The Power Wall

Can the grid build fast enough for the age of AI — and what happens to the world's largest portfolios if it can't?

The Frontier · Weekly Deep Dive · June 13, 2026

Executive summary

For two years the constraint on artificial intelligence has been narrated as chips. It is increasingly electrons. Across 2025 the electricity drawn by AI-optimised data centres grew by roughly 50%, while total data-centre demand rose about 17%, and the binding bottleneck on the next leg of the build-out is no longer fabrication capacity or capital — it is the ability to connect new load to a power grid that takes more than five years to interconnect and that, in the largest US market, has just cleared its capacity auction at the regulatory price ceiling. The world's largest hyperscalers are now restarting nuclear reactors and building their own gas plants rather than wait in the queue.

This matters to universal owners more than to almost anyone else, because the same institutions that hold the hyperscaler equities at index weight are also the dominant capital behind the physical build-out: sovereign and pension funds committed an estimated US$120bn to AI infrastructure across 2025–2026, and the GIC–CPP Investments–Equinix joint venture alone is sized at US$15bn. A long-horizon owner is therefore exposed to this story twice — once through the companies that need the power, and once through the wires, turbines, reactors and land that must deliver it.

The Desk's question is whether US data-centre electricity consumption reaches 300 terawatt-hours in calendar year 2027 — a level that would confirm the structural demand break is real and not, as in the 2007–2020 flat-demand era, another forecasting head-fake. Our answer, from a 50,000-path Monte Carlo, is about 68% yes, with a median outcome near 316 TWh, an 18% chance of touching roughly 350 TWh (around 8% of all US generation), and only a ~2% chance demand stalls below 250 TWh. The genuinely contested variable is not demand — it is whether the grid can physically deliver it. We assign a 45% probability that interconnection and build constraints bind hard enough in 2027 to cap realised consumption below underlying demand. That gap — forecast demand minus deliverable supply — is the Power Wall, and it is where the investable mispricing sits.

Desk view in one paragraph. The market has correctly identified that power is scarce — PJM capacity prices at the cap, hyperscaler nuclear restarts, and a rush to on-site gas all say so. What it is still under-pricing is duration and delivery: the assets that earn from scarce, dispatchable, grid-connected power over a decade — independent power producers, nuclear and gas-fired baseload, grid equipment, and the long-dated infrastructure debt that funds it — are priced for a cyclical surge, not a structural decade-long capacity shortage. The mirror-image risk is that the demand forecasts are wrong again, as they spectacularly were after 2007, and the same owners end up financing stranded megawatts.

The situation as of today

The numbers that define the problem are no longer speculative. The International Energy Agency's Energy and AI analysis puts global data-centre electricity consumption at roughly 415 TWh in 2024 (about 1.5% of world demand) and projects it to double to around 945 TWh by 2030 in its base case, reaching roughly 1,200 TWh by 2035. The United States and China together account for nearly 80% of that growth; US consumption alone is projected to rise by up to 240 TWh by 2030, a 130% increase over 2024.

In the US the trajectory has already bent sharply upward. The EIA in January 2026 forecast the strongest four-year growth in US electricity demand since 2000, explicitly attributing it to data centres, after a remarkable fifteen-year stretch in which demand was essentially flat. Data-centre consumption stood near 183 TWh in 2024, exceeded 200 TWh in 2025, and is tracking above 250 TWh in 2026 — versus a US total of roughly 4,283 billion kWh expected this year.

The supply side is where the strain shows. PJM Interconnection — the largest US grid operator, covering the mid-Atlantic and parts of the Midwest — cleared its 2026/2027 capacity auction at a record US$329.17/MW-day, the ceiling of the FERC-approved price collar, after the prior year's US$269.92 had itself been a record. PJM's forecast peak load rose by more than 5,400 MW year-on-year, driven largely by data centres, and the auction procured 134,311 MW — short of the reliability requirement. Interconnection queues now stretch beyond five years from request to commercial operation, and the IEA estimates roughly 20% of planned data-centre projects are at risk of delay because of grid constraints.

