AI Data Centre Power Demand Is Real, but $170B Has Been Blocked

Goldman Sachs projects U.S. data center electricity demand will hit 66 GW by 2027, more than double today’s 31 GW baseline. Yet over the same window, more than $170 billion in planned data center capacity has already been blocked, withdrawn, or stalled.

Both numbers are true at once, and that contradiction is the defining problem for anyone deciding where to put capital in this space right now. The demand growth is real. So is the structural filtering happening at the grid, the regulator, and the community level.

The gap between announced demand and realised project completions is where the investment risk actually lives. This analysis maps where capital is genuinely moving versus where it is merely being announced, and identifies the specific signals that separate projects likely to advance from those set to stall. Getting that distinction right is now the central question for any energy infrastructure decision tied to AI data center power demand.

The demand signal is real, but the official forecasts are inflated

Start with the credible near-term benchmark. Goldman Sachs Commodities Research, in projections issued 20 May 2026, put U.S. data center power demand at 31 GW in 2025, rising to 41 GW in 2026 and 66 GW in 2027. By 2027, that represents roughly 8.5% of U.S. summer peak demand, and the figures assume capacity utilisation of around 70%.

Extend the horizon and the case for treating this as a decade-long infrastructure theme rather than a passing cycle gets stronger. The Electric Power Research Institute (EPRI) estimated in February 2026 that data centers could consume between 9% and 17% of total U.S. electricity by 2030, potentially reaching 20% by 2035.

Here the picture starts to fracture. A November 2025 report from Grid Strategies, a power sector consulting firm, concluded that the combined data center load forecasts utilities had filed with the Federal Energy Regulatory Commission (FERC) exceeded credible industry estimates for build-out through 2030 by approximately 40%.

The forecasting warning that reframes everything Utility load forecasts filed with FERC ran approximately 40% higher than credible industry estimates, according to Grid Strategies. A significant slice of the data center capacity sitting in utility resource plans may never be built.

That overstatement has a structural cause, not a clerical one. Karl Rábago, principal at Rábago Energy and a former Texas Public Utility Commission commissioner, has argued that vertically integrated utilities carry a financial motivation to project high data center load growth, because infrastructure built to serve that load generates regulated returns for shareholders.

For an investor, the read is direct. Credible demand (hyperscaler-committed, contract-backed) behaves very differently from forecast demand (utility planning assumptions driven by speculative interconnection applications). Energy infrastructure plays tied to the latter carry embedded write-down risk that the headline GW figures conceal entirely. Getting the demand picture right at the source level is the first filter you apply.

For readers wanting the foundational context behind these demand projections, our full explainer on AI data center energy demand covers how the GW figures are constructed, what drives divergence between utility forecasts and industry estimates, and how the 2030 trajectory has been revised across major research sources.

Why $170 billion in announced capacity has already been blocked

The scale of the problem lands before the logic behind it does. According to a June 2026 dataset from Relae (formerly Carbon Direct), 46 announced data center projects across 20 U.S. states were blocked, withdrawn, or stalled between January 2024 and mid-May 2026, representing more than $170 billion in investment.

The pace is the alarming part. Relae counted 6 blocked projects in 2024, 25 in 2025, and more than 20 already by 15 May 2026. PJM-region projects account for a substantial share of the total.

This resistance is not random friction. It tracks predictably to project type, and the community opposition profile explains why. Projects draw the fiercest local pushback where their grid and environmental footprint looks large next to the economic benefit they deliver, particularly where commitments on clean energy, water, and jobs are absent.

The three drivers that consistently trigger opposition:

• A grid and environmental impact profile that outweighs local benefit
• The absence of clean energy procurement and water stewardship commitments
• Limited local job creation relative to the scale of the facility

Relae’s analysts note that projects linked to identifiable hyperscaler investment, and therefore to jobs, local procurement, and clean energy pledges, gain political support far more readily than anonymous speculative sites whose owners aim to flip entitlements once grid access is secured.

How the neocloud model is getting squeezed from both sides

Speculative neocloud operators, firms that lease computing capacity to larger technology companies, face pressure from two directions at once. Community opposition blocks their new sites, while utility financial filters (raised interconnection study fees, structured service agreements, and proof-of-customer requirements) block their grid access.

FERC Order 2023 and PJM’s reformed clustered interconnection process compound the squeeze by prioritising ready projects and filtering out speculative entries with no firm load commitment. Advait Arun of the Center for Public Enterprise put the outcome plainly: only projects with hyperscaler funding upfront are likely to advance through the development pipeline.

FERC Order 2023 interconnection reform introduced milestone payment requirements and cost allocation rules that make speculative queue entries substantially more expensive to maintain, directly reshaping which project types can realistically advance through the approval pipeline.

The acceleration through 2025 and into 2026 tells you the filters are tightening, not loosening. Late-entering speculative developers face a progressively more hostile approval environment than the historical record of announcements would suggest. For an investor, the practical consequence is that announced GW pipelines are not a reliable proxy for capital that will actually be deployed.

The infrastructure gap that the demand numbers do not show

The GW demand figures look clean on the surface. The infrastructure math behind them is where the real story sits, and it starts with the queue.

The national interconnection backlog stands at roughly 2,200 GW, according to S&P Global Commodity Insights cited via Reslink in July 2026. Timelines have not collapsed despite reform. Within PJM’s deregulated territory, the average interconnection wait runs about 8 years, against 1-5 years in vertically integrated regulated markets, according to Brad Viator, president of the utility trade association Power for Tomorrow.

