industrial customers—and hurricane-related capex diverting resources from data center projects.

NEE's dual structure—regulated utility (FPL) plus unregulated renewables (NextEra Energy Resources)—positions it to capture data center demand from both angles. FPL can sign traditional utility contracts with data centers in Florida, while NEE Resources can offer long-term renewable PPAs to hyperscalers anywhere in the U.S. The company's renewable portfolio is the largest in North America, giving it the scale to deliver multi-gigawatt PPAs that smaller developers cannot match. If NEE signs 2–3 GW of new data center contracts over the next 18 months—split between FPL and NEE Resources—the stock trades to $105, pricing in the structural growth from data center load that is not yet embedded in sell-side models.

Equinix — interconnection density and energized capacity as a moat

Equinix (EQIX) operates 260+ interconnection and colocation facilities globally with energized power and grid interconnections that took years to secure. As new data center construction stalls due to power constraints, demand for Equinix's in-place capacity should rise, supporting occupancy gains and pricing power. The interconnection density—multiple cloud on-ramps per facility—is a structural moat that pure colocation providers lack. Trading at 75.3x earnings and 7.49x price-to-book with a $107.4 billion market cap, EQIX is priced for growth but not yet for the scarcity premium that comes from being one of the few operators with energized, available capacity. The catalyst is pricing power and occupancy gains as new supply is delayed. The risk is if power constraints ease or if hyperscalers build their own campuses instead of leasing, eroding EQIX's scarcity premium.

Equinix's value proposition is the 260+ facilities already online with energized substations and utility contracts in place. When only 50% of 2026 planned capacity is under construction, hyperscalers cannot wait 4–5 years for interconnection—they lease from whoever has live capacity today. Equinix's interconnection density—its facilities serve as meet-me rooms where multiple cloud providers, ISPs, and enterprises interconnect—creates network effects that pure colocation providers cannot replicate. A hyperscaler leasing space in an Equinix facility gets not just power and cooling, but direct, low-latency connections to every other tenant in the building. This is worth a premium, and as new supply is delayed, that premium expands. The risk is the valuation: at 75x earnings and 30x EV/EBITDA, EQIX prices in significant growth, and if the power constraint eases faster than expected, the scarcity premium compresses.

Digital Realty Trust — 300+ facilities with established utility relationships

Digital Realty Trust (DLR) operates 300+ data center facilities across 26 countries with 50 million square feet of colocation space. The company's value proposition is the 50 million square feet already online with energized substations and utility contracts in place. When only 50% of 2026 planned capacity is under construction, hyperscalers cannot wait 4–5 years for interconnection—they lease from whoever has live capacity today. Trading at 46.8x earnings and 2.76x price-to-book with a $65.7 billion market cap, DLR is a lower-multiple alternative to Equinix with similar exposure to the power scarcity dynamic. The catalyst is occupancy and pricing gains in key markets as new supply is delayed. The risk is the same as Equinix: if supply catches up or demand moderates, the scarcity premium disappears.

Digital Realty's 300+ facilities with established utility relationships and take-or-pay power agreements across multiple continents provide broader geographic diversification than Equinix's 260+ interconnection hubs. DLR's facilities are larger on average—wholesale colocation serving hyperscalers and large enterprises—while Equinix skews toward retail colocation and interconnection. Both benefit from the power constraint, but DLR's wholesale focus means it captures more of the hyperscaler demand directly. At 21x EV/EBITDA, DLR's valuation is more reasonable than EQIX's 30x given similar thesis exposure. If DLR reports occupancy gains and pricing power in its next two quarters, the stock re-rates to $185, pricing in the scarcity premium that is not yet embedded in the current multiple.

Eaton Corporation — transformer shortages and electrical equipment bottlenecks

Eaton Corporation (ETN) supplies electrical distribution equipment, uninterruptible power supplies (UPS, battery backup systems that keep data centers online during grid outages), switchgear (electrical switches that control and protect circuits), and backup power infrastructure to data centers. Data centers cannot energize without Eaton's equipment, and transformer shortages—30% capacity deficit, 2.5-year lead times—directly translate into backlog and pricing power for Eaton. Trading at 39.9x earnings and 8.05x price-to-book with a $159.1 billion market cap, ETN is priced at a premium reflecting strong execution and market leadership. The catalyst is transformer and switchgear supply tightness driving pricing power as hyperscalers compete for limited equipment supply. The risk is if equipment supply chains normalize faster than expected, compressing margins and reducing backlog visibility.

