This essay is written for security leaders, infrastructure owners, and executives who seek a grounded understanding of how power supply will influence the growth of AI. I’ll analyze the assumptions, test the logic, and provide a hardening blueprint that you can act on now.

Global electricity use by data centers is projected to more than double by 2030, reaching approximately 945 TWh. The International Energy Agency attributes most of that surge to AI, with AI-optimized centers projected to more than quadruple their demand by the end of the decade.

In North America, the reliability community is already signaling stress. NERC’s 2025 summer outlook flagged record load growth driven by new data centers, electrification, and industrial activity. Its 2024 long-term assessment highlights the planning and operational challenges associated with large “single parcel” loads, such as AI campuses.

Assumption check. “Efficiency gains will offset most of this growth.” That’s unlikely. Model sizes, context windows, and inference volumes are increasing faster than efficiency gains. The IEA’s base case still projects data center electricity growth at about four times the rate of total grid demand.

Test the logic. If your 2026–2029 AI roadmap assumes “power will be there,” you’ve embedded unpriced risk. Treat power as a first-class dependency, the same way you treat capital and talent.

The U.S. bulk power system is aging. Large power transformers typically have an average service life of 38–40 years, which is at or beyond their typical design life, and distribution fleets exhibit similar age profiles in many territories. Replacement lead times and supply-chain constraints remain elevated.

The American Society of Civil Engineers’ 2025 report card lowered the energy grade to D+, citing shortfalls in condition and capacity. DOE is now using a uniform methodology to identify at-risk regions for reliability interventions.

Even where generation exists, getting steel in the ground and connected remains a slow process. The U.S. interconnection queues contain capacity that rivals or exceeds the entire installed fleet, which underscores how connection, not just construction, bottlenecks the transition.

Assumption check. “We’ll just buy more renewables.” Many projects sit in queues for years. Without transmission upgrades and firming, corporate PPAs won’t translate into a 24/7 supply that can feed GPUs at scale.

Counterpoint. Innovation can squeeze more out of existing wires. Dynamic line ratings, sensors, better forecasting, and storage can raise adequate capacity. You still need time, permits, and capital.

China treats electricity like a solved problem. Decades of overbuilding have left the country with massive reserves. New data centers are welcomed because they help consume that excess. China also built out ultra-high-voltage corridors to move bulk power and can restart idle coal plants or rapidly connect new solar installations. This ability to adjust supply quickly compresses the time between AI demand signals and actual power delivery.

The EU now requires large data centers to report energy performance and water metrics to a central database, with the option to tighten standards. Nordic projects pair abundant carbon-free power with waste-heat recovery into district heating, turning a liability into a civic asset.

After a pause, Singapore reopened data center approvals through a tightly managed program that caps capacity and rewards efficiency and the use of green power. It remains a hub, but on stricter terms.

Ireland is a cautionary case of concentrated demand outpacing grid headroom. Data centers consumed 22 percent of metered electricity in 2024, with new projects facing constraints into the late 2020s.

Alternative framing. Regions with clean, firm power, cooling advantages, and community integration models will become AI “compute provinces.” Others will face moratoria, curtailments, or costly make-ready work.

Leaders are no longer waiting for the grid to catch up.

In the U.S., many tech firms are even building their own private power plants to insulate operations from grid strain, while households are already experiencing higher bills as local grids stretch to cover new demand.

Assumption check. “Backup diesel covers our risk.” Diesel remains common and will persist, yet fuel logistics, permitting, localized emissions, and runtime limits make it a weak primary strategy for AI campuses. Expect regulators and communities to keep tightening around it.

Regulatory baseline. NERC CIP standards establish mandatory cybersecurity controls for the bulk power system, while CIP-014 focuses on physical security at critical transmission stations and substations. CISA has released cross-sector cybersecurity performance goals to raise the floor across essential services and has guidance on UAS threats to infrastructure.

Practical security blueprint for owners, operators, and large customers:

Counterpoints. A strict “self-power” posture can fragment the grid and raise social costs. The better path pairs private resilience with utility-scale upgrades that benefit the region.

Meanwhile, Goldman Sachs reports that technology demand is already growing faster than the expansion of the grid. McKinsey projects $6.7 billion in new data center investments by 2030, underscoring how money is flowing even as physical infrastructure lags.

Acceleration levers. Expand transmission, prioritize transformer procurement, align utility resource planning with forecasted AI loads, and incentivize firm, clean supply.

What leading firms are doing now is securing firm supply through nuclear and long-dated contracts, siting facilities in power-rich regions, and establishing on-site generation to de-risk interconnection timelines.

Your 90-day plan. Build a megawatt-by-megawatt map of your AI roadmap. Pressure-test interconnection assumptions. Open talks with utilities about firm service. Pre-negotiate load-shedding playbooks.

Test the logic. If model delivery dates don’t move when power delivery dates move, you’re not being honest with yourself.

Near term, 2025–2027. Bigger single-site loads. More gas peakers. More corporate nuclear deals. Early microgrids and fuel cells at AI campuses. Continued local resistance to new transmission.

Mid term, 2027–2030. Data center electricity use accounts for nearly 3 percent of global demand—first wave of advanced nuclear units. EU tightens standards. U.S. grid upgrades are maturing, but supply chain bottlenecks persist.

Strategic risk. China’s ability to scale power decades ahead, versus America’s reactive approach, creates a long-term competitive imbalance unless U.S. planning shifts from crisis response to proactive buildout; leadership in AI could erode.

Energy is now the fulcrum of AI power. China’s long planning horizon has created surpluses. The U.S. is still fighting permitting battles and household bill spikes. Without a structural change, America risks not just blackouts but losing the AI race itself.