U.S. data center power supplied to hyperscale, leased, and crypto-mining facilities reached about 64.4 GW in 2025, nearly triple the 2020 level, according to 451 Research, part of S&P Global Energy Horizons. That single number explains why 2026 hyperscale capacity planning has moved from real estate selection to power origination, interconnection risk, and utility negotiation. The limiting input for AI campuses isn't land, fiber, or even Nvidia supply in isolation. It's deliverable megawatts with a credible in-service date.
The market has entered a phase where a signed power contract can be worth more than a cheaper site. AI clusters want dense, round-the-clock electricity, often at a scale that looks more like industrial load than conventional cloud growth. BloombergNEF's July 2026 forecast puts 2030 installed U.S. data center capacity at 118 GW, up 52% from its December 2025 view, which means capacity planning models built in 2024 are already stale (BloombergNEF, 2026). The grid, however, isn't adding firm supply at the same speed. Lawrence Berkeley National Laboratory reported 2,061 GW of generation and storage still active in U.S. interconnection queues at year-end 2025, with median queue-to-commercial-operation duration above five years for projects built in 2025 (LBNL and GridTracker, 2026).
That timing mismatch is reshaping commercial behavior. Microsoft, Amazon, Google, Meta, Equinix, Digital Realty, Talen Energy, Brookfield Renewable, and utilities such as TVA and Dominion Energy are no longer treating power purchase agreements as sustainability paperwork. They are using PPAs, nuclear uprates, behind-the-meter designs, and utility tariffs as capacity weapons. For readers tracking digital infrastructure more broadly, MarketIntel sees this as the most important infrastructure constraint in AI through 2027.
$788 Billion Chases Megawatts
Gartner forecasts worldwide data center systems spending at $787.99 billion in 2026, up 55.8% from $505.63 billion in 2025. That is the broad spend pool around servers, storage, and data center hardware, and it defines the total addressable market for the infrastructure stack being pulled by AI (Gartner, April 2026). IDC narrows the lens to AI infrastructure: it reported $318 billion of global AI infrastructure spending in 2025, projected $487 billion in 2026, and forecast more than $1 trillion by 2029, implying roughly a 31% five-year CAGR from 2025 (IDC, April 2026). These estimates cluster around a simple conclusion: AI infrastructure has moved from a fast growth segment into one of the largest capital absorption markets in technology.
The serviceable market for power-constrained data center capacity is smaller but more strategically scarce. Goldman Sachs Research estimated global data center power use at about 55 GW in 2023 and forecast 84 GW by 2027, with AI rising from 14% of demand to 27% (Goldman Sachs Research, 2025). By 2030, Goldman expects global data center power demand to rise as much as 165% from 2023, while U.S. data centers could consume 8% of national power compared with 3% in 2022 (Goldman Sachs Research, 2024 and 2025). BloombergNEF is more aggressive on U.S. physical capacity, forecasting 118 GW installed by 2030 and 194 GW by 2035 (BloombergNEF, 2026).
Segmentation now matters. IDC's 2025 AI infrastructure spending was almost all compute-heavy: servers represented $87.7 billion of the $89.9 billion spent in Q4 2025, or 97.6% of quarterly AI infrastructure spend (IDC, 2026). The United States accounted for $69.2 billion in Q4 2025, equal to 77% of the global total, while China fell 8.1% year over year to $8.4 billion under export-control pressure (IDC, 2026). The Middle East and Africa grew more than 500% year over year to $1.8 billion, which shows sovereign AI campuses are emerging but still sit far below the U.S. scale.
The historical baseline is the contrast that makes 2026 different. Goldman Sachs noted that data center workloads nearly tripled between 2015 and 2019 while power demand stayed roughly flat near 200 TWh annually, helped by efficiency gains (Goldman Sachs Research, 2024). That era has ended. AI rack densities, liquid cooling, and high accelerator use have converted compute growth into electric load growth. The inflection isn't a spending curve alone, it's the point at which procurement teams must secure generation and transmission capacity years before models or enterprise demand are fully proven.
