AI is creating the largest surge in electricity demand since the electrification of American industry. It’s also becoming the most powerful tool for solving the problem it created. For the first time since the industrial era, electricity itself is becoming a strategic computing input. The limiting factor in AI is no longer computation alone. It is the physical infrastructure required to sustain computation at planetary scale.
David Goodnight | Founder, Comnet International | Author, Financing the World We Trade In
• AI data centers consumed energy equivalent to all of Pakistan.
• Nuclear, geothermal, fuel cells, rare earth magnets, solar, and grid optimization are all being reshaped by AI tools.
• AI found more geothermal sites in three years than the industry found in thirty.
• Data center waste heat is raising neighborhood temperatures by up to 4 degrees Fahrenheit.
• A startup just raised $140 million to build floating AI data centers powered entirely by ocean waves, operational by August 2026.
There is a central irony at the heart of the current AI-energy buildout: the same technology driving unprecedented electricity demand is now being deployed across laboratories, drilling rigs, and transmission control rooms to help solve the supply problem it created.
The scale of AI’s energy consumption is worth stating plainly. US data centers drew 183 terawatt-hours of electricity in 2024, more than 4 percent of total national consumption. The International Energy Agency projects that figure reaches 426 terawatt-hours by 2030, a 133 percent increase in six years, equivalent to adding the entire current electricity consumption of Germany to America’s grid in less than a decade.
America’s investor-owned utilities have responded by committing 1.4 trillion dollars in capital spending through 2030, the largest coordinated utility investment in American history. Goldman Sachs projects a 160 percent increase in data center power demand by decade’s end. Increasingly, the grid is being planned around a single emerging customer class: AI infrastructure.
Solar, wind, nuclear fission, and natural gas are all being deployed at maximum pace. They are necessary yet not sufficient. One of the deeper problems is timing. Building new transmission lines takes five to ten years. Permitting a gas plant requires three to five. Grid interconnection queues in key US markets now stretch seven to ten years. AI data center demand is growing on a trajectory that none of these timelines can match. OpenAI can train a frontier model in months. Building the transmission infrastructure to power the next generation of models can take a decade. No policy intervention has yet closed that gap.
What published research now confirms, across six distinct domains, is that artificial intelligence is simultaneously attacking the supply side of this equation. Not in theory. In peer-reviewed journals, in funded laboratory programs, in commercial deployments, and in signed deals that have cleared institutional capital committees.
1. Nuclear Fusion: Compressing Decades into Years
Researchers at Ames National Laboratory have built an AI tool called DuctGPT, published in Acta Materialia, that compresses the search for fusion reactor materials from months to hours. The core problem is brutal: reactor walls must withstand plasma temperatures in the hundreds of millions of degrees (hotter than the sun’s core) without degrading. Finding metal alloys that survive those conditions once required years of trial-and-error laboratory work. DuctGPT turns it into a search query: specify the required properties, and the system returns ranked candidates. What took a research team years now takes an afternoon.
AI is reshaping the physics of plasma control as well. One of fusion’s persistent engineering problems is plasma instability: sudden disruptions that can damage reactor walls and halt operation. AI systems are now being trained to detect the early signatures of these events and intervene before they occur. Google DeepMind is applying reinforcement learning to plasma stabilization in collaboration with Commonwealth Fusion Systems. NVIDIA and General Atomics are building a digital twin of the DIII-D National Fusion Facility in San Diego, allowing engineers to stress-test reactor conditions virtually before any physical exposure. The UK government committed 45 million pounds to build a dedicated AI supercomputer at the Culham fusion campus, beginning operations in June 2026.
Private capital is not waiting for government timelines. Global private investment in fusion now totals more than 9.7 billion dollars across approximately 50 projects. Helion Energy’s Polaris prototype achieved 150 million degrees Celsius in February 2026, ten times the sun’s core temperature, operating with deuterium-tritium fuel for the first time in a privately developed machine. Helion holds a power purchase agreement with Microsoft targeting grid delivery by 2028.
2. Geothermal: More Discoveries in Three Years Than in Thirty
Geothermal energy is firm baseload power. It operates 24 hours a day regardless of weather and exists nearly everywhere beneath the surface. Its problem has always been finding viable sites efficiently. Historically that required expensive brute-force drilling, most of which came up dry.