Eaton's backlog is the tell: the company has multi-year visibility into data center electrical equipment demand, and the 2.5-year lead times on transformers mean orders placed today won't deliver until 2028. This creates a natural pricing tailwind—when supply is constrained and lead times are long, buyers pay premiums to secure delivery slots. Eaton's UPS and switchgear products are mission-critical: a data center without backup power is not a data center, it's a liability. Hyperscalers cannot compromise on electrical infrastructure, so they pay whatever Eaton charges. The risk is the valuation: at 40x earnings and 29x EV/EBITDA, ETN is priced for perfection. If transformer supply normalizes or if hyperscaler capex deployment slows, the stock de-rates sharply. But if the power constraint thesis plays out and Eaton captures outsized pricing power from multi-year backlogs, the premium valuation is defensible.

Portfolio construction — why these weights

This is a 9-position portfolio structured to capture the power constraint thesis from three angles: regulated utilities with executed data center contracts (DUK, SO, CEG, NEE at 52% combined), independent power producers with signed hyperscaler PPAs and co-location advantages (AES, VST at 23% combined), and data center operators with scarce energized capacity (EQIX, DLR at 15% combined), plus electrical equipment suppliers benefiting from transformer and switchgear bottlenecks (ETN at 10%).

The weighting reflects conviction: DUK, SO, CEG, VST, AES, and DLR all received 'Core' verdicts with direct structural exposure to the thesis mechanism, while EQIX received 'Supporting' due to full valuation (30x EV/EBITDA) and negative free cash flow despite strong thesis fit. NEE and ETN are sized smaller—NEE because Florida exposure is narrower than Duke or Southern's multi-state footprints, ETN because the premium valuation (40x earnings, 29x EV/EBITDA) leaves limited margin for error if hyperscaler capex deployment slows.

The four regulated utilities (DUK, SO, CEG, NEE) are not identical—Duke and Southern have multi-GW executed contracts with take-or-pay structures already under construction, while NEE's Florida data center pipeline is earlier-stage and CEG's nuclear restart model, though proven with Microsoft, has not yet scaled to multiple hyperscaler PPAs. Similarly, VST and AES are both IPPs with hyperscaler contracts, but VST's firm dispatchable capacity in ERCOT and PJM (the two markets where AI data center construction is concentrated) and its ability to deliver power from existing plants without waiting for new transmission earns it higher conviction than AES, whose co-located generation projects are still in the build-out phase. EQIX and DLR are both data center REITs with in-place energized capacity, but DLR's 300+ facilities with established utility relationships and take-or-pay power agreements across multiple continents provide broader geographic diversification than EQIX's 260+ interconnection hubs, and DLR's 21x EV/EBITDA is more reasonable than EQIX's 30x given similar thesis exposure.

Ticker	Weight	Target	Horizon	Thesis fit
DUK	15%	$145	540d	7.6 GW executed contracts, two-thirds under construction, take-or-pay structures
SO	15%	$108	540d	7 GW contracted load driving 8% annual electric sales growth
CEG	12%	$135	450d	Largest U.S. nuclear fleet, Microsoft PPA proves reactor restart model
VST	13%	$200	450d	41 GW firm capacity in ERCOT/PJM, scarcity asset capturing premium rates
AES	10%	$18	540d	12 GW hyperscaler PPAs, co-located generation bypasses grid queues
NEE	10%	$105	540d	FPL serves Florida, renewable arm can bundle clean PPAs with utility service
EQIX	7%	$1,250	450d	260+ energized facilities, interconnection density as moat
DLR	8%	$185	450d	300+ facilities with locked power, wholesale colocation for hyperscalers
ETN	10%	$385	540d	Transformer shortages (2.5-year lead times) translate to backlog and pricing power

Assumptions and falsification conditions

1. Hyperscaler AI capex deployment continues at guided rates ($345 billion combined 2025 spend) and translates into data center construction demand through 2027–2028. Falsified if: hyperscalers cut capex guidance by >30% in 2026 due to demand weakness, model efficiency gains that reduce compute requirements by >50%, or strategic pivot away from owned infrastructure toward third-party cloud.