The Players Buying Scarcity
Microsoft is acting like the best-connected buyer in the power market because Azure growth now depends on booked energy as much as booked cloud revenue. Microsoft reported fiscal 2025 revenue of $281.7 billion, up 15%, and said Azure surpassed $75 billion in revenue, up 34% (Microsoft annual report, FY2025). Its most visible power move remains the Brookfield framework, announced in 2024, that gives Microsoft a path to more than 10.5 GW of new renewable capacity between 2026 and 2030 across the U.S. and Europe, a structure designed to match data center growth by region rather than buy generic certificates (Microsoft and Brookfield, 2024).
Amazon has become the nuclear-linked hyperscaler to watch. AWS sales reached $128.7 billion in 2025, up 20%, with AWS operating income of $45.6 billion (Amazon company filings, FY2025). In June 2025, Talen Energy expanded its PPA with Amazon to provide up to 1,920 MW of carbon-free nuclear power from Susquehanna through 2042, with full quantity expected no later than 2032 and possible acceleration (Talen Energy SEC filing, 2025). That deal converted a disputed behind-the-meter concept into a front-of-meter contract that still gives AWS a rare line of sight to gigawatt-scale firm power.
Google is using advanced nuclear as an option portfolio, not a near-term capacity fix. Alphabet reported 2025 revenue of $402.8 billion, Google Cloud revenue of $58.7 billion, and Google Cloud operating income of $13.9 billion (Alphabet annual report, FY2025). Google signed a 2024 agreement with Kairos Power for up to 500 MW of advanced nuclear capacity by 2035 and in August 2025 joined TVA and Kairos in a 50 MW Hermes 2 PPA targeting 2030 service for Tennessee and Alabama data centers (Google and Kairos Power, 2024 and 2025).
Meta is moving through balance sheet scale and long-dated leasing commitments. Meta disclosed $69.7 billion of 2025 property and equipment purchases, largely tied to servers, data centers, and network infrastructure, and guided to $115 billion to $135 billion of 2026 capital expenditures to support AI and core business (Meta filings, FY2025). Its roughly $103.8 billion of not-yet-commenced lease obligations, mostly for data centers, colocations, and network infrastructure, shows that Meta is locking capacity before all AI monetization paths are visible.
Equinix remains the premium interconnection platform, but it is now being pulled toward higher-power deployments. Equinix reported 2025 revenue of $9.2 billion and 280 data centers across 77 markets, including 23 xScale facilities (Equinix annual report, FY2025). In 2025 it opened 16 data centers and had 52 major development projects under way as of January 2026, expected to deliver more than 55,000 retail cabinets and over 100 MW of xScale capacity through 2028. Its land acquisitions could support roughly 1 GW of future retail and xScale capacity, which is a direct answer to power-per-cabinet inflation.
Digital Realty is competing through land bank, pre-leasing, and megawatt inventory. It reported $6.0 billion of rental and other services revenue in 2025 and said its portfolio was 84.7% leased at year-end (Digital Realty filings, FY2025). The company had 769 MW of projects under way across global metros, 64% pre-leased, and estimated its land and space held for development could support more than 3,500 MW of added capacity, including over 1,000 MW in Northern Virginia.
The share gain is going to players that can turn power uncertainty into contracted supply. Hyperscalers with investment-grade balance sheets are gaining first call on nuclear, renewable, gas, and utility-backed capacity because developers need bankable offtake. Among operators, the winners are Equinix and Digital Realty where they control scarce powered shells, while smaller developers without interconnection progress are being pushed into joint ventures or site sales.
Queues Become Strategy
The structural trigger is FERC Order No. 2023, paired with the visible failure of serial interconnection processes to handle AI-era load growth. The rule replaced much of the old first-come, first-served model with a first-ready, first-served cluster study process, increased study deposits, required site control, added commercial readiness deposits, and imposed withdrawal penalties when exits harm other projects (FERC, 2023). In plain terms, speculative projects lose room, and capital-ready projects with land, deposits, and offtake get better odds.