Zanskar, a geothermal exploration company, says its AI models have accelerated geothermal discovery rates dramatically, identifying more conventional geothermal sites in three years than the industry had found in the prior three decades. The approach has been validated with successful drilling at Pumpernickel and Big Blind in Nevada, both sites previously believed to be barren. A systematic review of 183 peer-reviewed papers published in ScienceDirect in November 2025 confirms AI is now being applied across the entire geothermal lifecycle, from exploration to drilling optimization to plant operations.
Fervo Energy, backed by Google and Breakthrough Energy Ventures, uses fiber optic sensing and AI-enhanced monitoring to improve drilling precision. The company raised 462 million dollars in December 2025 and is preparing a 1.33 billion dollar IPO at a valuation of up to 6.5 billion dollars in 2026. Geothermal is the baseload answer that solar and wind cannot provide, and AI found it hiding in ground the industry had already written off.
3. Solid Oxide Fuel Cells: AI Perfecting a Technology Nobody Noticed
Solid oxide fuel cells have been in commercial use since the early 2000s. They generate electricity through a chemical reaction rather than burning fuel, which makes them significantly more efficient than gas turbines or diesel generators, converting 50 to 60 percent of fuel to usable power, compared with 30 to 40 percent for conventional alternatives. They use no water in normal operation, can be installed in 90 days without waiting years for grid interconnection approval, and can run on hydrogen or biogas using the same hardware.
AI did not invent this technology. It made it precise for a new purpose. Bloom Energy’s sixth-generation platform was redesigned specifically for AI data center workloads using computational modeling against operating data from 1.5 gigawatts of deployed systems worldwide. The redesign produced cleaner power output, faster response to shifting load demands, and integrated cooling, three problems that conventional backup generators cannot solve simultaneously. The results are confirmed in Bloom’s 2025 SEC filings.
The capital markets noticed. Between October 2025 and January 2026, fuel cell companies closed 7.65 billion dollars in binding data center power agreements in 90 days. Brookfield Asset Management committed up to a 5-billion-dollar deployment of Bloom fuel cells at AI factories globally. American Electric Power followed with a 2.65 billion dollar, 20-year agreement. Goldman Sachs, Morgan Stanley, and Evercore ISI have all published research validating fuel cells as a bankable primary power source. Oracle’s Project Jupiter in New Mexico, where 2.45 gigawatts of Bloom fuel cells replace planned gas turbines and diesel generators, is the largest announced fuel cell deployment for an AI data center anywhere in the world.
Low-water generation technologies are also becoming strategically important as hyperscale campuses increasingly face cooling-water constraints and local permitting resistance.
4. Rare Earth Magnets: AI Reading 67,000 Compounds at Once
Rare earth magnets are inside every EV motor, wind turbine generator, defense platform actuator, and data center cooling system. China processes approximately 90 percent of global rare earth elements. In April 2025, Beijing imposed export licensing requirements on seven heavy rare earth elements, causing weeks-long delays for drone manufacturers and demonstrating the supply chain leverage in stark terms.
Researchers at the University of New Hampshire trained machine learning models to autonomously extract data from thousands of scientific papers and built the Northeast Materials Database, a catalogue of 67,573 magnetic compounds. The AI identified 25 previously unrecognized materials that retain magnetism at high temperatures. Seven of the screened candidates were subsequently confirmed in existing scientific literature, validating the model’s accuracy. The study was published in Nature Communications in October 2025, funded by the US Department of Energy’s Office of Basic Energy Sciences. A parallel 2026 study in the Journal of Magnetism and Magnetic Materials used AI screening to identify three additional rare-earth-free candidates.
5. Perovskite Solar and Green Hydrogen: Closing the Efficiency Gap
Researchers at the Karlsruhe Institute of Technology used AI to work backward from the solar cell performance they wanted, identifying new materials that could deliver it, a reversal of the traditional laboratory process. Published in Science in late 2024, the AI-discovered materials pushed one cell’s efficiency to 26.2 percent, roughly two points above the baseline. In 2025, next-generation perovskite-silicon tandem cells reached 34.5 percent efficiency in laboratory conditions, a threshold that conventional silicon alone cannot cross. The first commercial-scale perovskite production facilities were announced in 2025.
On the hydrogen side, Japan’s National Institute for Materials Science used AI to discover high-performance materials for splitting water into hydrogen without the rare platinum-family metals that have historically made the process expensive. A March 2026 paper in Angewandte Chemie describes an AI system that predicts catalyst performance before any physical experiment is run, then refines its own hypotheses based on lab results. This is not AI assisting science. It is AI conducting it.