2. U.S. electric grid capacity expansion remains constrained at 15–20 GW per year through 2028, with interconnection queue timelines of 4–6 years for new generation projects. Falsified if: FERC Order 2023 implementation or new DOE rulemakings materially shorten interconnection timelines to <2 years by 2027, or if utilities accelerate transmission and substation buildouts such that 40+ GW of new data center capacity can energize annually by 2027.

3. Data center operators without locked-in, long-term power contracts face either construction delays (12–24 months beyond planned timelines) or margin compression from paying spot power rates 50–100% above contracted PPA rates. Falsified if: spot power markets in key AI deployment regions (Northern Virginia, ERCOT, PJM) remain stable or decline, or if data center operators successfully negotiate short-term contracts at rates comparable to long-term PPAs signed by hyperscalers in 2024–2025.

4. Utilities and IPPs with executed data center power contracts (Duke's 7.6 GW, Southern's 7 GW, AES' 12 GW, Constellation's Microsoft PPA) convert those contracts into energized capacity and revenue on the timelines management has guided (2026–2029), without material regulatory pushback on cost recovery or project cancellations by hyperscaler counterparties. Falsified if: >30% of contracted data center load fails to materialize due to hyperscaler project cancellations, or if state regulators in North Carolina, Georgia, or Illinois block cost recovery for data center-related capex, forcing utilities to absorb costs.

5. Nuclear restarts and co-located generation projects (Constellation's Three Mile Island, AES' co-located renewables with Google and Meta) proceed on schedule and establish a replicable model for additional hyperscaler PPAs. Falsified if: Three Mile Island restart faces NRC delays or cost overruns that push commercial operation beyond 2028, or if co-located generation projects encounter interconnection or permitting obstacles that negate their speed-to-market advantage over traditional grid-connected projects.

Risks

Regulatory and permitting risk. Utilities' ability to deploy capex and recover costs through rate base expansion depends on state PUC approvals, which can be delayed or denied if regulators perceive data center load as benefiting a narrow customer class at the expense of residential ratepayers. If Duke, Southern, or NEE face regulatory pushback on cost recovery for data center infrastructure, the earnings growth embedded in the thesis weakens. Similarly, if FERC or DOE rulemakings intended to accelerate interconnections are delayed or watered down, the grid bottleneck persists longer but the political risk to utilities increases.

Hyperscaler execution and demand risk. The thesis assumes hyperscalers continue deploying $300+ billion annually in AI capex through 2027–2028. If AI workload growth disappoints—due to slower enterprise adoption, model efficiency gains, or competition from open-source models—hyperscalers could cut capex, delay data center projects, or renegotiate power contracts, reducing demand for utility and IPP capacity. The take-or-pay structures in Duke and Southern's contracts provide downside protection, but AES and VST, as merchant generators, have more exposure to demand fluctuations.

Technology and efficiency risk. If next-generation AI architectures (e.g., sparse models, neuromorphic chips, or quantum-classical hybrids) reduce compute requirements per inference by 50–70%, the power constraint thesis weakens because hyperscalers can scale workloads within existing grid capacity. Similarly, if liquid cooling and advanced thermal management reduce data center power consumption per rack by 30–40%, the urgency of securing multi-hundred-MW power contracts diminishes.

Grid capacity expansion faster than expected. If DOE's push to accelerate data center interconnections results in a new rulemaking that allows co-located load and generation requests to bypass traditional queues, and if transformer manufacturing capacity expands faster than the current 2.5-year lead times suggest, the grid bottleneck could ease by 2027–2028, compressing the scarcity premium on locked-in power contracts. This would hurt VST and AES (merchant generators capturing scarcity rents) more than DUK and SO (regulated utilities earning returns on capex regardless of scarcity).

Crowded trade and valuation risk. CEG, VST, and ETN have all re-rated sharply in the past 12–18 months as the market begins to recognize the power constraint thesis. CEG trades at 24.9x earnings, VST at 24x, ETN at 40x—all well above historical utility and industrial medians. If the thesis becomes consensus before the contracted revenue and earnings growth materialize, these names could face multiple compression even if fundamentals play out as expected. EQIX and DLR, as REITs, face interest rate sensitivity: if the Fed holds rates higher for longer, REIT valuations compress regardless of data center demand.