The response from the largest buyers tells you how serious the shortage is. Every major hyperscaler has now signed at least one nuclear deal; thirteen announced projects commit more than 9.8 GW of nuclear capacity. Microsoft's 20-year agreement to restart Three Mile Island Unit 1 (835 MW, roughly US$1.6bn restart cost; PPA value undisclosed; targeted for the second half of 2027) is the emblem of the era — note the widely-cited US$16bn figure is the project's projected 20-year statewide economic impact, not the contract or capex. Meta has contracted up to 6.6 GW across nuclear developers; Amazon bought a campus adjacent to the Susquehanna nuclear station. And because reactors take years, roughly one-fifth of planned US data centres have begun site work for on-site natural-gas generation — building their own power plants rather than waiting for the grid.

What changed in the last week

Three developments sharpened the picture in the seven days to June 13. First, the World Economic Forum published an analysis arguing explicitly that grid connectivity — not chips, capital or algorithms — is now the strategic bottleneck for AI, crystallising a shift in elite consensus. Second, fresh IEA Electricity 2026 grid analysis reiterated that annual global grid investment must rise about 50% from today's ~US$400bn toward roughly US$600bn by 2030 — and roughly double to over US$800bn a year by the early 2030s — to keep pace, against a backdrop of transformer and high-voltage-equipment shortages. Third, the institutional capital response continued to accelerate, with sovereign and pension allocators deepening digital-infrastructure positions — GIC's move into Goodman's data-centre portfolio and the CPP–GIC Equinix venture among them — confirming that the world's largest owners are treating power-and-compute infrastructure as a multi-decade allocation, not a trade.

Source Ledger

Institutional and primary sources only. Confidence: H (high) / M / L. "Moves model?" indicates whether the data point shifted the Desk's probabilities.

Key Data Table

The forecast question

Will US data-centre electricity consumption reach or exceed 300 TWh in calendar year 2027, as reported by the EIA / DOE?

• Horizon: calendar year 2027 (resolved when EIA/DOE 2027 consumption data is published, ~2028).
• Resolution criteria: YES if reported US data-centre electricity consumption for CY2027 is ≥300 TWh; secondary markers at ≥350 TWh (~8% of US generation) and a downside flag below 250 TWh.
• Base rate: demand-growth-turn forecasts have a poor accuracy record (see prior). Naïve extrapolation of 2024–26 growth clears 300 TWh easily; the binding question is whether the grid can deliver the power the demand implies.
• What success/failure looks like: YES confirms the structural break and validates a decade of dispatchable-power scarcity. NO (stall below 250) would imply either AI demand disappointment or a binding physical Power Wall — both highly consequential for owners.
• What would change the call: a sharp AI capex pause; a step-change in chip efficiency (demand destruction); or, on the supply side, an interconnection-reform breakthrough that releases the queue.

The prior and historical analogues

The single most important discipline here is humility about demand forecasting, because the recent record is humbling. Between 2007 and 2020 US electricity demand was essentially flat — growth of roughly 0.1% a year — even as the economy expanded about 8%. Efficiency gains (LED lighting, better motors, the shift from manufacturing to services) silently cancelled the growth that every utility integrated-resource plan of the 2000s had assumed. Forecasters spent that decade over-building and over-projecting; a generation of merchant gas plants was financed against load that never arrived.

Four analogues frame the prior:

1. The 2000s telecom/dot-com power build (over-forecast). Utilities and merchant generators built for an internet-driven demand surge that efficiency and offshoring erased. Relevance: the canonical warning that a credible technology-demand story can still be wrong. Why it may mislead: AI load is physically concrete and already metered, unlike the speculative 2000s projections.
2. The 2007–2020 flat era (structural break missed in the other direction). Relevance: shows demand can defy decades of "it must grow" priors. Implication: anchor the prior below naïve extrapolation.
3. The 1965–2000 electrification supercycle (sustained structural growth). Relevance: when a genuine new electricity-intensive use arrives (air conditioning, then computing), growth can run for decades. Implication: the AI break could be the start of a long up-cycle, not a blip.
4. The 2021–2025 reshoring + electrification + AI convergence (the turn). Relevance: EIA, PJM and IEA data all confirm the turn has already happened in the numbers. Implication: the question is no longer if demand grows but how fast and whether it can be served.