As of June 2026, PJM’s queue tracker showed roughly 118.8 GW total queued capacity, with about 55 GW sitting in the large-load queue. That is a substantial pipeline waiting on approval, financing, and construction.

Now the counterintuitive part. Relae’s modelling, drawing on data from LBNL, EPRI, Goldman Sachs, and the IEA, found that serving a 4 GW data center load on a 100% hourly carbon-free basis requires roughly 9.6 GW of additional clean capacity in ERCOT and about 10.5 GW in PJM.

The multiplier that resizes the opportunity Serving 4 GW of data center load at 100% hourly carbon-free energy requires roughly 9.6-10.5 GW of new clean capacity, according to Relae modelling. The generation build behind the load is roughly 2.5 times the nameplate demand.

That multiplier reframes the investment size. If you are evaluating generation and transmission plays tied to data center demand, the infrastructure opportunity should be sized at roughly 2.5 times the nameplate load, not one-to-one. The companies positioned to close that gap, especially in regulated markets with shorter interconnection timelines, represent a more durable opportunity than data center construction plays on their own.

Clean energy investment trends in 2026 reflect the multiplier dynamic directly: capital is flowing toward utility-scale solar, storage, and transmission rather than data center construction itself, consistent with the 2.5x generation requirement that confirmed hyperscaler load commitments create.

What PJM’s reliability warning means for the build timeline

PJM projects peak demand rising by roughly 32 GW between 2024 and 2030, with about 30 GW of that attributable to data centers, according to filings reported by Power Magazine in August 2026. The grid operator has explicitly flagged that demand commitments can materialise faster than the generation and transmission to serve them can be approved, financed, and built.

The Department of Energy’s repeated emergency grid directives to PJM in 2026 are a live indicator of how far actual grid stress has already outpaced the measured language in regulatory filings, reinforcing the case that the reliability gap the article describes is operational, not merely theoretical.

The pipeline exists but is not yet delivered. PJM’s reformed Cycle 1 process closed with 811 new generation projects totalling roughly 220 GW.

For an investor, that gap is the signal. The demand is real, the infrastructure intent is real, and the space between the two is precisely where reliability risk and investment opportunity sit together.

Where the utility gatekeeping model is creating a durable two-tier market

The filtering mechanisms running through this whole story are not temporary. They are hardening into permanent market architecture, and it helps to name the specific tools utilities now deploy against speculative load:

1. Raising interconnection study fees as application volumes climb
2. Structuring service agreements that stop large customers ramping demand at other ratepayers’ expense
3. Keeping potential large-customer load out of forecasts until a contract is executed or close to it

The read on utility incentives is genuinely contested, and both sides deserve their due. One view, articulated by Rábago and echoed in Relae’s analysis, holds that vertically integrated utilities have regulated-return incentives to encourage large capital programmes and may not rigorously vet load forecasts, exposing ratepayers to stranded-asset risk if demand fails to arrive.

The competing view holds that PJM’s revised forecasting and FERC Order 2023 are grounding the process in firmer commitments. Order 2023 enforces milestone requirements, makes projects bear appropriate upgrade costs, and reduces speculative entries.

What both readings produce, though, is the same practical output.

The balance-sheet gap is stark. Goldman Sachs estimates hyperscalers will collectively spend roughly $581 billion on U.S. AI infrastructure in 2026, giving them the standing to negotiate bespoke grid upgrade arrangements that speculative developers cannot match. Bain and Co. noted in October 2025 that the industry is shifting from an early-stage scramble to a more measured, execution-oriented phase, which implies the filter between credible and speculative projects will only sharpen.

For you as an investor, the gatekeeping architecture works two ways at once. It is a risk signal for speculative-tier exposure and a quality screen for identifying the hyperscaler-anchored projects most likely to drive durable load growth.

What the two-tier split means for energy infrastructure capital right now

The structural finding is now clear. Genuine demand growth concentrates in hyperscaler-anchored projects sited in markets with shorter interconnection timelines and adequate clean generation. Speculative capacity is being progressively filtered out at the grid, the regulator, and the community.

The quantitative frame captures the divide cleanly: roughly $581 billion in hyperscaler AI infrastructure spending against more than $170 billion already blocked. The demand is real and large; the investable slice within it is narrower and more specifically located than the aggregate projections imply.

The AI capacity expansion cycle shares structural features with prior infrastructure build-outs where announced investment consistently outran deliverable supply, a pattern that historically preceded valuation compression in both the developer tier and the equipment supply chain before the genuine demand signal reasserted itself.

To tell durable plays from announced-but-stalled noise, monitor the variables that actually lead outcomes:

• Contract execution, not announcement
• Interconnection queue position and timeline in the relevant market
• Utility service agreement structure
• Community permitting status

The clean capacity multiplier is where the biggest implication sits. Because each GW of confirmed data center load needs roughly 2.5 times that in clean generation to serve it on an hourly carbon-free basis, the generation and transmission opportunity is substantially larger than the data center construction opportunity itself. With EPRI’s 2035 projection of up to 20% of U.S. electricity as the structural ceiling, and PJM’s roughly 30 GW of data center demand growth by 2030 anchoring regulated-market generation investment, this is a multi-year theme. Investors positioned on the infrastructure required to serve confirmed hyperscaler load, rather than on headline construction announcements, sit on the side where capital deployment is structurally more likely to follow through.

This article is for informational purposes only and should not be considered financial advice. Investors should conduct their own research and consult with financial professionals before making investment decisions. Past performance does not guarantee future results, and financial projections are subject to market conditions and various risk factors.