When Builds Outweigh Power: The Coming Grid Crisis Behind AI-Driven Data Centres
Nov 20, 2025
1. The Set-Up: Big Demand on a Slow Supply
Here’s the simple projection: U.S. data-centre load today runs at roughly 35 GW. (BloombergNEF) According to BloombergNEF (BNEF) the figure could swell to 78 GW by 2035 — a jump of +43 GW over 11 years. (BloombergNEF)
My assumptions:
- Baseline (2024) load = 35 GW
- Target (2035) = 78 GW
- Incremental needed = 43 GW over 11 years
Now: utilities must generate, deliver and connect that extra capacity — build new plants, transmission lines, substations, and handle permitting.
2. The Conservative Scenario: “We Keep Up”
If grid build-out keeps pace, then the math is manageable.
- Add ~3.9 GW/year (43 GW / 11) on average
- Demand and supply roughly align
- Outcome: markets tighten (“higher power prices”), but projects energise within normal lead times
In this scenario, bottlenecks exist but aren’t catastrophic: projects move, grid expansions happen, and risk remains moderate.
3. The Stress Scenario: “Build-Out Lags Behind”
Here’s where things get dangerous. Big quoted pipelines of contracted load already exceed what utilities can handle. For instance, Dominion Energy cited ~40 GW in various contracting stages.
- If only 10 GW is energised in the next 3 years, backlog = 30 GW
- Against the incremental need of 43 GW, backlog ≈ 70% (30/43 ≈ 0.6976)
Outcome: large volumes of finished “shell” capacity with no power, bids rising for energisation slots, legal/regulatory fights, and projects shifting to regions with spare capacity (often not ideal).
4. Practical Implications
Short term (next 1-3 years):
- More “dark shells” of data-centres: buildings finished but not fully energized
- Longer connection timelines, bidding wars for power grids, and premium pricing for “energised racks”
- More regulatory friction: utilities push back, transmission policy gets spotlighted
Medium term (3-7 years):
- Utility capex skyrockets: transmission upgrades, substations, generation capacity
- Permitting and interconnection become strategic bottlenecks: who gets served first will matter
- Data-centre site‐selection shifts strongly toward regions with spare capacity, cheaper power intrusions
Longer term (by 2035):
- If BNEF demand materialises and build-out doesn’t accelerate materially, you have a structural gap between demand and supply
- Result: persistent high power prices, grid reliability risks, and geographically skewed power availability
Nice — I’ll do this in two parts: (A) a compact regional heat-map table you can drop into the story, with numeric anchors and sources; (B) exactly where to insert the visual (and a caption + short explainer) so the map strengthens the narrative.
I kept the scoring rule simple and reproducible: heat score = backlog / incremental need + queue pressure + recent contract growth, scaled into four bands. Sources follow each row so you can cite the map.
Regional heat map (table for the choropleth)
“Regional friction isn’t evenly distributed. The map below shows where interconnection queues, contracted pipelines, and recent contract growth combine to create acute power-to-build mismatches. Red zones are the highest risk: finished shells but no wires.”
Legend: Red = Extreme risk, Orange = High, Yellow = Moderate, Green = Low
Region (operator / market) | Heat score (qual) | Key numbers (digit-by-digit) | Why it’s hot (short) | Source |
| Northern Virginia / Dominion (PJM) | Red | Dominion contracted ≈ 40,000 MW (40 GW) of data-centre capacity under contract (Feb 2025); incremental US need = 43,000 MW. Dominion backlog ~30,000 MW example = ~70% of US incremental need. | Massive contracted pipeline concentrated in a single utility; substation & transmission constraints; big capex plans but long lead times. | Dominion / Data Center Dynamics; Dominion testimony. (Datacenter Dynamics) |
| Texas (ERCOT + Permian areas) | Red | ERCOT tracked interconnection requests from 56 GW (Sep 2024) to ~205 GW in 2025 (requests, not guaranteed builds); Texas data-centre demand projected to 14.5 GW by 2030 (451 Research). | Extremely high requests vs existing peak (~85.5 GW); Permian region power shortage; competition with oil & gas electrification. | Texas Tribune; S&P/451 Research; FT. (The Texas Tribune) |
| PJM (broader Mid-Atlantic / Midwest) | Orange → Red | PJM interconnection timelines rose from <2 years (2008) to >8 years (2025); large queue/backlog that delays generation and connection. | Long study & queue times create bottlenecks; many data-centre requests concentrate here. | RMI / PJM commentary. (RMI) |
| California (CAISO) | Yellow → Orange (localized hot spots) | CEC projects data-centre load +2.3 GW by 2030, +3.3 GW by 2035; CAISO stressing interconnection reform & local constraints. | Smaller absolute GW growth than VA/TX, but tight local transmission, land constraints, and permitting friction. | CAISO / CEC. (caiso.com) |
| MISO / SPP (Midwest, central) | Yellow | MISO interconnection queue large (hundreds GW of generation proposals), but load requests for data centres less concentrated; queue churn high. | Generation queue backlog could slow clean supply arrival for new loads; regional capacity more distributed. | MISO / RTO Insider. (RTO Insider) |
| Southeast (Duke, Southern, non-Dominion areas) | Yellow → Orange | Pockets of data-centre demand; utility capex rising (Duke, others) but constrained transmission corridors. | Growing pipelines; regulatory landscape variable by state. | S&P Global / utility filings. (S&P Global) |
| Other (Midwest small clusters, New England) | Green → Yellow | Smaller absolute GW requests; interconnection still slow but less concentrated data-centre demand. | Less immediate risk, but still vulnerable to local permitting and transmission timelines. | BNEF / regional notes. (BloombergNEF) |
“Heat map of U.S. data-centre power risk (2024–25 inputs): red = extreme backlog/contract pressure; orange = high; yellow = moderate; green = managed. Sources: BNEF, Dominion, ERCOT, PJM, CAISO, Data Center Dynamics.”
5. Why This Matters for Markets & Infrastructure
- Valuation risk: Data-centre developers assume power will be available; if it isn’t, cost overruns and delays erode returns.
- Utility investment risk: Big capex ahead — regulatory regimes, cost recovery, stranded asset risk.
- Geographic risk: Regions with constrained grids get bid up, site diversification becomes more important.
- Energy pricing risk: Higher delivered cost of electricity reduces margins for hyperscalers and colocation providers.
6. Summary
The math is clean, the risk evident: 43 GW of incremental load by 2035 for U.S. data-centres, per BNEF.
If the grid keeps up at ~3.9 GW per year, the market tightens but remains sustainable.
If build-out lags, we face delays, stranded shells, premium power pricing and strategic risk.
In short: the build-boom may outrun the power-boom, and the mismatch is where the risk lies.