Net prior on "≥300 TWh in 2027": we start at ~45% — deliberately below naïve extrapolation, weighted by the flat-era cautionary record and by the real possibility that the grid caps realised consumption.

Updating the model (Bayesian)

The Monte Carlo (below) puts the same probability at 67.6%. We report the Desk call as ~68%, reflecting that the demand momentum and the buyers' willingness to self-supply power outweigh the grid-delay drag for the 300 TWh threshold specifically — though the delay drag is exactly what makes the higher 350 TWh threshold a long shot.

Scenarios (MECE, summing to 100%)

Scenarios A and D are the "demand confirmed" states (50% combined); B is the investable mispricing (scarcity priced as cyclical, not structural); C is the tail that the flat-era prior insists we keep on the board.

Simulation results

We ran a 50,000-path Monte Carlo of US data-centre electricity consumption in CY2027 (TWh). Design: a 2026 base of ~250 TWh (lognormal, ±6%); a year-ahead demand-growth rate drawn from a Triangular(6%, 27%, 55%) distribution chosen to span the genuine forecaster disagreement; a chip-efficiency / demand-destruction multiplier (Normal, mean 1.00, sd 6%); and — the distinctive feature — a grid-constraint mechanism: with 45% probability the interconnection/build constraint binds in 2027, applying a 5–30% haircut to the increment of demand over the prior-year base (capping realised consumption below underlying demand). Realised consumption is floored near the prior year (demand rarely falls outright).

Outputs:

• Median: 315.8 TWh; P10 273.7 / P25 292.6 / P75 341.0 / P90 365.6; mean 318.1.
• P(≥250 TWh): 98.4% · P(≥300 TWh): 67.6% · P(≥350 TWh, ~8% of US gen): 18.3% · P(<250 TWh): 1.6%.
• The constraint binds in 45.1% of paths; median uncapped demand is 321.8 TWh, so the Power Wall shaves roughly 6 TWh off the median and far more in the right tail.

Main drivers: the demand-growth draw dominates the centre of the distribution; the grid-constraint haircut dominates the right tail (it is why 350 TWh is only an 18% event despite strong demand). Limitations: the constraint probability and severity are Desk judgement, not estimated from a delivered-vs-requested megawatt panel; the growth distribution is calibrated to published forecast ranges, not to a structural model of compute demand; reporting-definition changes (what counts as "data centre") could move the resolved number independent of physical reality. This is a model output, not a forecast.

Market pricing vs. the Desk view

The market has priced scarcity but not duration. PJM capacity at the cap, gas firming, and US$16bn nuclear restarts all say power is short today. What looks under-priced is the persistence of that shortage: interconnection queues and transformer lead-times mean the imbalance is a multi-year, arguably decade-long, condition — yet much of the dispatchable-power and grid-equipment complex still trades on normalised mid-cycle multiples rather than structural-scarcity multiples.

• Most under-priced: long-lived dispatchable generation (nuclear/SMR, efficient gas) and grid equipment (transformers, HV switchgear) — assets that earn from scarce, grid-connected, around-the-clock power for years.
• Most over-priced relative to delivery risk: hyperscaler equity and data-centre REITs on the assumption that contracted compute can actually be powered on schedule. The Power Wall is a margin and timing risk the market under-weights.
• What consensus is missing: that the binding variable has moved from chips to interconnection — a slower-moving, harder-to-fix constraint than capital.
• What would prove the Desk wrong: an AI capex air-pocket (Scenario C) that turns today's scarcity into tomorrow's stranded megawatts — the 2000s telecom analogue made real.

Universal-owner portfolio implications

This is strategic framing, not personalised advice. The exposure map (visual appendix) scores twelve asset groups across the three live states — build-out delivers, Power Wall binds, demand disappoints.

The structurally favoured exposures across both demand-confirmed and Power-Wall states are dispatchable generation, grid equipment, and natural-gas/midstream — they win whether the grid keeps pace or falls behind, because either way power is scarce and dispatchability is paid for. Long-duration infrastructure debt funding these assets is attractive on a risk-adjusted basis if structured with inflation and merchant-price protection. The most state-dependent exposures are hyperscaler mega-cap equity and data-centre REITs, which are tailwinds only if power is delivered on time and become headwinds under the Power Wall. The clearest tail-loser is anything levered to continued exponential AI demand with no fallback — Scenario C re-runs the flat-era stranding.

