The next decade of artificial intelligence will be decided by who can build power, finance it, connect it to the grid, and bring the neighbors along. That is the work I have done for more than twenty years.

By Neil Auerbach, Chairman and CEO, Hudson Sustainable Group

Artificial intelligence is now limited by electricity, and the limit has two parts. The first is power: the plants and wires that generate and deliver energy. The second is capacity: power that is available in the few hundred hours a year when the grid is stressed. Both are short. The models are ready, and the physical system behind them is not. I have spent more than twenty years financing this kind of buildout, first building the U.S. renewable energy investment business at Goldman Sachs and then at Hudson Sustainable Group. What I see today is one of the most important shifts of my career, and it rewards the people who know how to get projects built.

The intelligence is here. The buildout is the work.

On a recent episode of Moonshots with Peter Diamandis, investor Chamath Palihapitiya put it plainly: “I do think that we have achieved AGI. It cannot diffuse.”

Whether or not one accepts the label, his second sentence is the one that matters. Intelligence in a lab changes nothing until the physical system behind it is built, and today that system lags far behind the software. Every new model needs data centers, and every data center needs a steady supply of electricity. The winners of the next ten years will be the firms that can deliver that electricity on time and on budget.

How much power does AI need?

The numbers are moving fast. The U.S. Department of Energy and Lawrence Berkeley National Laboratory found that data centers used about 4.4 percent of all U.S. electricity in 2023, or 176 terawatt hours, and project that share could reach 6.7 to 12 percent by 2028. The Electric Power Research Institute puts the 2030 range at 9 to 17 percent. In plain terms, data center demand could nearly double or more than triple within five years, on top of everything else Americans are electrifying, from cars to homes to factories.

Will efficiency solve it?

The most common objection is that it will: models are getting smaller and more efficient, so the same work will take less electricity. The first half is true. The conclusion does not follow.

Economists have known why since 1865, when William Stanley Jevons observed that more efficient steam engines did not reduce Britain’s coal consumption. They increased it, because cheaper power found new uses faster than efficiency removed old ones. The same thing happened to electric lighting: three centuries of efficiency gains left the world using more light, not less. Every time intelligence gets cheaper per query, the number of queries, agents, and workloads worth running grows faster than the savings. Efficiency does not shrink the demand for power. It widens the market for it.

That is why I expect power to remain scarce in the United States for years, even as the models improve, and why the queue, the permits, and the transformers, none of which move at software speed, will keep setting the pace.

Where does the buildout slow down?

The power exists on paper. Getting it onto the grid is the hard part. Even after easing somewhat in 2025, about 2,061 gigawatts of generation and storage were waiting in U.S. interconnection queues at year-end. Of the capacity that asked to connect between 2000 and 2020, three quarters withdrew and about 13 percent was operating by the end of 2025. The wires are slower still: the country built roughly 900 miles of high-voltage transmission in 2024 and about 400 in 2025, against the roughly 5,000 miles a year the Energy Department’s own scenarios imply.

And the shortage is sharper than the annual numbers suggest. What the grid lacks most is capacity, meaning power available in the few hundred stressed hours a year. That is why capacity prices, not energy prices, are hitting the auction price cap and driving household bills: the PJM market, from Virginia to Chicago, cleared at the cap for the third auction running in July and fell 6.8 gigawatts short of its reliability requirement, its second shortfall in a row. I made the broader case in RealClearEnergy on November 6: the grid will not build what AI needs in time, on-site generation can supply the energy but not the capacity, and the fastest capacity in the country can be built in American homes.

Why solar and storage are leading new power

Developers have already chosen. Of the 86 gigawatts of new utility-scale capacity planned in the U.S. for 2026, solar makes up 43.4 gigawatts, or 51 percent, and battery storage makes up 24 gigawatts, or 28 percent. Together, they account for nearly four of every five new gigawatts.

The split matters. Solar supplies energy. Batteries supply capacity, the thing the grid is most short of, and they can be sited close to where power is used, which reduces the strain on the grid and shortens the path to energization. That is where the next move is.

The capacity is in the neighborhood, and so is the vote

On Election Day, November 3, voters decided 41 local data center measures across ten states. Forty of them went against the industry, most by 60 to 80 points. The one that tells you what comes next was in Perry Township, Ohio, where 87 percent of voters approved a requirement that hyperscale developers post financial guarantees before they receive incentives. Communities are not saying never. They are saying show us the money, and guarantee it, before we vote.

There is a way to do that which also solves the data center’s capacity problem, and it starts with being precise about what the data center is buying. Not the capacity that grid operators accredit and auction, but relief in the few hundred hours a year when its utility would otherwise order it to curtail, verified at the meter, from the households around it. Researchers at Duke University found that in most of those hours less than half of the new load needs to curtail, and that the average event lasts about two hours. A home battery can do that job.

