Only about 50% to 60% of U.S. data center capacity scheduled for the next one to two years is expected to arrive on time, even as U.S. data center power demand is forecast to rise from 31 GW in 2025 to 66 GW in 2027 (Goldman Sachs Research, 2026). That single mismatch now explains more about hyperscale AI strategy than GPU allocation, cloud pricing, or model choice. The bottleneck has moved from chip availability to electricity that can be contracted, interconnected, firmed, and defended before a rival locks up the same substation.
By August 2026, AI data center power demand has become a board-level constraint because the interconnection queue is no longer a back-office utility process. It's a capital allocation filter. A hyperscaler can announce a 1 GW campus, order accelerators, secure land, and sign anchor tenants, yet still lose two years if transmission studies, capacity accreditation, or local load rules don't line up. That lag changes the economics of every hyperscaler PPA contract. Power is no longer an operating input bought after site selection; it's the site selection thesis.
The market is splitting into three groups. The first has contracted power at scale, often through nuclear, renewables, or gas-backed structures. The second has land and capital but uncertain queue position. The third has demand but is becoming a price taker in cloud and GPU-as-a-service markets. For CFOs and infrastructure investors, the question in 2026 isn't whether AI compute demand exists. IDC says AI infrastructure spending reached $318 billion in 2025 and is projected to reach $487 billion in 2026 (IDC, 2026). The more important question is which megawatts become billable before financing costs, grid charges, and local opposition consume the margin.
This briefing treats power as the scarce asset in AI infrastructure. For related coverage on capital spending and enterprise technology allocation, see MarketIntel.
$487 Billion Chases Scarce Megawatts
IDC projects global AI infrastructure spending of $487 billion in 2026, up roughly 53% from $318 billion in 2025, with a five-year compound annual growth rate near 31% through 2029 as the market exceeds $1 trillion (IDC, 2026). That is the total addressable market for the compute stack, but the serviceable market for power-constrained AI capacity is smaller and more valuable. In practice, the monetizable slice is the portion of AI infrastructure that can be matched with firm power, cooling, fiber, and permits within a commercial window.
Gartner's power forecast gives that bottleneck a hard boundary. Worldwide data center electricity consumption is projected to reach 565 TWh in 2026, up from 447 TWh in 2025, while global data center power demand is expected to rise from 104 GW in 2025 to 132 GW in 2026 and 290 GW by 2030 (Gartner, 2026). That implies a roughly 23% compound annual growth rate in power demand from 2025 to 2030, faster than most regulated utilities can add dispatchable capacity and transmission in congested zones. AI-optimized servers are expected to account for 175 TWh of consumption in 2026, or 31% of data center power use, and to exceed conventional server power consumption in 2027 (Gartner, 2026).
The U.S. serviceable available market is moving even faster. Goldman Sachs Research forecasts U.S. data center power demand at 41 GW in 2026 and 66 GW in 2027, versus 31 GW in 2025, with data centers rising to 8.5% of U.S. peak summer power demand in 2027 from 4.1% in 2025 (Goldman Sachs Research, 2026). BloombergNEF's July 2026 U.S. capacity outlook puts installed U.S. data center capacity at 118 GW by 2030 and 194 GW by 2035, after a 52% upward revision to its 2030 forecast since December 2025 (BloombergNEF, 2026). Using BloombergNEF's earlier 2024 base of almost 35 GW and its 2035 forecast of 194 GW, the implied 2024 to 2035 CAGR is roughly 17%, an analyst calculation based on BloombergNEF published figures.
Regional differences matter because power markets clear locally. Northern Virginia and the broader PJM region face the hardest constraint because data center load growth, capacity market scarcity, and transmission congestion are colliding. Texas remains more flexible because ERCOT can add generation faster, but it still exposes buyers to price volatility and extreme-weather risk. Georgia, Ohio, Pennsylvania, Oklahoma, and North Carolina are attracting AI campuses because they combine land, fiber routes, industrial zoning, and politically acceptable generation additions. Europe is more constrained by permitting and energy-price sensitivity, while the Middle East is gaining share through sovereign AI programs backed by power and land, a pattern visible in IDC's report that Middle East and Africa AI infrastructure spending grew more than 500% year over year in Q4 2025 to $1.8 billion (IDC, 2026).
