In June 2026, Reuters reported that parts of the Netherlands were expanding electricity connection restrictions because the grid could no longer keep pace with demand. New residential connections in Utrecht were being limited. Businesses were waiting years for service. Data centers, electrification, renewable generation, and economic growth were colliding with the physical limits of infrastructure.¹
Most readers will interpret this as an electricity story. It is not. It is a decision-making story. The most important question in the AI economy is not how much electricity artificial intelligence will consume. It is who will make infrastructure commitments before they know the answer.
Across utilities, developers, investors, regulators, and technology companies, leaders are increasingly being asked to commit capital, approve projects, allocate risk, and shape infrastructure systems years before demand forecasts can be proven. A utility may commit to transmission upgrades expected to last forty years based on demand projections extending only a few years into the future. A developer may secure land and permits before power delivery is certain. An investor may underwrite infrastructure based on assumptions about AI adoption, chip efficiency, customer behavior, and regulatory treatment that remain deeply uncertain.
The challenge is not demand uncertainty. The challenge is commitment uncertainty. That distinction matters because most discussions about AI infrastructure focus on forecasting risk. Utilities, investors, and policymakers are confronting something more consequential: commitment risk. The defining question is not whether forecasts will prove imperfect. They always do. The defining question is which commitments should be made before uncertainty is resolved and which commitments should wait.
The organizations that thrive in the AI economy will not necessarily be those with the best forecasts. They will be the organizations that make the best decisions while the evidence remains incomplete.
The Evidence Is Arriving After the Decisions
The International Energy Agency projects that global data center electricity consumption could more than double to approximately 945 TWh by 2030, roughly equivalent to Japan's current electricity consumption.² EPRI estimates that U.S. data centers could consume between 9 percent and 17 percent of national electricity by the end of the decade, depending on AI adoption, efficiency improvements, and deployment patterns.³
These forecasts matter, but they are not the central challenge facing decision-makers. Electricity systems do not experience demand globally. They experience demand at specific substations, transmission corridors, distribution feeders, water basins, and utility service territories. A region may have adequate generation capacity in aggregate while still being unable to serve another 300 MW campus. A utility may have sufficient energy resources but lack transformer availability. A state may support economic development while struggling to deliver infrastructure on the timeline customers expect.
The result is a dangerous zone where leaders must act before the evidence has matured. Waiting for certainty can mean losing investment opportunities. Acting too early can strand capital, raise customer costs, create reliability risks, or lock regions into poor infrastructure decisions.
Across utilities, developers, investors, policymakers, and technology companies, the same decision framework increasingly appears.
The AI-Energy Decision Stack
The challenge is not forecasting AI demand. The challenge is making infrastructure commitments before demand is fully known.
| Decision Test | The Executive Question | Failure Mode | Signal from the Market |
|---|---|---|---|
| Demand Credibility | Is the demand real enough to justify commitment? | Infrastructure is built for load that never materializes or arrives years late. | Microsoft's decision to slow or re-phase portions of planned data center expansion illustrates how demand assumptions evolve as markets change. |
| Grid Deliverability | Can the system actually serve the load? | Demand exists, but power cannot reach the customer when and where it is needed. | Dutch grid congestion demonstrates the difference between demand existing and infrastructure being able to serve it. |
| Cost Causation | Who bears the risk if assumptions prove wrong? | Customers, investors, utilities, or taxpayers inherit costs they never agreed to bear. | The rise of large-load tariffs reflects growing concern about allocating AI infrastructure costs fairly. |
| Flexibility | How much uncertainty can be absorbed before new infrastructure is required? | Organizations make irreversible commitments before uncertainty is resolved. | ERCOT's focus on flexible large loads demonstrates how optionality can defer costly infrastructure investments. |
| Resource Constraints | What becomes scarce after power is secured? | Projects stall because leaders solved the wrong constraint first. | Water availability is becoming a defining siting consideration across parts of the Southwest. |
| Institutional Durability | Will the decision survive regulatory and public scrutiny? | Technically viable projects become politically or regulatorily vulnerable. | Ireland's efforts to balance data center growth, grid reliability, and climate objectives illustrate the governance challenge. |
| Strategic Value | Is the opportunity worth the infrastructure risk? | Capacity is built without creating durable advantage. | Northern Virginia became the world's largest data center cluster because multiple constraints were solved simultaneously. |
The Shift from Forecasting Risk to Commitment Risk
For decades, infrastructure planning focused on predicting future demand. AI is changing the problem. The central challenge is no longer forecasting demand accurately. It is determining which infrastructure commitments should be made before demand is fully known—and who bears the consequences if assumptions prove wrong.
