Research the degree to which power availability, grid constraints, water cooling capacity, and data center construction…

Power and grid constraints have become the dominant physical bottleneck for AI data center expansion in 2025-2026, often outpacing the pace of chip or networking hardware availability. Hyperscalers can procure GPUs and high-bandwidth memory (HBM) with lead times measured in quarters, but energizing facilities requires grid interconnections that routinely take 5–7 years due to queues, transmission upgrades, and permitting—creating a mismatch where announced capacity far exceeds operational megawatts.[1][2]

• Deloitte’s April 2025 survey of 120 US power and data center executives found grid stress as the leading challenge, with 72% rating power/grid capacity as “very or extremely challenging.”[2]
• As of early 2026, ~190 GW of hyperscale capacity had been announced across 777 projects (~148 GW planned), yet only ~12 GW was operational and ~21 GW under construction.[1]
• US data center power demand is projected to rise from ~80 GW in 2025 to 150 GW by 2028 (Bloom Energy report).[3]
• In Virginia alone, Dominion Energy reported ~70,000 MW of data center requests (roughly triple its system peak load), with 25,000 MW having projected connection dates through 2031.[4]
• Sightline Climate analysis (widely cited in 2026 reporting) projected 30–50% of planned 2026 US data center capacity delayed or canceled due to power, permitting, and related constraints.[5][5]
• Interconnection queues in key regions (PJM, ERCOT) stretch years; high-power transformer lead times have extended to as long as five years (from 24–30 months pre-2020).[5]

This forces hyperscalers and developers to prioritize “speed to power” in site selection, pursue behind-the-meter generation (e.g., natural gas peakers or small modular reactors), or shift to less-constrained but higher-latency or costlier locations. For new entrants or competitors, power access now functions as a de facto moat or barrier to entry in prime markets.

Water cooling capacity is a growing but generally secondary constraint, most acute in water-stressed regions and driving rapid adoption of liquid cooling systems. AI rack densities (projected to reach 50x traditional levels) overwhelm air cooling, increasing reliance on evaporative or hybrid systems that consume significant water—though closed-loop and direct-to-chip liquid cooling mitigate usage while introducing new infrastructure requirements.[1]

• A 100 MW AI data center can use ~2 million liters of water daily (equivalent to ~6,500 households); projections show cooling water demand potentially rising sharply.[6]
• Goldman Sachs forecasts liquid-cooled AI servers rising from 15% in 2024 to 54% in 2025 and 76% in 2026.[7]
• Liquid cooling can reduce site energy use by 25–30% and improve PUE (near 1.1 in best cases), with some closed-loop designs cutting freshwater consumption materially.[8]
• Concerns are heightened in the US West/Southwest and areas with municipal competition; however, power and grid issues are cited far more frequently as the primary limiter in utility filings and analyst reports.[9]

Implication: Cooling adds cost and complexity (especially retrofits), but does not appear to rival power as a nationwide deployment gate. Operators investing early in liquid-ready designs or alternative cooling gain an edge in high-density AI clusters.

Data center construction timelines are extended well beyond the 12–18 month build window by power infrastructure, equipment lead times, and permitting, amplifying the physical constraint. Shell construction is relatively fast, but full energization and fit-out for AI workloads face multi-year delays from transformers, substations, and local approvals.[1]

• Over 25% of projects slated for 2025 online dates were delayed; 30–50% of 2026 pipeline capacity faces similar slippage.[1][10]
• Average global construction costs rose to ~$10.7M per MW by 2025 and are forecast at $11.3M per MW in 2026 (6% increase).[11]
• Between March 2024 and 2025, at least 16 developments were delayed or denied, often due to community opposition or permitting.[1]
• In PJM territory, projects entering service around 2025 averaged more than seven years from initiation to operation (including ~3+ years in queue and ~4 years post-approval).[12]

For competitors: Speed advantages now hinge on pre-positioned power contracts, modular/on-site generation, or greenfield sites with faster utility processes rather than pure construction efficiency.

Physical infrastructure constraints (power/grid primary, water and timelines secondary) rival or exceed networking/memory as binding limits on the pace of energized AI capacity in 2025-2026, though the picture is nuanced by timeframe and specific bottleneck. Hardware supply (HBM, CoWoS advanced packaging, networking interconnects) constrains what can be manufactured and shipped, but without grid power, racks remain unenergized. Some 2026 analyses note a shift toward chips as the tighter near-term limit after earlier power dominance, yet deployment delays remain heavily infrastructure-driven.[13][14]

• Multiple sources (Deloitte, Sightline, utility filings) consistently rank power/grid as the top or among the top challenges for actual buildout.[2][15]
• Chip/memory constraints (e.g., HBM allocation prioritizing AI) affect production velocity, but power queues directly gate when capacity becomes usable.[16]
• Hyperscaler capex remains high (~$650B+ projected for 2026 across major players) despite delays, indicating capital is not the limiter—conversion to operational MW is.[17]

Hyperscalers and utilities are adapting through rate reforms, on-site generation, and new market strategies, signaling that physical constraints will shape AI scaling trajectories through at least 2027–2028. Examples include Dominion’s GS-5 rate class (data centers commit to long-term payments for requested power) and broader shifts toward behind-the-meter solutions.[18][19]

Implications for entrants or competitors: Success requires integrated strategies combining power procurement (PPAs, co-generation, nuclear/SMR partnerships), liquid-cooling-ready designs, and geographic diversification beyond saturated markets like Northern Virginia. Those who treat infrastructure as a core product moat—rather than a commodity—will capture disproportionate share of deployable AI capacity. Additional research into specific 2026 utility integrated resource plans or hyperscaler earnings commentary would further refine regional variances.