The historical baseline shows why 2026 is the inflection year. Cloud data centers once optimized for unit cost, latency, and tax incentives. AI campuses optimize for guaranteed megawatts, often hundreds at a time. Lawrence Berkeley National Laboratory reported 2,061 GW of generation and storage capacity active in U.S. interconnection queues at the end of 2025, down from a peak near 2,600 GW in 2023 but still a massive backlog (LBNL and GridTracker, 2026). That means AI demand is arriving while the supply-side queue remains clogged, turning interconnection status into a balance-sheet variable.
The Players Locking Up Supply
Alphabet is behaving like power procurement is part of cloud product management. Google signed two long-term PPAs with TotalEnergies in February 2026 for 1 GW of solar capacity in Texas, equal to 28 TWh over 15 years, with construction scheduled to begin in Q2 2026 (TotalEnergies SEC filing, 2026). Alphabet reported $350.0 billion in 2025 revenue and disclosed $149.1 billion of purchase commitments at year-end 2025, while warning that AI infrastructure will raise depreciation, energy, equipment, and network costs (Alphabet Form 10-K, FY2025). Its December 2025 agreement to acquire Intersect for $4.8 billion shows the strategy: own more of the data center and energy development chain before the queue prices that optionality away (Alphabet Form 10-K, FY2025).
Microsoft is the most exposed to the gap between AI demand and energy availability because Azure growth is absorbing capacity almost as fast as the company can build it. Microsoft reported fiscal 2026 revenue of $331.8 billion, up 18%, while server products and cloud services revenue reached $129.4 billion (Microsoft Form 10-K, FY2026). The company warned that electricity availability, connection delays, outages, and higher power costs could restrict data center expansion, which is unusually direct language for a mega-cap software company (Microsoft Form 10-K, FY2026). Its power strategy has shifted toward long-duration clean power, including the previously announced Brookfield renewable energy framework, because AI margins depend on multi-year energy cost visibility.
Amazon is taking a wider portfolio approach, mixing utility-scale investments, nuclear agreements, and self-designed silicon to reduce both power and accelerator constraints. Amazon disclosed that certain energy contracts extend 20 years and covered roughly 200 million MWh subject to derivative accounting at the end of 2025, with a weighted-average remaining duration of about 16 years (Amazon Form 10-K, FY2025). In June 2025, Talen Energy expanded its nuclear power relationship with Amazon, agreeing to supply up to 1,920 MW of carbon-free electricity from Susquehanna to support AWS data center operations (Talen Energy SEC filing, 2025). Amazon's 2025 U.S. investment announcements included $11 billion in Georgia, $10 billion in North Carolina, and $20 billion in Pennsylvania data center campuses, confirming that site choice is following available infrastructure, not just customer density (Amazon Economic Impact Report, 2026).
Meta has become the clearest buyer of long-duration firm clean capacity. It guided to 2026 capital expenditures of about $115 billion to $135 billion to support AI and its core business, after $69.7 billion of 2025 property and equipment purchases (Meta Form 10-K, FY2025). Vistra disclosed in January 2026 that it signed 20-year PPAs with Meta for 2,609 MW of carbon-free power and capacity from PJM nuclear plants, including 2,176 MW of operating energy and capacity and 433 MW of planned uprates (Vistra Form 10-K, FY2025). This is not ordinary renewable procurement; it's an attempt to reserve accredited capacity in the market where data center load is under the most political and reliability pressure.
Oracle is turning power access into a cloud share weapon. Oracle reported fiscal 2026 revenue of $67.4 billion, with cloud revenue of $34.0 billion, up from $24.5 billion in fiscal 2025 (Oracle Form 10-K, FY2026). The company's pitch to AI customers rests on large cloud regions and dedicated capacity for model training, which means its competitive position depends on getting multi-hundred-megawatt blocks energized quickly. Oracle doesn't have the same consumer cash engine as Alphabet or Meta, so slower interconnection would hit it harder through financing cost and customer delivery schedules.