That matters in 2026 because the queues are still immense even after reform. LBNL and GridTracker counted about 8,200 active U.S. projects seeking interconnection at year-end 2025, representing 1,312 GW of generation plus 749 GW of storage (LBNL, 2026). Only 13% of capacity that submitted interconnection requests from 2000 to 2020 had reached commercial operation by the end of 2025, while 75% had withdrawn (LBNL, 2026). A hyperscaler can't base a 2027 AI deployment plan on a power project that may never survive the queue.
PJM has become the clearest stress test. Its 2027/2028 capacity auction cleared at the FERC-approved cap of $333.44 per MW-day, yet procured capacity plus fixed-resource commitments fell 6,517 MW short of PJM's reliability requirement, the first such shortfall since the auction structure began in 2007 (PJM annual report, 2025). PJM has said data centers are the primary driver of demand growth and can be developed two to three times faster than many generation technologies needed to serve them (PJM, 2026). That mismatch has turned interconnection timing into the gating item for hyperscale planning.
Three Risks Are Mispriced
The first risk is contract timing slippage, with a roughly 60% probability that at least several large U.S. AI campuses announced for 2027 or 2028 miss initial power delivery schedules. The mechanism is simple: generation projects can have offtake, land, and equipment orders yet still wait for network upgrades, transformer supply, permitting, or capacity accreditation. The affected players are hyperscalers, wholesale colocation developers, and utilities in PJM, ERCOT, MISO, and the non-ISO West. The timeline is immediate through 2028 because today's capacity plans are being written against queue data that still shows a median above five years from request to commercial operation for built projects (LBNL, 2026).
The second risk is power cost pass-through, with a roughly 50% probability that enterprise cloud buyers see AI infrastructure pricing absorb more locational energy cost by 2027. PJM's capacity price movement from $28.92 per MW-day for 2024/2025 to $269.92 for 2025/2026, then capped near $329 to $333 for later delivery years, shows how tightness can move into market prices (PJM, 2024 and 2025). Affected players include CIOs signing reserved AI instances, SaaS vendors buying inference capacity, and private equity owners underwriting margin expansion from automation.
The third risk is regulatory backlash, with a roughly 40% probability that at least five large U.S. utility territories impose stricter large-load requirements, special tariffs, or milestone rules before the end of 2027. The mechanism is political: residential and industrial customers resist paying for grid upgrades tied to speculative data center load. Utilities will ask large-load customers for deposits, minimum bills, curtailment terms, or direct contributions to network upgrades. That hits developers with weak balance sheets first, but it can also slow hyperscaler campuses in fast-growth counties.
The tail risk most analysts are underweighting is stranded power procurement. If model efficiency improves faster than inference demand or if AI revenue fails to cover the capital cycle, a hyperscaler could hold long-duration power obligations tied to campuses that ramp slowly. Probability is lower, roughly 20% through 2029, but the financial effect could be severe because PPAs, leases, and network commitments can outlast the first generation of GPUs.
Enterprise Buyers
Enterprise buyers need to treat AI capacity as a location-sensitive input, not a generic cloud SKU. CFOs should ask cloud providers where training and inference workloads will run, whether the region sits in a constrained power market, and how energy charges or capacity scarcity may appear in committed-use pricing. CTOs should split workloads by latency and energy sensitivity: latency-light batch inference can be routed to cheaper regions, while regulated or low-latency workloads may justify higher-cost local capacity.
Procurement teams should also negotiate performance and availability credits that reflect power risk. A standard cloud service credit doesn't compensate a business when a model rollout slips because the reserved cluster isn't energized. Buyers with large AI commitments should ask for region substitution rights, price protection if energy adders rise, and transparency on whether the provider relies on grid supply, on-site generation, or contracted PPAs.
Investors
Investors should underwrite powered capacity, not announced capacity. The key diligence question is whether a project has an executed interconnection agreement, utility service agreement, substation plan, transformer slot, and credible generation source. LBNL's finding that only 13% of 2000 to 2020 requested capacity reached commercial operation by year-end 2025 is a warning against valuing early-stage queue positions like finished assets (LBNL, 2026).
