Back
Artificial intelligence’s bubblitude fizzes with circular transactions, risk concealment, and exotic real-estate debt finance. In a frenzy to build AI data centers, Big Tech recently borrowed and bonded more money in 11 weeks than in the previous three years combined. More than a thousand new data centers are under construction or planned nationwide. Though they don’t yet know how many of those facilities will eventually materialize, energy suppliers are using AI data centers’ ravenous appetite for electrons to justify vast new investments in gas and nuclear power plants and the revival of uneconomic coal plants, claiming that all are needed to win the AI arms race and keep the lights on.
This trillion-dollar surge is transforming not only equity and capital markets but also the future U.S. power mix, locking in decisions that will shape energy affordability for decades. Smarter, cheaper, cleaner, less-risky options for powering data centers exist — if decision-makers choose them.
To meet all the expected new electricity demand, the U.S. has rapidly proliferated its gas-fired capacity under development in 2025. For context, at the start of 2024, only 4 gigawatts of gas-fired power in the U.S. development pipeline were explicitly earmarked for powering data centers. Today, over 100 gigawatts are.
And developers are proposing to invest over $400 billion to build more than 250 gigawatts of new U.S. gas-fired power plants — nearly tripling the gas power pipeline in a year, mostly driven by speculative AI projects subsidized by 37 heavily lobbied state governments.
Some data centers are even being mandated as “critical defense facilities” to be built on federal land, alongside otherwise uneconomical nuclear plants exempted from strict Nuclear Regulatory Commission scrutiny, all at taxpayer expense. This is happening, ironically, in Texas — the nation’s free-enterprise leader in solar, wind, and batteries. These renewable resources totaled 97% of its 2025 capacity additions, while fossil fuels amounted to 3%, and nuclear 0%. But in the past two years, planned gas plants in Texas nearly quadrupled, to 80 gigawatts. Only China has more gas plants under development than Texas, and nearly half the Texas plants are meant to power data centers directly.
We’ve seen this movie before. A quarter century ago, the coal industry warned that the Internet would overwhelm the grid without massive new coal capacity. Demand proved to be over tenfold lower. The dot-com bubble burst in 2000, permanently vaporizing $120 billion of electricity investments and embalming another $80 billion in infrastructure built long before it was needed. Today’s AI mania rhymes: Gas and nuclear vendors that can’t beat energy efficiency and renewables in competitive markets are leveraging hype into mandates and subsidies to rescue their losers.
Yet capital markets increasingly fear that AI looks like a bubble set to pop. That’s because each new data center effectively bets against at least 10 plausible outcomes that make the investment unwise: Scaling large language models could fail to achieve superintelligence; customer revenue could disappoint; inaccuracy may persist; smaller and leaner models might keep outperforming giants; copyright infringements may have to be paid for; data centers may go on quadrupling their energy efficiency every year; and flexible interconnection might stretch existing grid assets to serve all new demand.
Each new power plant also bets against the ways that data centers may access cheaper electricity, such as adding pop-up microgrids, colocating renewables and storage at idle gas plants, and buying efficiency, flexible load, storage, and clean supply from other customers. Betting against any one of these realities is risky. Betting against all of them strains credulity.
Many utilities are already trimming projections toward reality. Regulators in data-center hot spots are scrambling to shield customers from accelerating and politically sensitive rate hikes — already up 16% in Illinois, 13% in Virginia, 12% in Ohio, and 6% nationwide. Meanwhile, actual data-center demand still barely shows up in national totals. U.S. weather-adjusted electricity use fell in 2023, then rose by 2% in 2024, about one-twentieth due to new data centers. Nearly all the growth comes instead from air conditioning, electrifying buildings and vehicles, and reshoring industry. These needs can all be more cheaply met by better efficiency, and by another vast and potent competitor to fossil fuels: renewables.
Globally, data centers — roughly one-ninth of which are devoted to AI — use about 1.5% of today’s electricity. The International Energy Agency forecasts they’ll grow in this decade while renewable supplies grow 11 times more. Thus, solar and wind power, now swiftly displacing costlier fossil-fueled and nuclear power, dwarf the AI boom. Speed to market is paramount for AI developers, so many smart tech companies choose renewables to get their data centers built and running quickly and cheaply.