---

Recent Findings Supplement (June 2026)

Power availability and grid constraints have emerged as the dominant physical bottlenecks for AI data center buildout in 2025–2026, often rivaling or exceeding networking and memory hardware constraints. Hyperscalers and developers face multi-year delays in securing reliable power, driving shifts to on-site generation, co-location models, and alternative siting—even as physical construction timelines remain comparatively short (12–36 months). Water cooling capacity adds localized pressure, prompting regulatory pushes for closed-loop systems and liquid cooling transitions. These infrastructure realities are reshaping project viability, with 30–50% of planned 2026 U.S. capacity at risk of delay or cancellation.[1][2]

1. Power Demand Surge Creating Acute Grid Strain

U.S. data center power demand is projected to more than double from 31 GW in 2025 to 66 GW by 2027 (Goldman Sachs, May 2026), driven by AI workloads that concentrate massive, constant loads. This pushes data centers' share of national peak summer demand from 4.1% to 8.5%, tightening markets and forcing utilities and developers into reactive planning.[3]

• IEA analysis indicates ~20% of planned data center projects risk delays without grid fixes; transmission build times of 4–8 years and component lead times (e.g., transformers doubled in recent years) compound the issue.[4]
• ERCOT is tracking ~410 GW of large loads seeking interconnection as of March 2026 (~87% data centers), with 198 GW applied in Q1 2026 alone—roughly matching current peak load.[5]
• PJM's 2026 long-term load forecast shows accelerated growth from AI data centers, contributing to capacity market price spikes (e.g., 2026–2027 delivery year clearing at $329/MW vs. prior lows).[6]

Implications: New entrants or competitors must prioritize "speed to power" locations or hybrid models over traditional grid-dependent sites; failure to secure firm power early can strand projects despite available land or capital.

2. Interconnection Delays and Project Slippage

Grid connection queues now routinely exceed physical build times, with average waits of 5–7 years (or more) versus 12–18 months for data center construction itself. Over 25% of 110 projects slated for 2025 online were delayed due to power, permitting, and related issues; similar patterns persist into 2026.[2]

• PJM data shows projects averaging >7 years total to operation, with more time post-approval than in queue; as of early 2026, >21 GW in engineering/procurement and 8.2 GW under construction.[7]
• Sightline Climate tracking (early 2026) of 777 hyperscale projects (>50 MW, announced since 2024) projects 30–50% of 2026 pipeline capacity delayed or canceled.[1]
• FERC-directed reforms (e.g., PJM compliance filings in Feb 2026 for co-located loads) and DOE emergency curtailment orders (May 2026) highlight ongoing strain, including rules for backup generation and expedited tracks.[8][9]

Implications: Developers able to navigate or bypass queues (via policy advocacy, alternative generation, or less-constrained regions like parts of Texas) gain decisive timing advantages; pure reliance on traditional interconnection is increasingly uncompetitive.

3. Shift to On-Site, Co-Located, and Alternative Power

Grid constraints are accelerating behind-the-meter generation, PPAs, and co-location, with hyperscalers accepting added complexity for timeline certainty. On-site power expectations among hyperscalers/colos rose 22% in the six months prior to Bloom Energy’s January 2026 report.[10]

• Hyperscalers (Microsoft, Google, Amazon, others) pursuing nuclear offtakes and large PPAs; Alphabet announced energy innovation acquisitions to support on-site management.[11][2]
• JLL (Jan 2026) notes average grid waits >4 years in primary markets, spurring colocated battery storage and natural gas as bridges (despite sustainability concerns for some tenants).[12]
• PJM proposals (2026) include backstop generation procurement, connect-and-manage frameworks with earlier curtailment options, and expedited tracks for state-sponsored generation.[13]

Implications: Competitors with expertise in on-site generation, financing hybrids, or regulatory navigation can accelerate deployments where grid-dependent players stall; this favors well-capitalized or vertically integrated players.

4. Water Cooling Capacity Adding Localized Pressure

Direct water use for evaporative cooling, combined with indirect use via power generation, is straining resources in high-growth areas, prompting efficiency shifts and new regulations. A 100 MW AI data center can use ~2 million liters daily (on-site portion ~725,000 liters).[14]

• March 2026 state legislation in South Carolina (HB 4583) and Kansas (SB 400) mandates closed-loop systems with zero net water withdrawal/discharge for data centers.[15]
• UT Austin white paper (May 2026) projects Texas data centers could account for 3–9% of state water use by 2040 (cooling + power generation).[16]
• Industry shift toward liquid cooling and no-water systems (e.g., new facilities in Arizona/Wisconsin saving ~125 million liters/year each starting 2026); AWS expanding reclaimed water use.[17]

Implications: Site selection must now incorporate water stress mapping and cooling tech roadmaps; regions or operators adopting closed-loop/liquid cooling early avoid regulatory or community pushback.

5. Construction Timelines and Supply Chain Realities

While physical builds are faster than grid connections, rising costs (global average to $11.3M/MW in 2026, +6% YoY), labor shortages (~439k workers industry-wide), and equipment lead times (transformers, generators) extend effective timelines.[12]

• Data center construction costs rose at 7% CAGR 2020–2025; skilled labor shortages and peak crew sizes (now 4,000–5,000) add friction.[2]
• Supply chain issues (e.g., transformers) cited as key 2026 delay factors alongside power.[18]

Implications: Modular or standardized designs, automation, and early procurement lock-ins help mitigate, but these are secondary to power access for overall speed-to-market.

Overall, physical infrastructure—especially power and grid access—has become a more binding constraint than networking or memory hardware for AI scale-up in the near term. Entities that treat power strategy as a core development input (rather than a downstream utility task) will capture disproportionate share of the buildout. Additional research into specific utility IRPs or hyperscaler earnings calls could further quantify regional variances.