Test One: Demand Credibility
One of the most common mistakes in the AI economy is treating requested capacity as equivalent to actual future demand.
A hyperscale developer may request 500 MW of service. An investor presentation may describe a gigawatt-scale pipeline. An economic development agency may announce a transformative project. None of these statements necessarily mean that the load will arrive on schedule, operate at projected utilization levels, or persist long enough to justify the infrastructure required to serve it.
Electricity infrastructure is built differently than software. Transmission upgrades, substations, transformers, and generation resources often require large upfront commitments with recovery periods measured in decades. The consequences of overbuilding and underbuilding are both significant.
Recent reports that Microsoft slowed or re-phased portions of planned data center expansion provide a useful reminder that even the most sophisticated organizations continuously reassess infrastructure requirements as markets evolve.⁴ The lesson is not that forecasts are wrong. The lesson is that demand credibility must be evaluated separately from demand potential.
The organizations that succeed will understand the difference between requested capacity, contracted capacity, energized capacity, and utilized capacity.
Test Two: Grid Deliverability
Even credible demand does not guarantee deliverable demand. The Netherlands illustrates this distinction clearly. Grid operators faced growing congestion because infrastructure expansion could not keep pace with demand growth.¹ The challenge was not the absence of power. It was the inability to move power where and when customers required it.
Generation availability and deliverability are not the same thing. A region can possess sufficient energy resources while lacking transmission capacity, transformer availability, substation upgrades, or interconnection capability.
For AI infrastructure, the relevant question is rarely whether power exists somewhere in the system. The relevant question is whether power can be delivered to a specific location, at a specific time, under acceptable reliability conditions.
Test Three: Cost Causation
The third test is fundamentally economic. Who pays if the forecast is wrong?
This question sits at the center of growing debates over large-load tariffs, data center service agreements, and utility investment strategies. As AI infrastructure expands, utilities and regulators increasingly face pressure to determine whether existing customers, future customers, developers, or investors should bear the costs associated with new infrastructure.
The rapid expansion of large-load tariff proceedings across the United States reflects a broader effort to answer this question.⁵ Minimum contract terms, collateral requirements, phased service structures, and exit fees all represent attempts to align infrastructure commitments with customer obligations.
The issue is not whether growth should occur. The issue is whether risks and rewards are allocated fairly. In many ways, this may become the defining infrastructure question of the AI era. The debate is no longer simply about how much demand AI creates. It is about who bears the consequences if demand arrives differently than expected.
Test Four: Flexibility
Traditional infrastructure planning often assumes a binary choice: build or do not build. The AI economy increasingly requires a different approach.
Reuters reported in 2026 that utilities, grid operators, and data center developers were exploring flexible operating arrangements that could significantly reduce future infrastructure investment requirements.⁶ Non-firm service, phased energization, interruptible structures, and demand flexibility are increasingly being viewed as tools for managing uncertainty rather than eliminating it.
This represents an important shift in thinking. The most sophisticated organizations are no longer trying to predict the future perfectly. They are designing systems capable of adapting to multiple futures.
Optionality is becoming a strategic asset.