Public-market investors should separate cloud demand from power conversion. Microsoft, Amazon, Alphabet, and Meta can fund multi-year procurement programs, but their return spread will depend on how much AI revenue is produced per megawatt. Infrastructure investors should prefer developers with pre-leased capacity, balance sheet partners, and utility relationships. Digital Realty's 64% pre-leased development activity and Equinix's global power sourcing footprint are qualitatively different from merchant land positions marketed as future AI campuses.
Vendors
Vendors selling into data centers should target the bottleneck layer. Switchgear, transformers, liquid cooling, backup generation controls, grid software, power forecasting, and demand-response tools have better pricing power than generic shells. Equipment suppliers should reserve production slots for customers with executed offtake and financing because queue withdrawals can turn soft backlog into noise.
Power vendors should design products around reliability and accounting at the same time. Hyperscalers need hourly matching, emissions attributes, capacity value, and regulatory defensibility. That favors portfolios that combine renewables, storage, nuclear, gas peakers, and demand flexibility, with contracts explicit about deliverability rather than annual certificate matching.
The 2027 Constraint Test
The base case, at 55% probability, is that U.S. AI data center capacity keeps expanding through 2027 but at a visibly slower pace than announced pipelines imply. BloombergNEF's 118 GW U.S. 2030 installed capacity forecast is plausible only if utilities, developers, and regulators convert a large share of current projects into powered campuses (BloombergNEF, 2026). Under this base case, hyperscalers keep signing large PPAs, more nuclear and gas-linked deals appear, and premium capacity concentrates in PJM, Texas, the Southeast, and selected Midwest markets.
The contrarian view, at 25% probability, is that power scarcity becomes less binding by late 2027 because AI chip efficiency, workload scheduling, and model distillation lower megawatts per unit of useful output. This doesn't mean electricity demand falls. It means hyperscalers get more revenue and model capacity from each megawatt, which would reduce panic buying for marginal sites and hurt developers holding weak power positions.
The downside scenario, at 20% probability, is a permitting and affordability backlash. In that case, utilities win approval for large-load tariffs, counties pause projects, and merchant data center developers face higher deposits and slower service dates. The leading indicators are PJM and ERCOT large-load forecast revisions, executed interconnection agreements in LBNL data, and the spread between announced data center capacity and utility-confirmed service dates. Capacity auction prices, transformer lead times, and utility filings for special data center tariffs should be tracked monthly.
Seven Executive Takeaways
- Power access is now the main constraint on AI data center deployment, with S&P Global reporting U.S. supplied data center power at about 64.4 GW in 2025.
- Gartner's $787.99 billion 2026 data center systems forecast shows hardware spending is scaling faster than grid delivery capacity.
- IDC's $487 billion 2026 AI infrastructure forecast implies cloud buyers should expect AI pricing to reflect power scarcity, not only GPU cost.
- Amazon's 1,920 MW Talen nuclear PPA is the clearest template for firm, long-duration hyperscaler power procurement.
- Google's 500 MW Kairos pathway is strategically important, but it won't solve near-term 2026 or 2027 capacity shortages.
- Interconnection queue positions without executed agreements, deposits, and utility upgrade plans deserve steep valuation discounts.
- PJM's capacity market is the warning signal: data center demand can arrive faster than accredited supply, forcing price caps, tariffs, or delayed load.
Questions Capital Committees Are Asking
Question: Should a CFO approve a three-year AI cloud commitment when power costs are moving this quickly?
A CFO should approve only if the contract separates compute price, energy exposure, region rights, and service commitments. AWS reported $128.7 billion of 2025 sales and $45.6 billion of operating income, so large cloud providers can absorb some volatility, but they won't carry all power risk forever (Amazon filings, FY2025). PJM's capacity prices show the issue: the RTO-wide clearing price moved from $28.92 per MW-day for 2024/2025 to $269.92 for 2025/2026, then later auctions hit FERC-approved caps above $329 (PJM, 2024 and 2025). Buyers should request region substitution rights, caps on energy-related adders, and disclosures on whether reserved AI capacity is tied to constrained grids. A lower headline GPU price can be poor economics if the provider later throttles availability or steers workloads to higher-latency regions.
