The AI Data Center Energy Crisis: Can We Power the Future?

The Scale of the Problem

Something extraordinary is happening to the global power grid, and most people have no idea. While consumers debate which AI chatbot writes the best emails, the infrastructure behind those chatbots is triggering the largest energy buildout in modern history. We are not talking about a modest uptick in electricity usage. We are talking about a fundamental reshaping of how the world generates and consumes power.

U.S. data center energy demand is projected to nearly double between 2025 and 2028, jumping from 80 to 150 gigawatts. To put that in perspective, that is equivalent to adding a country with the energy needs of Spain to the American grid in just three years. Globally, AI-driven data centers are contributing nearly one-fifth of all power growth, consuming roughly 126 GW annually through 2028, which is almost as large as Canada's total annual power demand.

The investment numbers are staggering. Large technology companies committed more than $1 trillion in spending during the 2025-2026 period alone, and Goldman Sachs Research estimates that approximately $720 billion in grid spending through 2030 will be needed just to keep pace. When we first started tracking these numbers in early 2025, the projections seemed aggressive. They have since been revised upward multiple times.

In our analysis, three forces are converging simultaneously: the exponential growth of AI compute, the aging state of existing power infrastructure, and the sheer physical limits of how quickly new generation capacity can come online. The result is an energy crisis that is already here, even if most consumers have not felt the full impact yet.

How Much Power AI Actually Needs

Not all computing is created equal. A traditional Google search uses roughly 0.3 watt-hours of electricity. A ChatGPT query consumes approximately 10 times that amount. Training a single large language model can consume as much electricity as 100 American homes use in an entire year. And that was the baseline in 2024. The models being trained in 2026 are orders of magnitude larger.

The real power drain comes from inference, not training. Every time someone asks an AI to generate an image, write code, summarize a document, or power an autonomous agent, that inference request hits a GPU cluster in a data center somewhere. Multiply that by hundreds of millions of daily users across dozens of AI products, and the aggregate demand becomes enormous.

By 2028, data centers could consume 12 percent of all electricity generated in the United States. That is a remarkable figure for an industry that barely registered on the national energy radar a decade ago. PJM, the largest grid operator in the U.S., projects a 6 gigawatt shortfall by 2027, equivalent to the output of six large nuclear power plants simply not existing when they are needed.

We have spoken with data center operators who describe the current situation as a land grab. The companies that secure power purchase agreements now will have the capacity to train next-generation models. The ones that do not will fall behind. Energy access is becoming as important as chip access in the AI race.

The Nuclear Bet

When three of the richest companies in history simultaneously pivot toward the same energy source, it is worth paying attention. Microsoft, Google, and Amazon have all made massive bets on nuclear power, and the logic is straightforward: nuclear provides 24/7 carbon-free baseload power that renewables simply cannot match.

Microsoft and Three Mile Island

Microsoft committed to a 20-year, 835-megawatt power purchase agreement worth $16 billion to restart the Three Mile Island nuclear facility in Pennsylvania. Yes, that Three Mile Island. The site of America's most infamous nuclear accident in 1979 is being resurrected to power AI. The restart targets 2028, and if successful, it would provide dedicated, carbon-free electricity to Microsoft's data center operations in the region.

We find the symbolism remarkable. The facility that nearly killed the American nuclear industry is now being positioned as its savior. Microsoft's willingness to attach its brand to Three Mile Island signals just how desperate the power situation has become.

Google and Small Modular Reactors

Google took a different approach by signing the first U.S. corporate deal for a fleet of small modular reactors (SMRs) with Kairos Power. The agreement covers 500 MW of capacity, with operations expected by the early 2030s. SMRs are a newer technology that promises faster deployment, lower upfront costs, and enhanced safety compared to traditional nuclear plants.

In our assessment, Google's bet is the riskiest of the three. SMRs remain largely unproven at commercial scale, and the timeline to operational capacity stretches several years into the future. But if the technology delivers, it could become the template for how tech companies power their infrastructure going forward.

Amazon and Susquehanna

Amazon has arguably made the boldest move, investing over $20 billion to convert the area around the Susquehanna nuclear facility into an AI-ready data center campus powered entirely by carbon-free nuclear energy. The scale of this investment dwarfs most individual energy projects and signals Amazon's belief that nuclear adjacency will be a defining competitive advantage in the AI era.