CoreWeave sits between hyperscalers and specialist AI demand, and it shows how power has become a traded asset. The company reported 2025 revenue of $5.1 billion, up from $1.9 billion in 2024, with remaining performance obligations of $60.7 billion at year-end 2025 (CoreWeave Form 10-K, FY2025). CoreWeave operated 43 data centers with more than 850 MW of active power and had about 3.1 GW of contracted power capacity as of December 2025 (CoreWeave Form 10-K, FY2025). Its customer contracts are mostly take-or-pay, which means power delivery timing directly controls revenue recognition and debt service.
Vistra and Constellation are gaining share on the supply side because nuclear capacity has become the premium product in AI power contracting. Vistra paired the Meta PPAs with nuclear uprates and a planned acquisition of Cogentrix Energy's roughly 5,500 MW natural gas portfolio, while guiding to about $2.6 billion of 2026 capital expenditures and nuclear fuel purchases (Vistra Form 10-K, FY2025). Constellation told investors it submitted 5,000 MW of new capacity into PJM processes and highlighted long-term contracting options for baseload clean generation (Constellation SEC investor materials, 2026). The share gain mechanism is simple: companies with accredited capacity, nuclear operating history, and interconnection credibility can sell certainty, while renewable-only developers sell energy that still needs firming.
FERC Forced The Queue Fight
The decisive 2026 trigger is FERC's June 18, 2026 show-cause action on large-load integration. The Federal Energy Regulatory Commission ordered all six regional grid operators under its jurisdiction to justify or reform tariffs governing how data centers, manufacturing plants, and other large energy users connect to the grid (FERC, 2026). That action changed the market from a quiet negotiation among utilities, developers, and large customers into a regulated contest over who pays, who waits, and which projects count as real.
The trigger matters because speculative load requests had started to distort planning. FERC Commissioner David Rosner said large loads were larger, more concentrated, and able to change consumption in seconds, while speculative projects could clog queues and inflate forecasts (FERC Commissioner remarks, 2026). In plain English, some developers were shopping projects across utilities, asking for studies in several places, then keeping optionality without committing capital at the same pace. That creates a read-then-act problem for grid planners: utilities study load that may not happen, build forecasts around it, and risk charging other customers for infrastructure that was sized for phantom demand.
PJM is the test case because it combines the largest U.S. data center concentration with tight capacity conditions. PJM serves 67 million people and, in July 2026, filed a Reliability Backstop Procurement proposal tied to a 6,831 MW shortfall in the 2028/2029 delivery year capacity auction (PJM, 2026). The proposed procurement would seek resources for terms of up to 15 years with a maximum willingness to pay of $555 per MW-day (PJM, 2026). The policy signal is blunt: large loads that don't bring capacity, agree to curtailment, or sign qualifying bilateral contracts may face direct cost allocation.
That is why hyperscaler PPA contracts are shifting from renewable energy certificates and hourly matching toward capacity-backed structures. A solar PPA can satisfy a clean-energy goal, but it may not solve the reliability charge if it doesn't provide accredited capacity during tight hours. Nuclear, gas with carbon strategy, geothermal, storage-backed renewables, and behind-the-meter structures are becoming more valuable because they answer the regulator's core question: can this load be served without raising reliability risk for everyone else?
Three Risks Are Mispriced
The base risk is schedule slippage, with an analyst-estimated probability near 60% for large U.S. AI campuses planned for 2026 to 2028. Goldman Sachs expects only about 50% to 60% of data center capacity scheduled for the next one to two years to come online on time (Goldman Sachs Research, 2026). The mechanism is not one delay; it's stacked delays in interconnection studies, transformers, gas turbines, substations, local permits, and cooling systems. The affected players are hyperscalers, GPU cloud firms, colocation landlords, and utilities in PJM, MISO, and selected Western markets. The timeline is already visible in 2026 procurement plans, but the revenue impact will be felt most sharply in 2027 and 2028 as committed AI capacity misses delivery dates.