Question: Which hyperscaler has the strongest power position for AI growth?
There isn't one universal winner, but Amazon has the most concrete near-term firm-power signal because the Talen deal covers up to 1,920 MW of nuclear power through 2042, with full volume expected no later than 2032 (Talen Energy filing, 2025). Microsoft has the broadest renewable framework through Brookfield's more than 10.5 GW pathway from 2026 to 2030, but renewable PPAs still need grid deliverability and hourly matching to serve AI clusters (Microsoft and Brookfield, 2024). Google has the strongest advanced nuclear option through Kairos, yet its 50 MW Hermes 2 first deployment targets 2030, so it's more strategic than immediate. The practical answer is regional. The best provider for a buyer is the one with firm capacity in the required geography, not the largest global PPA headline.
Question: Are data center developers overbuilding, or is the market still short?
The market is short on powered capacity and at risk of overbuilding unpowered announcements. BloombergNEF raised its 2030 U.S. installed data center capacity forecast to 118 GW in July 2026, but it also noted a widening gap between AI chip-implied demand and what energy constraints allow to be built (BloombergNEF, 2026). Digital Realty's 769 MW of projects under way with 64% pre-leasing is much higher quality than speculative land because customers have already committed to much of the capacity (Digital Realty filings, FY2025). LBNL's interconnection data is the caution: most proposed generation capacity historically withdraws before operation. Investors should ask whether a site has utility-confirmed load service, network upgrades, and equipment slots. Without those, a project is an option on future power, not a data center asset.
Question: Do nuclear PPAs solve the AI data center energy bottleneck?
Nuclear PPAs help with firm, carbon-free supply, but they don't solve the bottleneck at 2026 scale. Amazon's Talen agreement is meaningful because it uses an existing nuclear asset and reaches up to 1,920 MW over time, with the contract running through 2042 (Talen Energy, 2025). Google's Kairos pathway is important for the 2030s, with up to 500 MW by 2035 and a 50 MW Hermes 2 project targeted for 2030 through TVA (Google and Kairos, 2025). The constraint is timing. New nuclear takes years, while data center shells and GPU clusters can be planned much faster. Through 2027, the practical supply stack will still include grid power, renewables, batteries, gas, demand response, and selective nuclear uprates rather than new reactors alone.
Question: What should a PE investor diligence before buying a data center platform?
A PE investor should diligence power rights before revenue multiples. The required file set includes executed utility service agreements, interconnection study status, upgrade cost allocation, transformer procurement, backup generation permits, water approvals, and customer contracts tied to megawatts rather than square feet. Equinix reported 280 data centers and more than 500,000 interconnections in 2025, which gives it a demand base and utility credibility many private platforms lack (Equinix annual report, FY2025). Digital Realty disclosed over 3,500 MW of potential additional capacity in its development holdings, but even there the investor must separate developable potential from energized capacity (Digital Realty filings, FY2025). A platform with lower EBITDA but firm power can be worth more than a faster-growing developer whose sites sit behind uncertain grid upgrades.
The Scarcity Premium Widens
AI data center planning in 2026 is no longer a race to announce campuses, it's a race to prove deliverable electricity. The winners will be the companies that convert capital into firm power with credible dates, not the ones with the largest press releases. Gartner, IDC, Goldman Sachs, BloombergNEF, S&P Global, LBNL, FERC, and PJM all point to the same operating reality from different angles: AI demand is rising faster than the grid can absorb without contract discipline, queue reform, and new utility economics.
For enterprises, the action is to make AI procurement location-aware and to price energy risk before signing multi-year commitments. For investors, the action is to value executed power rights, not announced megawatts. For vendors, the action is to sell into the constraint layer: cooling, grid equipment, scheduling software, and capacity-backed energy products. The market won't wait for a clean planning cycle. By August 2027, at least three major U.S. hyperscale AI campus announcements from 2025 or 2026 will be publicly delayed, resized, or moved because utility-confirmed power arrives later than the original capacity plan.
Sources include Gartner, IDC, BloombergNEF, Lawrence Berkeley National Laboratory, FERC, PJM, and company filings.