A data center that needs several hundred megawatts of that relief can contract for it from its neighbors: a smart thermostat in every enrolled home and a battery in most of them, installed at no cost to the homeowner, with the data center carrying the financing on investment-grade credit and paying each household a capacity dividend, every month, for fifteen years. The homeowner gets backup power, a lower bill, and a monthly payment from the company next door. The data center gets verified capacity in its own grid zone, a basis to seek a faster connection, and a community with a direct financial stake in the project getting built. Buying that capacity from local homeowners has the potential to turn opponents into partners. We call it a community capacity compact, and it is a better answer to a moratorium than a lawsuit.

Two honest points. First, this is not the cheapest capacity per kilowatt. A gas plant is cheaper on that measure, if you can get one before 2030. Household capacity is faster and cheaper than waiting for a gas turbine: the added cost to the data center works out to a few cents for every kilowatt-hour the campus uses, a small fraction of what a year of idle chips costs. Second, it does not replace what keeps the lights on inside the campus. The batteries in the garages keep the grid from curtailing the data center; the equipment on site keeps the data center running. Those are different jobs.

It is also feasible. Roughly 250,000 home batteries are installed in the United States each year; a one-gigawatt campus needs about 22,000 a year in its own grid zone for four years, less than a tenth of that. The constraint is installers, not homeowners, and a data center that funds their training has solved its own constraint and created the most durable local benefit it can offer.

The ingredients already exist: investment-grade buyers, the third-party ownership and securitization markets that put solar on five million roofs, and a tax credit that still applies to batteries in homes. Regulators can help: FERC has ordered the grid operators to rewrite their rules for connecting large loads, with PJM’s filing due November 16, and those rules should let verified household capacity in a data center’s own zone count against its curtailment obligations. What has been missing is the contract and the financial architecture that binds all of it together, and that is what we are building. I will have more to say about it soon.

Capital decides what gets built

A good project with the wrong capital structure stays on paper. A good project with the right structure gets built. That is where Hudson does its work.

Much of our current work is in distributed solar and storage: projects measured in megawatts rather than gigawatts, sited close to load, that can often be permitted and energized in months rather than years. They add capacity in smaller, faster steps, close to the people who need it, and they are the kind of asset that rewards careful structuring over size.

The same thinking runs through our work in residential power. Hudson’s leadership helped found Sunlight Financial in 2014. Today, SolSource carries those lessons forward, financing solar and batteries for homes, including through Propel, launched this year with TriBeam Financial, with Enphase and Greentech Renewables as hardware and distribution partners.

What I would tell a developer today

Own the constraint. Power is the scarcest input in AI, and Jevons tells us it will stay scarce even as the models get leaner.

Separate energy from capacity. The grid can often supply the first. The second is what sets the price, and it does not have to come from a turbine.

Favor contracted, deliverable capacity over nominal megawatts. A gigawatt in a press release and a gigawatt that is verified at the meter and reaches a substation are different assets.

Pay the neighbors. After November 3, a community with a financial stake in your project is not a nice-to-have. It is the permit.

Value structuring. The financing plan often decides whether a project is built at all.

The AI era will be built in steel, silicon, and electricity, and it will be built in communities that have agreed to host it. The firms that know how to finance both will define it.

Work with Hudson

Hudson works with management teams on capital structuring, strategic development and institutional positioning, including as advisor, co-sponsor or co-investor, where applicable. If you have a project that is real but stuck on structure, or a community that is real but not yet on your side, that is the conversation we want to have. Explore the Hudson track record or reach out to the team directly.

This article is provided for general information and does not constitute tax, legal, or investment advice.

Frequently asked questions

Why is power the main limit on artificial intelligence?

AI models run in data centers, and data centers need large, steady amounts of electricity. New power plants and grid connections take years to permit, finance, and build. That means the speed of AI adoption now depends on how fast new power reaches the grid, more than on the models themselves.

How much electricity will U.S. data centers use?

The U.S. Department of Energy and Lawrence Berkeley National Laboratory project that data centers could use 6.7 to 12 percent of U.S. electricity by 2028, up from about 4.4 percent in 2023. The Electric Power Research Institute projects 9 to 17 percent by 2030. Either way, demand roughly doubles or triples within five years.

Will more efficient AI models reduce power demand?

Efficiency lowers the electricity used per query, but history shows cheaper computing creates more uses faster than efficiency removes old ones, the pattern economists call the Jevons paradox. As intelligence gets cheaper, more agents and workloads become worth running, so total power demand keeps rising even as models improve.

