Executives once rushed to embed artificial intelligence across their operations. Now some call it a monster of their own making. Costs have climbed faster than expected. Budgets strain under the weight of compute expenses that few fully anticipated just 18 months ago.
Amazon, Walmart and Uber rank among the early adopters now dialing back AI deployment at scale, according to a Financial Times report. One unnamed leader told the paper, “We created a monster.” The quote captures a sudden shift. What began as competitive necessity has become a line item demanding constant scrutiny.
But the story runs deeper than corporate belt-tightening. Global data centers already consume around 415 terawatt hours of electricity. That equals roughly 1.5% of worldwide demand. Projections show this figure doubling to 945 TWh by 2030. AI accounts for nearly half the added load.
The International Energy Agency laid out the numbers in detail. Its analysis projects data center electricity use growing 15% annually through the end of the decade. Accelerated servers tied to AI workloads expand at 30% per year. Conventional servers trail at 9%. The gap reveals where the real pressure builds.
The Surge in Demand Meets Grid Realities
Numbers alone don’t convey the strain. A single hyperscale facility can draw 100 megawatts. That’s enough electricity for 100,000 households. Some planned campuses push toward 5 gigawatts. Meta’s Hyperion project in Louisiana will require at least that much once complete. Think of it as powering an entire major city with one installation.
In the United States the picture sharpens. Data centers took 4.4% of national electricity in 2023. Lawrence Berkeley National Laboratory, in a Department of Energy-backed study, sees that share climbing to between 6.7% and 12% by 2028. Absolute consumption could jump from 176 TWh to as much as 580 TWh. Growth like this doesn’t fit neatly into existing transmission systems.
Utilities feel the pinch first. Goldman Sachs analysts warned that data centers will drive 40% of U.S. electricity demand growth through the end of the decade. Prices, they predict, stay elevated. No quick relief for households or manufacturers. But the real bottlenecks appear in permitting, transformers and high-voltage equipment. Lead times stretch years.
Recent developments underscore the tension. On June 20, 2026, the Federal Energy Regulatory Commission directed six regional grid operators to speed connections for large loads including AI data centers. The unanimous vote signals growing official concern. Transmission upgrades can’t wait. Yet building them takes time, capital and political will.
Water adds another constraint. Cooling these facilities consumes vast quantities. UN researchers, as reported by Reuters in early June, forecast both power and water use from data centers doubling by 2030. AI explains 40% of the power side. Local communities in drought-prone areas already push back on new projects.
Companies don’t ignore these signals. Some quietly cancel or delay builds. Others explore onsite generation. Bloom Energy’s 2026 report finds more than one-third of data centers may rely on 100% onsite power by 2030. Gas turbines, fuel cells and even small nuclear reactors enter the conversation. The shift reflects pragmatism. Grid connections remain uncertain. Self-generation offers control. It also raises emissions questions that boards must answer.
Anthropic offers a stark benchmark. The company estimates that training one frontier model by 2027 could demand 5 gigawatts. The entire U.S. AI sector might need 50 GW of fresh capacity by 2028. Former Google CEO Eric Schmidt told Congress that data centers will require 29 GW more by 2027 and another 67 GW by 2030. These aren’t abstract forecasts. They represent real requests landing on utility desks today.
And efficiency gains only complicate the picture. Model optimizations have cut inference costs dramatically at some firms. Alphabet reported a 78% reduction in Gemini serving expenses through improvements. Such progress sounds positive. Yet cheaper queries often trigger more usage. Total energy draw can still rise. Economists call it the Jevons paradox. History shows efficiency rarely reduces absolute consumption when demand is elastic.
Training a large model once burned through millions of kilowatt-hours. Inference now dominates. The IEA calculates that AI servers, mostly running inference, will drive almost half the net increase in data center electricity between 2024 and 2030. A single ChatGPT-style query once took 2.9 watt-hours. Newer measurements put the median closer to 0.3 watt-hours. Multiply by billions of daily interactions and the aggregate still climbs fast.
Tech suppliers feel the ripple effects. Memory makers allocate 70% of certain production to AI data centers in 2026. High-bandwidth memory soaks up capacity. Prices spike. Power electronics face shortages too. The components that manage voltage in servers now prioritize AI customers. Everyone else waits longer or pays more.
Investors watch closely. Hyperscalers including Microsoft, Amazon and Google plan hundreds of billions in capital spending. BNEF tracked data center IT capacity under construction exceeding 23 gigawatts earlier this year. Leasing deals with specialized providers topped $100 billion in value. Five-year commitments give builders confidence. They also expose those builders if utilization falls short.
So far utilization holds. Major model providers reported triple the active users and five times the revenue over the past year. Demand appears real. But corporate users grow more selective. They measure return on each AI feature. Marketing pilots that once launched quickly now face review. Departments get budgets for AI experiments, then hit limits when invoices arrive.
The FT story highlights this reckoning. Early enthusiasm gave way to governance. Chief financial officers now sit in meetings once reserved for technology leaders. They ask simple questions. How much does this feature actually cost per user? What’s the measured productivity lift? Answers don’t always satisfy.
Yet abandonment isn’t the theme. Refinement is. Companies trim low-value applications. They optimize models. They move workloads to times or locations with cheaper power. Some explore smaller, specialized models that require less compute. The monster gets leashed, not slain.
Longer term the industry bets on breakthroughs. More efficient chips. Liquid cooling at scale. Even nuclear restarts or new small modular reactors tailored for data center campuses. Virginia and other states court these projects with tax incentives and streamlined approvals. The economic upside justifies the effort. Jobs, tax revenue and technological edge hang in the balance.
Still, near-term limits persist. Grid operators in Texas project peak demand hitting 145 GW by 2031, up sharply from recent levels. Data centers contribute heavily to that curve because they run 24 hours a day. Their load profile is steady, predictable and enormous. Traditional generation and transmission weren’t built for this shape of growth.
Policy makers respond unevenly. The FERC action on June 20 aims to prevent bottlenecks from slowing critical infrastructure. It won’t add megawatts overnight. Transmission projects face local opposition, environmental reviews and supply chain delays for transformers that can take two years to manufacture.
Private capital steps in where public systems lag. Hyperscalers sign direct power purchase agreements with renewable developers and gas plant owners. Some explore co-location with nuclear facilities. The arrangements bypass congested utility queues but raise questions about who pays for grid upgrades that benefit everyone.
One fact stands clear. The AI buildout has moved beyond hype. It now collides with physical limits on energy, water and equipment. Companies that master this intersection, through efficiency, smart siting or creative financing, gain lasting advantage. Those who treat compute as free will face painful corrections.
The monster, as one executive put it, has been created. Taming it requires more than better algorithms. It demands fresh thinking about power markets, infrastructure planning and the true economics of intelligence at scale.
WebProNews is an iEntry Publication