The second risk is cost socialization backlash, with an analyst-estimated probability near 45% in constrained regions. Residential and industrial customers are unlikely to accept rising grid charges if they believe data centers caused the need for new generation or transmission. PJM's 2026 process already points toward allocating backstop procurement costs to data centers that haven't procured capacity or accepted curtailment (PJM materials, 2026). The mechanism is political pressure that turns into tariff reform, interconnection deposits, readiness screens, or special rate classes. The affected players are merchant data center developers, cloud providers without firm supply, and utilities exposed to public utility commission hearings. The relevant timeline is 2026 through 2029, because rate cases and capacity auctions move slower than AI campus announcements.
The third risk is contract mismatch, with an analyst-estimated probability near 35% for newer AI infrastructure platforms. CoreWeave reported roughly $60.7 billion of remaining performance obligations and about 3.1 GW of contracted power capacity at year-end 2025 (CoreWeave Form 10-K, FY2025). Those numbers are powerful, but they also create delivery pressure. If a provider signs take-or-pay customer contracts, raises debt, and then faces delayed power delivery, the cash-flow model can tighten quickly. The affected players are GPU-as-a-service providers, single-customer colocation conversions, and lenders relying on campus-level contracts. The timeline is 2026 to 2027 because many contracts begin when capacity is delivered, not when the press release lands.
The underweighted tail risk is a reliability event blamed on flexible AI load. FERC has already highlighted that large loads can change consumption in seconds (FERC Commissioner remarks, 2026). If a regional outage, voltage event, or emergency dispatch episode is politically tied to AI campuses, regulators could impose mandatory curtailment protocols, minimum self-supply requirements, or tougher readiness rules. The probability is low, analyst-estimated at 15% over 24 months, but the impact would be high because it would lower the valuation of queue positions that don't include firm capacity and grid services.
Enterprise Buyers
Enterprise buyers should stop treating AI capacity as a pure cloud sourcing exercise. Procurement teams need to ask cloud and GPU vendors for region-level power status, not just availability zones, GPU types, and service-level terms. A CFO signing a multi-year AI platform contract should require disclosure of whether the capacity is energized, under construction, subject to utility interconnection approval, or dependent on a third-party campus conversion. That language matters because Goldman Sachs expects U.S. data center demand to rise to 66 GW in 2027 while only 50% to 60% of near-term scheduled capacity arrives on time (Goldman Sachs Research, 2026).
Buyers should split workloads by power risk. Latency-sensitive inference can sit in constrained premium regions if the business case supports it, but batch training, synthetic data generation, and model evaluation can move to regions with better power visibility. Contracts should also include capacity substitution rights, price reopeners tied to material energy surcharges, and exit rights if a provider misses energization milestones. That won't remove risk, but it prevents a vendor's queue problem from becoming the buyer's product delay.
Investors
Investors should underwrite megawatts like leased square footage used to be underwritten. The first diligence question is no longer acreage or shell construction cost; it's whether the project has executed interconnection agreements, firm capacity treatment, transformer delivery slots, and credible offtake. A 500 MW announced campus with weak queue status deserves a higher discount rate than a smaller site with energized capacity and contracted expansion rights.
Debt investors should pressure sponsors for stress cases that assume 12 to 24 months of delayed billing, 15% to 25% higher electrical equipment cost, and partial curtailment during peak periods. Equity investors should favor platforms that control more of the power chain, such as nuclear-linked suppliers, utility partners, or campuses with behind-the-meter generation. The market is still rewarding headline GW announcements, but by late 2026 public filings are starting to reveal which contracts create cash and which create construction obligations.
Vendors
Vendors selling into AI data centers need to map their go-to-market plans to power readiness. Server, cooling, switchgear, and software vendors should segment accounts by energized power, committed interconnection, and speculative pipeline. The best customers are not always the biggest announced campuses; they are the projects with capital, power, and a realistic commissioning sequence.
Cooling vendors should prioritize high-density retrofit and heat rejection systems because AI-optimized servers are expected to account for 175 TWh of 2026 data center consumption (Gartner, 2026). Electrical equipment vendors should price scarcity openly because transformers, switchgear, and substations are now schedule-critical assets. Software vendors should build tools that forecast workload curtailment, energy price exposure, and carbon matching because those issues are becoming commercial terms in cloud contracts, not sustainability footnotes.