Meta's Nuclear Ambitions

Not to be outdone, Meta issued a request for proposals seeking 1 to 4 GW of new nuclear capacity. That is an enormous amount of power for a single company, enough to supply a mid-sized city. Meta has been quieter about its AI ambitions than its competitors, but this RFP suggests its infrastructure plans are just as aggressive.

Renewable Energy Alternatives

Nuclear is not the only path forward. Solar and wind capacity continue to grow rapidly, and in many regions they offer the cheapest electricity available. But renewables face a fundamental challenge for AI workloads: intermittency.

AI data centers run 24 hours a day, 365 days a year. They cannot throttle down when clouds cover the solar panels or wind speeds drop. Battery storage technology has improved dramatically, but storing enough energy to bridge multi-hour or multi-day gaps in renewable generation remains prohibitively expensive at the scale data centers require.

That said, several promising developments deserve attention. Large-scale battery installations in Texas and California are beginning to demonstrate that four to six hours of grid-scale storage is commercially viable. Enhanced geothermal systems, which tap heat deep underground regardless of weather conditions, are attracting significant investment from companies like Fervo Energy. And next-generation solar panels with efficiency ratings above 30 percent are entering commercial production.

In our view, the most realistic path forward is a hybrid approach: renewables for daytime and high-wind periods, nuclear for baseload, and battery storage to smooth the transitions. No single energy source can solve this problem alone.

The Water Problem

Energy consumption gets the headlines, but data centers have another resource problem that receives far less attention: water. Cooling these massive facilities requires enormous quantities of water, and the demand is growing at an alarming rate.

Without new efficiencies, data centers across America may require 697 million to 1.45 billion gallons of extra peak water capacity per day by 2030. That upper figure matches New York City's entire daily water supply. Expanding public water systems to meet this demand is projected to cost between $10 billion and $58 billion by 2030.

The geographic problem is acute. More than 160 new AI data centers have been built across the U.S. in the past three years, and many of them sit in regions with already-scarce water resources. Data centers in Texas alone will consume an estimated 49 billion gallons of water in 2025, potentially rising to 399 billion gallons by 2030.

Traditional evaporative cooling, which is the most energy-efficient cooling method, loses approximately 80 percent of the water it uses to evaporation. The remaining 20 percent is sent to wastewater facilities. Closed-loop cooling systems can reduce freshwater use by up to 70 percent, but they cost more to build and operate.

We have visited data center sites where the tension between the facility and the local community is palpable. Residents see their water bills rising while a massive tech campus next door consumes water at industrial scale. This is a political problem as much as an engineering one.

Next-generation cooling technologies, including cold plates, single-phase immersion, and two-phase immersion cooling, are being deployed at increasing scale. These liquid cooling methods can dramatically reduce water consumption while also improving energy efficiency, but retrofitting existing facilities is expensive and time-consuming.

Impact on Consumer Electricity

Here is where the data center energy crisis becomes personal. Residential electricity prices in the U.S. have risen by more than 36 percent since 2020, from 12.76 cents per kilowatt-hour to 17.44 cents per kilowatt-hour in early 2026. In many cities, the average price has climbed to roughly 19 cents per kilowatt-hour, an increase of nearly 50 percent in just a few years.

Data centers are not the only cause of rising electricity prices. Inflation, aging infrastructure, extreme weather events, and the cost of the clean energy transition all play roles. But data center demand is becoming a significant contributing factor, expected to account for 40 percent of electricity demand growth through the end of the decade.

Goldman Sachs projects that data center growth will boost core inflation by 0.1 percent in both 2026 and 2027, and household electricity prices are expected to rise an additional 6 percent through 2027. The effects are already visible in specific regions. Baltimore residents saw their average bill jump by more than $17 per month after a record-setting power auction by PJM, with another increase of up to $4 coming in mid-2026.

The distributional impact is concerning. Lower-income households spend a larger share of their income on electricity, so these price increases hit hardest where they can be least afforded. Goldman Sachs analysts have noted that the income and spending drags will be disproportionately larger for these households.