Test Five: Resource Constraints
Electricity is only one constraint. Water availability, cooling requirements, workforce capacity, construction timelines, permitting schedules, and supply-chain limitations increasingly shape infrastructure outcomes. In parts of Arizona and the broader Southwest, water availability has become a central consideration in data center siting decisions.⁷
Many projects that appear viable from an electricity perspective encounter delays elsewhere. The lesson is straightforward. Infrastructure systems rarely fail where attention is focused. They fail where constraints are underestimated.
In some regions, electricity may become the limiting factor. In others, permitting may dominate. In still others, water, workforce availability, or cooling requirements may emerge as the binding constraint. AI infrastructure is a systems challenge, and systems fail at their weakest point.
Test Six: Institutional Durability
Every infrastructure decision eventually encounters institutions. Utilities face regulators. Developers face permitting agencies. Technology companies face communities. Investors face policymakers.
Ireland provides a useful example. The country's efforts to balance data center growth, grid reliability, climate objectives, and economic development have demonstrated that infrastructure decisions eventually become governance decisions.⁸ Technical feasibility alone is rarely sufficient.
Projects must survive regulatory review, legislative scrutiny, community concerns, and changing political priorities. Institutional durability is not a communications issue. It is an infrastructure issue.
Test Seven: Strategic Value
The final test is the most important. Does the opportunity justify the risk? Northern Virginia offers perhaps the clearest example. The region did not become the world's largest concentration of data centers because demand happened to appear there first. Demand followed because fiber infrastructure, transmission access, substations, market structures, workforce capabilities, customer ecosystems, and institutional support evolved together over time.⁹
Its success was not primarily a story about demand. It was a story about alignment. Not every megawatt creates equal value. Not every project strengthens a regional economy. Not every AI investment improves long-term competitiveness.
The strongest decisions emerge when leaders evaluate both sides of the equation: what risks are being accepted and what durable advantages are being created.
The Leadership Challenge
For two years, the AI conversation has focused on models, chips, and capital. The next decade will be shaped by something less glamorous: infrastructure commitments made under uncertainty.
Forecasts matter. The IEA projects that global data center electricity consumption could exceed 945 TWh by 2030.² Yet history suggests that infrastructure transitions are rarely won by the organizations with the most accurate forecasts.
Railroads were not built with perfect demand visibility. Interstate highways were not planned with complete certainty. The modern electric grid itself emerged long before its ultimate uses were understood.
What distinguished successful infrastructure systems was not predictive accuracy. It was the ability to make intelligent commitments under uncertainty. The winners in the AI economy will not necessarily be the organizations with the best forecasts.
They will be the organizations that know which commitments to make before the evidence is complete—and which commitments to avoid. The next phase of the AI economy will be shaped by electricity, interconnection, transmission, cooling, water, land, permitting, capital, regulation, and public trust.
Those constraints are no longer background conditions. They are becoming the terms on which AI infrastructure can scale. That is no longer simply an energy problem. It is becoming the central leadership challenge of the AI economy.
References
- Reuters. “Dutch Power Moratoriums Highlight Challenge Facing Grid Operators.” June 16, 2026.
- International Energy Agency. Energy and AI. Paris: International Energy Agency, 2025.
- Electric Power Research Institute. Powering Intelligence: Updated U.S. Data Center Scenarios. Palo Alto, CA: EPRI, 2026.
- Reuters. “Microsoft Shelves AI Data Center Deals, Sign of Potential Oversupply, Analyst Says.” February 24, 2025.
- Edison Electric Institute. Large Load Projects and Tariffs. Washington, DC: EEI, May/June 2026.
- Reuters. “Stressed U.S. Grid Forcing Data Centers to Get More Flexible.” March 26, 2026.
- Associated Press. “Proposed Data Center Prompts Tucson to Regulate Large Water Users.” August 21, 2025.
- Commission for Regulation of Utilities. Large Energy User Connection Policy Decision Paper. Dublin: CRU, December 12, 2025.
- Virginia Economic Development Partnership. “Data Centers.” Accessed June 18, 2026.