The Next Two Years Decide Winners
The base case carries a 55% analyst-estimated probability: U.S. AI data center growth continues, but delivery spreads away from the most congested markets. Under this scenario, Goldman Sachs' 41 GW U.S. data center demand forecast for 2026 and 66 GW forecast for 2027 remain directionally right, but a meaningful share of capacity shifts toward Texas, Georgia, Pennsylvania, Ohio, Oklahoma, and selected secondary markets (Goldman Sachs Research, 2026). Hyperscalers keep signing nuclear, renewable, gas, and storage-backed contracts, while utilities demand more site control, deposits, and curtailment commitments.
The contrarian view carries a 25% analyst-estimated probability: model efficiency and inference optimization slow power growth faster than expected. Amazon has argued that Trainium could save tens of billions of dollars of capital expenditures per year at scale and provide margin benefits versus relying on third-party chips (Amazon shareholder letter, 2026). If custom silicon, smaller models, and better inference routing reduce watts per token, the most speculative campuses could be deferred. This wouldn't end the power race, but it would separate necessary capacity from option-like land grabs.
The downside scenario carries a 20% analyst-estimated probability: regulatory pushback and financing stress collide. In this case, PJM-style backstop cost allocation spreads, interconnection readiness rules tighten, and lenders demand more equity for campuses without firm capacity. Data center developers with customer concentration, weak balance sheets, or unproven off-grid power systems face refinancing risk. Public cloud leaders still build, but marginal GPU cloud and colocation players lose negotiating power.
The leading indicators are concrete. First, watch PJM's Reliability Backstop Procurement implementation and whether large loads sign bilateral capacity contracts before central procurement begins (PJM, 2026). Second, track executed interconnection agreements, not announced data center GW. LBNL's queue data showed 2,061 GW of generation and storage active at the end of 2025, so queue movement is more informative than project announcements (LBNL and GridTracker, 2026). Third, follow hyperscaler purchase commitments and energy disclosures in filings. Alphabet's $149.1 billion of purchase commitments and Meta's $131.0 billion of contractual commitments at year-end 2025 show how much future capacity is already financially embedded (company filings, FY2025).
Seven Signals For Decision Makers
- Power has replaced GPUs as the binding constraint for many AI capacity plans because grid connection timing now controls revenue timing.
- U.S. data center power demand is forecast to rise from 31 GW in 2025 to 66 GW in 2027, but only about 50% to 60% of near-term scheduled capacity is expected on time (Goldman Sachs Research, 2026).
- AI-optimized servers are expected to consume 175 TWh in 2026 and surpass conventional servers in 2027, making power density a procurement issue (Gartner, 2026).
- Hyperscaler PPA contracts are shifting toward capacity-backed supply, including nuclear and firmed renewables, because energy-only deals don't solve reliability charges.
- PJM's proposed 6,831 MW backstop procurement is the clearest signal that large data center loads may need to bring capacity or pay for scarcity (PJM, 2026).
- Specialist AI cloud providers can grow faster than hyperscalers, but delayed energization can strain debt and take-or-pay contract models.
- Investors should discount announced gigawatts unless interconnection status, equipment delivery, and capacity accreditation are verified.
How should a CFO test whether an AI cloud contract carries hidden power risk?
A CFO should ask for the physical status of the capacity behind the contract: energized, under construction, utility-approved, or only planned. The distinction affects delivery risk and pricing power. CoreWeave's filings show why this matters. The company reported $60.7 billion of remaining performance obligations and about 3.1 GW of contracted power capacity at year-end 2025, but its revenue depends on bringing capacity online across future periods (CoreWeave Form 10-K, FY2025). A buyer should require remedies if a named region misses capacity dates, including substitution into another region, service credits tied to delayed availability, and the right to reduce committed spend. For mission-critical AI workloads, the contract should also disclose whether power cost pass-throughs, curtailment rights, or grid emergency rules can affect runtime availability.
Are nuclear PPAs now the premium structure for hyperscaler AI demand?