For a universal owner specifically, three structural conclusions follow. First, look-through your power exposure: the same fund often owns the hyperscaler (needs power), the utility/IPP (sells power), the infra fund building the data centre (consumes power), and the grid-equipment maker (enables power). Net that book deliberately rather than by accident. Second, treat the build-out as an inflation and reliability story for the whole portfolio: data-centre load is already lifting retail electricity bills and capacity costs, with political and EM-importer spillovers. Third — the universal owner's distinctive lever — stewardship of the build: pushing investee utilities and developers toward firm, low-carbon, grid-additive power (rather than back-up diesel and queue-jumping) is both a systemic-risk and a returns question for an owner that will hold these assets for decades.

Second- and third-order effects

A binding Power Wall does not stay inside the data-centre sector. Capacity-market spikes and on-site gas build feed into retail electricity inflation, already visible at ~17.65¢/kWh residential, with regressive distributional effects and rising political pressure on regulators to socialise or wall-off data-centre costs. Natural gas becomes the swing fuel of the AI era, tightening the link between Henry Hub, power prices and AI economics. Emissions trajectories bend as gas and delayed coal retirements fill the gap, complicating net-zero commitments held by the very owners financing the build. Transmission and transformer supply chains — a global shortage — become a geopolitical and industrial-policy contest. Capital flows intensify: sovereign and pension funds, already ~US$120bn into AI infrastructure, deepen a generational rotation into real-asset, power-adjacent infrastructure, with the second-round financial-stability concern the ECB has flagged for long-horizon balance sheets. And water, land and local-politics constraints layer on top of the electron constraint, making siting a multi-variable optimisation that favours a handful of power-rich regions (the US Sun Belt, the Gulf, the Nordics).

What we're watching

Visual appendix

1. Monte Carlo distribution of 2027 realised US data-centre demand (50,000 paths; P10/median/P90 marked, 300 TWh threshold).
2. Threshold probabilities — ≥250 / ≥300 / ≥350 TWh / <250 TWh.
3. Forecast demand vs deliverable supply, 2024–2035 — the Power Wall gap (Desk-modeled constraint, clearly labelled).
4. Bayesian prior → posterior path — how each evidence item moved the call.
5. Universal-owner exposure heatmap — twelve asset groups × three states.

Red-team: how this could be wrong

The weakest assumption is the 45% constraint probability and its severity — these are Desk judgement, not estimated from a megawatt-by-megawatt delivered-vs-requested dataset, and a reader is entitled to disagree. If interconnection reform moves faster than expected (FERC and state action are already in train), the Power Wall thesis weakens and Scenario A dominates. Conversely, the demand path could be over-stated: the 2007–2020 flat era is a standing rebuke to anyone confident about electricity-demand growth, chip efficiency is improving fast, and a meaningful share of announced AI capex may never be built or monetised — in which case today's scarcity becomes tomorrow's stranded gas plants and the owners financing the build wear the loss. Reporting-definition drift could also move the resolved 2027 number independent of physical reality. The honest posture is that demand growth is highly likely but its level is genuinely uncertain, and the supply side is the swing factor the market is mis-weighting in both directions.

Methodology box

Probabilities are the Desk's subjective estimates, formed by anchoring on documented base rates and updating on institutional evidence (IEA, EIA, DOE, PJM, FERC, WEF) in a Bayesian frame, with the path of updates shown rather than asserted. The simulation is a real 50,000-path Monte Carlo in NumPy with disclosed distributions, a labelled grid-constraint mechanism, and stated limitations; outputs are model outputs, not forecasts. Live energy figures are drawn from EIA and IEA publications dated through June 2026. No facts, quotes, numbers or sources were invented; figures the review process could not stand behind were removed or labelled. Scenario probabilities are MECE and sum to 100%, rounded to 5%.

Disclaimer

This report is for informational and research purposes only and does not constitute investment, legal, tax, or financial advice. Editorial scenario analysis only — not investment, actuarial, or geopolitical advice.