6. Grid Optimization and the Transmission Bottleneck
All of the generation technologies above require time to scale. In the gap, AI is doing something less visible but immediately valuable: reducing the friction inside the grid that already exists. But before optimization comes a more fundamental problem that energy professionals know, and the broader conversation has largely missed: the physical infrastructure to move power from source to load is itself the binding constraint.
Power transformer lead times in the United States reached 128 weeks in 2025, up from roughly 24 weeks in 2021, according to Wood Mackenzie. Generator step-up transformers average 144 weeks, with some orders extending to four years. Transformer prices have risen 77 percent since 2019. Sightline Climate tracked 12 gigawatts of 2026 US data center capacity announced across 140 projects; only 5 gigawatts was actually under construction. High-voltage substations, switchgear, and HVDC equipment face lead times of three to five years. Electrical equipment represents less than 10 percent of total data center cost and 100 percent of the bottleneck.
AI is being applied directly to this constraint. PJM Interconnection approved an 11.8 billion dollar transmission expansion plan in February 2026 and deployed AI-enabled tools to compress interconnection review cycles that currently stretch seven to ten years. High-voltage long-distance transmission, the infrastructure that moves power from where it is generated to where AI needs it, is a market projected to grow from 15.6 billion dollars in 2025 to over 22 billion dollars by 2030, with AI now being applied to accelerate both project design and real-time grid control. AI-driven planning tools are compressing timelines on transmission infrastructure that was previously constrained by the pace of human engineering teams. The DOE committed 1.9 billion dollars to transmission modernization in 2025. The constraint is structural, documented, and not yet solved.
A peer-reviewed scenario analysis published in Nature Communications Sustainability in May 2026 projects that AI-enabled grid optimization could generate net electricity savings of up to 130 percent of AI’s own consumption under high-efficiency scenarios, meaning AI could theoretically save more power than it uses, a finding the researchers themselves flag as requiring careful interpretation. Separately, researchers at Tufts University demonstrated in April 2026 that a new AI architecture can reduce the energy required to run AI models by up to 100 times without sacrificing accuracy.
AI Is Reindustrializing the Physical World
The AI boom is not just a software story. It is the largest physical infrastructure buildout since the interstate highway system. Every frontier model runs on a foundation of power transformers, substations, transmission lines, cooling systems, and the copper and rare earth metals that connect them. Morgan Stanley Research projects a power shortfall of roughly 49 gigawatts against projected US data center demand by 2028, a gap no algorithm can close. It requires steel, concrete, and industrial timelines. The United States has not faced an infrastructure expansion challenge of this scale since rural electrification. The AI buildout increasingly resembles the railroad era more than the software era.
The national security implications run deeper than most coverage acknowledges. AI dominance requires energy dominance. Energy dominance increasingly requires grid modernization. Grid modernization requires transformer manufacturing. Transformer manufacturing depends on mineral supply chains. And mineral processing, for the rare earths, copper, and specialty alloys that run through every element of this chain, is heavily concentrated in China. Beijing’s April 2025 export licensing restrictions on seven heavy rare earth elements were not an isolated trade action. They were a demonstration of leverage over the entire physical stack that AI depends on. The AI race is quietly becoming a grid race, a transformer race, and a minerals-processing race. The software industry is discovering, with some urgency, that it is now in the infrastructure business.
Geography is being rewritten in the process. AI data centers increasingly cluster around cheap power, abundant land, cooling access, and transmission proximity. Electricity-rich regions such as Nevada, Texas, Wyoming, parts of Appalachia, and the Nordic countries are becoming the next digital capitals. The next generation of technology hubs may be determined less by venture capital density than by megawatts available at the substation. AI is forcing the software industry into the capital intensity of the utility industry. The geography of computing is being pulled back toward the physical world.
Some of that geography is now moving offshore. Oregon-based Panthalassa, backed by Peter Thiel in a $140 million round announced in May 2026, is building self-propelled floating platforms called Ocean-3 that generate electricity from wave motion and run AI computing onboard. No land. No fuel. No grid connection. Results are transmitted to shore via low-Earth-orbit satellite. The ocean provides both the power and the cooling. The first units are targeted for deployment in the northern Pacific by August 2026. Panthalassa estimates 2,500 terawatt-hours of wave energy sit untapped along US coastlines alone. If even a fraction of that becomes recoverable, the geography of AI infrastructure expands from a land-constrained problem to a planetary-scale resource question.