Nuclear PPAs are becoming premium contracts where capacity accreditation, carbon goals, and delivery certainty overlap. Meta's 20-year agreement with Vistra covers 2,609 MW of carbon-free power and capacity from PJM nuclear plants, including 2,176 MW from operating energy and capacity and 433 MW from uprates (Vistra Form 10-K, FY2025). Amazon's expanded Talen agreement covers up to 1,920 MW from the Susquehanna nuclear plant for AWS operations (Talen Energy SEC filing, 2025). These deals matter because they address both energy and capacity in regions where reliability charges can become material. They aren't easy to replicate because nuclear plants are scarce, politically sensitive, and already contracted in many markets. That scarcity is exactly why the companies that secure these agreements gain strategic insulation.
Why doesn't signing a renewable PPA fully solve the data center energy bottleneck?
A renewable PPA can lower carbon exposure and secure long-term energy, but it doesn't always guarantee deliverable capacity during system stress. Google's 1 GW TotalEnergies solar agreement in Texas is large and economically meaningful, equal to 28 TWh over 15 years (TotalEnergies SEC filing, 2026). Yet solar output depends on time of day and must be paired with storage, load flexibility, market purchases, or other firming resources to cover around-the-clock AI operations. In PJM, the issue is sharper because capacity market rules focus on accredited reliability contribution, not only annual MWh. That means a hyperscaler may need a portfolio: renewables for cost and carbon, nuclear or gas-backed supply for firmness, storage for ramping, and contractual curtailment terms to satisfy grid operators.
Which public companies have the best exposure to the AI power shortage?
The strongest exposure sits with companies that control scarce capacity or can convert power into billable compute quickly. Vistra has direct upside through the Meta PPAs, nuclear uprates, and its planned Cogentrix gas acquisition of roughly 5,500 MW (Vistra Form 10-K, FY2025). Constellation benefits from nuclear scarcity and has highlighted 5,000 MW of new capacity submissions in PJM-related materials (Constellation SEC investor materials, 2026). Alphabet, Amazon, Microsoft, and Meta have balance sheets that let them prepay, guarantee equipment, and sign long-duration energy deals. CoreWeave offers higher growth exposure, with 2025 revenue of $5.1 billion and 3.1 GW of contracted power, but it also carries greater delivery and financing sensitivity (CoreWeave Form 10-K, FY2025). Investors should separate power ownership from power dependence.
What should PE investors diligence before buying a data center platform in 2026?
PE investors should diligence the queue before the model. That means reviewing interconnection agreements, utility study status, substation scope, transformer delivery dates, capacity accreditation assumptions, water and cooling permits, local tax agreements, and customer contract start triggers. Core Scientific's 2026 filing shows the kind of detail that matters: it had 590 MW of contracted leased customer power capacity and 395 MW actively billing as of June 30, 2026, with conversion depending on equipment lead times, labor, permitting, interconnection sequencing, and customer deployment (Core Scientific Form 10-Q, 2026). The gap between contracted and billable MW is where returns are won or lost. Buyers should also stress customer concentration, since a single AI cloud customer can support financing but also concentrate renegotiation risk if capacity arrives late.
The Megawatt Becomes The Moat
The AI data center market in 2026 is no longer best understood as a race for land, chips, or cloud logos; it's a race for credible megawatts. IDC's $487 billion 2026 AI infrastructure spending forecast shows that capital is available, and Gartner's 565 TWh data center electricity forecast shows that demand is real (IDC, 2026; Gartner, 2026). The constraint is the grid's ability to turn announcements into energized campuses without shifting unacceptable cost or reliability risk onto other customers.
The strategic answer is not a single technology. Renewables will win where land, interconnection, and storage economics work. Nuclear will earn premium pricing where capacity and carbon matter together. Gas will return in regions that value speed and firmness, though emissions and fuel risk will cap its appeal. Demand response, workload shifting, and better AI chips will reduce pressure at the margin, but they won't erase the need for large blocks of firm power.
Executives should watch three things through 2027: which hyperscalers disclose larger energy commitments, which grid operators adopt PJM-style readiness and cost rules, and which announced campuses move from contracted MW to billable MW. The companies that can prove all three will set the clearing price for AI capacity. By December 2027, at least 20 GW of announced U.S. AI data center capacity will be delayed, relocated, or repriced because interconnection and capacity obligations couldn't be secured on the original schedule.