However, other AI firms have rushed for gas power, and that stampede has doubled gas-plant costs and backlogged gas turbine deliveries to past 2030, to the point that two-thirds of gas-plant project proposals have no named turbine manufacturer. This jam has pushed about a fifth of projects to substitute off-grid gas power, often using adapted aircraft jet engines. These turbine generators are easily available but engineered to meet peak demand, so they’re inefficient, noisy, and dirty. Running them constantly to power data centers would quickly inflate electricity costs and magnify public health damages. U.S. data centers were already projected to cause more than $20 billion per year in asthma and cardiopulmonary disease costs by 2030. Communities will not welcome additional pollution, water stress, noise, and rate hikes.
Gas markets magnify the financial risks of turning to gas to power data centers. New gas wells decline faster than old ones, while falling oil prices can make new drilling and refracking unattractive. At the same time, exuberant exports of liquefied American gas (and gas pipelined to Mexico) are pushing gas toward both global glut and domestic scarcity. The analysts at BloombergNEF predict that new gas-fired AI power could tip the 2025–30 U.S. gas surplus into a deficit, making volatile gas prices for heating, industry, and utilities spike. Indeed, BloombergNEF says wholesale gas futures for 2028–30 are unsustainably priced below production cost. And whatever the gas price, new gas-fired power plants are likely to become underutilized, subsidized assets that burden electricity customers long after today’s AI ebullience fades. While many data centers will be built, many won’t, and many won’t actually run at full tilt for decades to come — stranding gas plants and pipelines built to power them.
Even as national policy reinforces a gas lock-in, power choices that can scale at AI speed already dominate actual markets. Renewables captured over 92% of the world’s new generating capacity in 2024 and (including storage) about 90% of U.S. additions in 2025, with 93% expected in 2026. They are far cheaper than gas power, keep getting cheaper, sell on constant-price contracts for decades, and finance like low-risk annuities. They’re virtually unlimited and deploy at industrial speed.
Last May, China added 1 gigawatt of solar and wind power roughly every six hours around the clock. Pakistan displaced 30% of its utility power with solar in four years. Vietnam added solar equivalent to half of its coal generation in two years. South Australia generates 75% of its annual electricity from renewables and will reach 100% by 2027, driving 37 firms to propose relocating there to secure stable, low-cost power. Global metals giants Rio Tinto and BHP are relying on “renewable baseload” power to smelt aluminum and mine copper. Apple’s data centers have run on fully renewable energy for more than a decade. Google just announced that on-site solar, wind, and battery power will get its new 850-megawatt Texas data center online in 18 months, not five-plus years.
Critics have long claimed that variable renewables are too unreliable: The wind doesn’t always blow, and the sun doesn’t always shine. But evidence shows that intermittency concerns are now generally unfounded. Ten proven carbon-free balancing methods already make high-renewable grids reliable and economic in many countries. One of those methods, batteries, costs 96% less today than it did in 2010. BloombergNEF finds that battery-firmed solar and wind deliver steady power more cheaply than any new fossil or nuclear plants, and many operating ones. That’s why three-fourths of India’s new firm capacity today is solar-plus-storage.
Renewables also offer essential speed. In Sparks, Nevada, the world’s largest solar-powered microgrid continuously powers modular data centers. Solar panels laid on desert ground feed hundreds of second-life electric-vehicle batteries joined to form a superbattery. It was all built in four months and delivers electricity that’s cheaper, quieter, and more reliable than grid power; uses virtually no water; emits nothing; and is even portable. This is what clean, scalable, market-speed power looks like. Gas isn’t it.
AI does have some valuable applications. No one yet knows, though, if its revenues can repay the immense and swiftly depreciating investments required. But while markets are answering that trillion-dollar question, the AI boom must not be allowed to undermine American energy affordability and security.
Utilities and regulators can protect existing customers with a simple safeguard, giving teeth to vague qualitative pledges: Sell power to new data centers only under “take or pay” contracts that repay the entire electricity investment regardless. Those agreements should be backed by robust bonds or insurance, priced by capital-market risk experts (not by developers), to ensure that if an AI venture collapses, losses fall on the developer, not on households and small businesses.
If markets, and not mandates, determine the outcome, the conclusion is already clear. Gas, coal, and nuclear are too slow, too costly, and too risky to anchor the next wave of U.S. power demand. The only technologies that scale quickly enough, cheaply enough, and reliably enough for AI already dominate global additions. Policy will now decide whether Americans will enable the new energy system or protect the old — and whether they’ll pay for stranded gas plants or profit from the cheapest and most secure electricity in history.
.svg)