What This Means for Infrastructure Finance
The research direction across all six domains is increasingly validated. The less-examined question is what happens next: when these technologies work at scale, who finances them, and how?
Many of these projects may ultimately follow the same financing logic that has underpinned every prior generation of first-of-a-kind energy infrastructure. The power purchase agreements already in place at Helion, Commonwealth Fusion Systems, and Fervo Energy are not just commercial milestones. They are the foundation of a project finance capital stack. A developer with signed offtake from Microsoft, Google, or a major utility has the same fundamental financing asset as an LNG terminal developer with a signed capacity agreement.
Brookfield’s involvement signals growing institutional confidence that fuel cells may become financeable infrastructure at hyperscale. When a firm of that scale underwrites a global AI factory portfolio on fuel cells, a new asset class has crossed from interesting to bankable.
Across all six domains, the pace of AI-driven development is outrunning the institutions built to finance and regulate it.
Many of these claims may not survive contact with execution. Fusion timelines have disappointed before and may disappoint again. Geothermal scaling at the speed Zanskar implies has not yet been demonstrated at national scale. Transformer bottlenecks may worsen before AI-assisted engineering compresses them. Permitting reform, the single most powerful lever available, has stalled repeatedly across administrations. And AI’s own energy efficiency gains, while real, have not historically outpaced demand growth; there is no strong reason to assume they will now. A peer-reviewed ASU study published this month found data center waste heat is already raising air temperatures in downwind Phoenix neighborhoods by up to 4 degrees Fahrenheit, in a city that recorded a 97-degree overnight low in 2023. The infrastructure problem is not only about watts. It is about heat, water, and the communities being built around these facilities.
But the deeper risk may not be execution failure at all. If AI makes power grids more efficient, accelerates materials discovery, improves data-center operations, and reduces the cost of computation, the result may not be lower electricity demand. It may be the opposite. In economics, Jevons Paradox describes the phenomenon where efficiency gains make a resource cheaper to use, which increases total consumption rather than reduces it. The National Academies has noted this risk explicitly in the context of data-center efficiency and electricity demand. AI may become the most consequential modern test of that principle. If intelligence becomes cheaper, more industries will use more of it. The question is not simply whether AI can reduce the energy intensity of computation. The question is whether civilization will consume every efficiency gain by asking machines to think more often, in more places, for more purposes.
For two decades, the digital economy appeared detached from the physical world. That detachment is ending. The next technology supercycle will not be constrained by code. It will be constrained by electricity, steel, copper, transformers, and the speed at which civilization can build.
The deeper paradox is that AI may not reduce humanity’s demand for energy at all. It may accelerate it.
Infrastructure is fate.
Sources: Acta Materialia, DuctGPT, Ames National Laboratory (2026); Nature Communications, University of New Hampshire (October 2025); MIT Technology Review, Zanskar geothermal (December 2025); ScienceDirect geothermal lifecycle review, 183 papers (November 2025); Carbon Credits, Fervo Energy IPO (2026); Science / KIT, perovskite inverse design (January 2025); Applied Physics Reviews, AI green hydrogen (September 2025); Angewandte Chemie, self-improving catalyst discovery (March 2026); Oracle / BorderPlex / Bloom Energy press release (April 27, 2026); Bloom Energy SEC filings FY2025–26; Helion Energy press release (February 13, 2026); Fusion Industry Association Global Report 2025; Wood Mackenzie, transformer lead time survey (Q2 2025); Sightline Climate, data center capacity tracking (2026); PJM Interconnection transmission expansion plan (February 2026); Markets and Markets, HVDC transmission market report (2025); DOE transmission modernization funding (2025); Power Magazine, AI grid delivery (April 2026); Nature Communications Sustainability, AI grid savings (May 2026); ScienceDaily / Tufts University, AI efficiency (April 2026); IEA Energy and AI (April 2025); Brookings Institution (April 2026); PowerLines Utility Capital Expenditure Analysis (April 2026).
David Goodnight is the founder of The Goodnight Group and Comnet International, and author of Financing the World We Trade In (2026). He has arranged more than $3 billion in project financings across twenty countries. A portion of book sales are contributed to the David Goodnight Scholarship Fund.