What is the difference between power and capacity?

Power, as used here, means energy delivered across the year. Capacity is power available in the few hundred hours when the grid is most stressed. A region can have enough energy and still be short of capacity on a hot afternoon, which is why capacity prices in markets like PJM have hit their price cap and are driving household bills.

Can homes supply capacity to data centers?

Yes. Home batteries and smart thermostats can discharge and flex in the few hundred hours a year when the grid would otherwise ask a data center to curtail. A data center can contract for that verified capacity from households in its own grid zone and share the value through a monthly capacity dividend.

Is household capacity more expensive than gas?

Per kilowatt, a gas plant is cheaper, but new turbines are largely sold out until about 2030. Household capacity starts delivering within months and grows in step with the campus, and for a one-gigawatt campus it costs a few cents on every kilowatt-hour the campus uses, far less than the cost of waiting a year to operate.

How can data centers win local approval after the 2026 midterms?

Voters decided 41 local data center measures in November 2026, and 40 went against the industry. In Perry Township, Ohio, 87 percent backed financial guarantees from developers. Projects that give households a direct, guaranteed financial stake, such as paid capacity contracts with batteries installed at no cost, answer that demand and turn opponents into partners.

How does Hudson Sustainable Group invest in AI power demand?

Hudson focuses on capital structuring for clean energy and power projects, including solar, battery storage, and distributed generation. Hudson works with management teams as advisor, co-sponsor or co-investor, where applicable. Its leadership has mobilized nearly $15 billion across more than 25 investments in over 25 countries since 2002.

Sources

Moonshots with Peter Diamandis, EP #304, October 21, 2026 (quote at approx. 00:44:44): https://www.youtube.com/watch?v=kAwG9E1W_ME

Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report (Dec. 2024): https://energyanalysis.lbl.gov/news/berkeley-lab-report-evaluates-increase-electricity-demand-data-centers

Electric Power Research Institute, Powering Intelligence 2026 (Feb. 2026): data centers 9 to 17 percent of U.S. electricity by 2030

Lawrence Berkeley National Laboratory, Queued Up: 2026 Edition (July 2026): https://emp.lbl.gov/news/backlog-power-plants-seeking-transmission-grid-connection-eased-somewhat-2025-amidst

Grid Strategies / ACEG, Fewer New Miles (2025): 888 miles of 345 kV+ transmission in 2024; ~5,000 miles a year implied by DOE’s most conservative scenario. Oilprice.com / Yahoo Finance (Jul. 2026): ~400 miles in 2025

U.S. Energy Information Administration, New U.S. electric generating capacity expected to reach a record high in 2026 (Feb. 20, 2026): https://www.eia.gov/todayinenergy/detail.php?id=67205

PJM 2028/2029 Base Residual Auction results (July 14, 2026): cleared at the $325/MW-day cap, 6,831 MW short of the reliability requirement: https://insidelines.pjm.com/pjm-capacity-auction-procures-138318-mw-of-generation-resources-as-work-continues-to-address-growing-electricity-demand/

2026 local data center ballot measures: The Grid Sight tally of 41 votes in 10 states (Nov. 6, 2026): https://thegridsight.com/data-center-votes-2026; CNBC (Nov. 4, 2026): https://www.cnbc.com/2026/11/04/data-center-ballot-measures-2026-midterm.html; Perry Township, OH Issue 16 (87 to 13)

Norris, T. H., et al., Rethinking Load Growth, Duke University Nicholas Institute (Feb. 2025): 76 to 126 GW of new load integrable at 0.25 to 1.0 percent annual curtailment; average events 1.7 to 2.5 hours; in roughly 88 percent of curtailment hours less than half the new load must curtail: https://nicholasinstitute.duke.edu/publications/rethinking-load-growth

SEIA / Wood Mackenzie, U.S. Energy Storage Monitor (Mar. 2026): roughly 250,000 residential storage systems installed in 2025

FERC show-cause orders to the six RTOs and ISOs on large-load interconnection (June 18, 2026); PJM compliance filing due November 16, 2026

Tsao, J. Y., et al., “Solid-state lighting: an energy-economics perspective,” Journal of Physics D (2010): lighting consumption rose with efficiency over three centuries

W. S. Jevons, The Coal Question (1865): the Jevons paradox

Neil Auerbach, RealClearEnergy op-ed (Nov. 6, 2026): https://www.realclearenergy.org/articles/2026/11/06/ai_needs_power_it_needs_capacity_even_more__and_your_house_has_it_1215218.html

Hudson Sustainable Group track record: https://hudsonsustainable.com/track-record/