In our assessment, there is a growing disconnect between the companies profiting from AI and the communities absorbing the infrastructure costs. Big Tech negotiates favorable long-term power contracts, while residential ratepayers face the grid upgrade costs passed through by utilities. This dynamic is generating political backlash that will likely intensify over the coming years.

What Happens Next

The next three to five years will be critical. Several scenarios are plausible, and the outcome depends on decisions being made right now.

In the optimistic scenario, nuclear restarts and new SMR deployments begin delivering power by 2028-2030. Renewable capacity continues to expand rapidly. Battery storage costs fall enough to make multi-hour storage economically viable. Liquid cooling technology reduces water consumption significantly. Grid upgrades proceed on schedule, and the power crunch eases by the end of the decade.

In the pessimistic scenario, nuclear projects face regulatory delays and cost overruns, as they historically have. Renewable growth hits supply chain constraints. NIMBYism slows new transmission line construction. The gap between data center demand and available supply widens, leading to more aggressive competition for power, higher prices, and potential brownouts in vulnerable regions.

The most likely outcome, in our view, sits somewhere in between. Some nuclear projects will deliver on time. Others will not. Renewables will continue growing but will not fully close the gap. Electricity prices will continue to rise, but not catastrophically. The AI industry will face periodic capacity constraints that slow the pace of scaling but do not halt it.

What we are watching most closely is the political response. Several states are already considering legislation to require data center operators to pay a larger share of grid upgrade costs. If this movement gains momentum, it could reshape the economics of data center siting and potentially slow the U.S. buildout in favor of international locations with cheaper, more abundant power.

Can Innovation Solve This

The technology industry's track record of solving resource problems through innovation is genuinely impressive. But energy is different from software. You cannot scale a power plant the way you scale a cloud service. Physical infrastructure takes years to build, requires permits and community buy-in, and is subject to the laws of thermodynamics.

That said, several innovations could meaningfully change the trajectory.

More efficient AI hardware. Every generation of AI chips delivers more compute per watt. NVIDIA's Blackwell architecture is significantly more energy-efficient than its Hopper predecessor, and custom chips from Google (TPU), Amazon (Trainium), and Microsoft (Maia) are all optimized for power efficiency. If hardware efficiency gains outpace the growth in model size, total energy demand could moderate.

Model efficiency improvements. Techniques like quantization, distillation, and mixture-of-experts architectures allow AI models to deliver comparable performance with significantly less compute. The shift from dense to sparse models is particularly promising, as it means not every parameter needs to be activated for every query.

Edge computing. Moving more AI inference to user devices reduces the load on centralized data centers. Apple, Google, and Qualcomm are all investing heavily in on-device AI capabilities, and the latest generation of smartphones and laptops can handle many AI tasks locally.

Fusion energy. This is the wildcard. Companies like Commonwealth Fusion Systems and Helion Energy claim they are within years of demonstrating net energy gain from fusion reactors. If fusion delivers on its promise, it would provide virtually unlimited, clean, baseload power. But fusion has been "ten years away" for decades, so skepticism is warranted.

In our analysis, hardware and model efficiency improvements are the most likely near-term contributors. They will not eliminate the energy problem, but they could slow the rate of demand growth enough to give infrastructure buildout time to catch up.

Conclusion

The AI data center energy crisis is not a future problem. It is a present one. The decisions being made today by tech companies, utilities, regulators, and policymakers will determine whether the AI revolution accelerates or stalls.

The numbers are sobering: demand doubling in three years, trillion-dollar investments required, water consumption rivaling major cities, and electricity prices rising for hundreds of millions of consumers. But the response has been equally dramatic. The nuclear renaissance, the acceleration of renewable deployment, and the push for more efficient AI hardware all represent genuine progress.

What concerns us most is the equity dimension. The benefits of AI are broadly distributed, but the costs of powering it are concentrated in specific communities, often those least equipped to absorb them. Addressing this imbalance should be a priority for every stakeholder in the AI ecosystem.

The energy crisis will not stop AI. The economic incentives are too powerful, and the technology is too useful. But it will shape AI's trajectory in ways that pure technologists often underestimate. Energy access is becoming the defining constraint of the AI era, and the companies and countries that solve it first will lead what comes next.