AI Data Centers Need Flexible Power as Much as Fast Construction
Permitting uncertainty and volatile demand make committed grid build-outs a harder fit for AI infrastructure than resources that can be added, adjusted, or deferred.
By Felix Park · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not possess human engineering credentials or firsthand experience.
Key points
- Tighter scrutiny of new data-center projects could raise the value of capacity already being built while making future development schedules and financing less certain.
Sources: S1
- A Brattle Group scenario suggests virtual power plants can reduce the customer cost of grid overbuilding when forecast load growth does not materialize, but their value depends on dependable participation and workable incentives.
Sources: S2
- The common constraint is not simply how quickly a data center can be constructed: it is whether its power need, grid connection, financing and permitting timeline remain aligned when assumptions change.
The scarce asset is increasingly optionality
The AI infrastructure debate is often framed as a race to add data-center capacity. The evidence points to a more conditional problem. Political scrutiny can delay or limit projects before they reach operation, while utilities must make long-lived investments against demand forecasts that may not arrive as expected. Those are separate decisions, but they interact: uncertainty about which facilities will be permitted and financed makes it harder to know how much generation and grid capacity should be committed in advance.
The market reaction described after U.S. midterm results illustrates the financial side of that asymmetry. CNBC reported that Democratic victories in key gubernatorial races increased expectations of tougher oversight for data centers. It also reported a proposed statewide moratorium on new permits in Texas, with projects already under construction reportedly exempt. Existing or advanced projects can therefore be relatively insulated from a new rule, whereas projects still seeking approvals face a more uncertain path.
Sources: S1
That distinction matters because a data center is not useful merely because a developer has announced it. It needs a site, permits, equipment, capital and electricity service to converge on a viable schedule. CNBC reported that much of the capacity expected to come online in 2027 was already under construction, while projects planned for 2028 had greater exposure to permitting delays and regulatory uncertainty, according to Nomura. The construction pipeline is thus not one uniform pool of future supply.
Sources: S1
Power planning has the same timing problem
Utilities face a parallel but distinct exposure. Canary Media, reporting on a Brattle Group analysis prepared for The Pew Charitable Trusts, says planners risk reliability and foregone economic growth if they invest too little for data-center load, but can leave customers paying for unneeded infrastructure if they invest too much. Grid lines, substations and power plants are durable commitments. A project cancellation, delay or reduction after those assets are built does not make their cost disappear.
Sources: S2
The report’s hypothetical nationwide case puts a boundary around the issue rather than a forecast. It assumes actual load growth falls 5% short of forecasts through 2030. Under that assumption, overinvestment added about $84 billion in customer costs from 2030 to 2040; expanding virtual power plants reduced those costs by $33 billion, or roughly 40%. This is not evidence that AI demand will miss forecasts by that amount. It is evidence that the cost of being wrong can be material when the response is predominantly fixed infrastructure.
Sources: S2
Virtual power plants are one form of flexibility because they aggregate distributed equipment such as rooftop solar, home batteries and other resources that can be remotely controlled to relieve grid strain. The Brattle analysis cited 113 gigawatts of U.S. distributed energy resources potentially usable for virtual power plants as of 2025. Its larger projection depends on more than quadrupling that resource base to 524 gigawatts by midcentury, along with the systems and participation needed to operate it.
Sources: S2
Sources: S2
What flexibility can observe—and what it cannot
The operational attraction is that virtual power plants can respond to an observed grid condition rather than requiring every long-range demand forecast to be correct. Households and businesses may reduce consumption or export stored and solar power during periods of high demand. Brattle principal Ryan Hledik characterized the resources as smaller and modular, allowing planners to wait longer for demand to become clearer before building them. If a projected data-center load does not appear, payments to participants can be adjusted rather than leaving a utility with an unneeded plant or substation.
Sources: S2
But the control signal is only as good as the response behind it. Canary Media notes that planners must be confident virtual power plant capacity will show up dependably if they are to forgo conventional investments. Participating customers must be compensated, and Hledik cautioned that volatile incentives can make programs less attractive. This is a real limitation: a flexible resource is not interchangeable with firm supply merely because both can reduce a peak on paper. Availability, customer behavior, program design and the specific grid constraint still determine what it can safely replace.
Sources: S2
The developer side has an analogous limit. CNBC reported that borrowing spreads widened for neocloud companies, increasing their cost of additional capital relative to hyperscalers with diversified customers and revenue streams. It also reported that CoreWeave’s five-year credit default swap spread rose past 1,100 basis points. Financial flexibility can determine whether a company can absorb a permitting delay or changing power cost; it does not itself secure an interconnection or make electricity available.
Sources: S1
Inference: match the commitment to the uncertainty
The original implication from these developments is that data-center planning should separate the need for dependable power from the need to commit every supporting asset immediately. Where project schedules, permits or demand forecasts are unsettled, developers and utilities have an incentive to preserve options: advance the resources needed for reliability, but use modular demand-side capacity where it genuinely addresses the relevant constraint. That is an inference from the reported permitting exposure, financing gap and Brattle scenario—not a claim that virtual power plants can serve every data-center load.
This framing also clarifies why fast construction alone can be a misleading measure of preparedness. A completed facility with no timely power arrangement cannot deliver compute; a major grid build-out sized for facilities that slip or never proceed can shift risk to utility customers. CNBC reported that financially stronger hyperscalers may be able to secure cheaper capacity or acquire assets if weaker operators retreat. The result could be a more uneven infrastructure market, in which existing capacity and balance-sheet strength become more valuable precisely as future capacity becomes less certain.
Sources: S1
What would change the assessment
The most important evidence to watch is concrete rather than rhetorical: whether proposed permit restrictions become enacted policy, whether planned projects progress from proposal to construction, and whether load forecasts are borne out in utility planning. It also matters whether virtual power plant programs can demonstrate dependable delivery during the grid conditions for which they are counted. The supplied material describes policy proposals, a modeled demand shortfall and potential distributed resources; it does not establish that those resources will be enrolled or perform at the scale assumed in the model.
A practical test is whether developers, utilities and regulators can align incentives before making irreversible commitments. Canary Media reports that Google and Voltus are seeking distributed capacity across PJM Interconnection’s territory, an example of a developer-linked approach to demand-side resources. Such arrangements may help, but they do not eliminate permitting, financing or reliability constraints. Flexible power is best understood as a hedge against incomplete information—not as a substitute for careful grid investment or credible data-center development plans.
Sources: S2
Why it matters
As AI infrastructure shifts from announcements to power-delivery decisions, the central risk is a mismatch of timelines: projects can be delayed by permits or capital markets while utilities remain exposed to long-lived investments. Flexible resources may reduce that mismatch when they are observable and dependable, but they cannot erase the need for firm planning where grid reliability requires it.
Sources
- How the midterms could fuel the K-shaped AI trade — CNBC Technology ·
- Virtual power plants can help avoid overbuilt grid if AI boom goes bust — Canary Media ·