While markets seem immune to gravity, they aren't actually immune to physics. If the hype cycle is right about anything: AI will need massive amounts of electricity. Most people assume that this outsized (and largely outsourced) input demand will curtail growth and ultimately crush profitability. They're wrong. AI companies will never pay a penny for their electrical use. But you will.
America doesn't have a power grid. It has three. At least, three main ones: the Eastern Interconnection, the Western Interconnection, and ERCOT, or simply: Texas. There are also Alaskan grids, Hawaiian grids, and a smattering of municipal or tribal micro-grids.
Any grid's ability to deliver power comes down to the same three principal components: generation, transmission, and storage capacities. In other words, how much can the grid produce, how much can be sent, and how flexibly can it be adapted for demand. A grid is only ever as good as its weakest link. And our grids? They aren't so good for AI.
The U.S. has roughly 1.25 terawatts of nameplate generation capacity. On a typical day, that's about two to three times what we actually use. But not all of that capacity is available on demand, and not every day is typical. On a peak demand day, the grid can see a 750 GW load against an actual available capacity of about 850–900 GW. That's workable, but barely—recommended reserve margins are 20%. Once you drop below about 10%, brownouts become likely. So while we’re usually well below maximum load, the true headroom at peak is much tighter than it looks on paper.
But generation isn’t the whole game. We also need to get the power from where it’s produced to where it’s consumed. The U.S. has over 600,000 miles of high-voltage lines, but they’re unevenly distributed and often congested. Over long distances, resistance causes electricity to leak as heat, so the farther you send it, the more you lose—making distant surplus less useful in a local crunch. The three major grids are connected by only a handful of low-capacity DC ties, which makes shifting large amounts of electricity across regions slow and limited. In many cases, peak-day constraints aren’t about a lack of generation—they’re about not being able to move enough of it to the right place, at the right time, which is why brownouts are typically localized.
Electricity moves through the grid at nearly the speed of light, so the power you’re using right now is being generated right now. The only way to buffer that instant balance between supply and demand is through storage. Today, that mostly means batteries or pumped hydro—literally using surplus power to pump water uphill, then releasing it through turbines when demand spikes. Storage smooths short-term fluctuations and can shift energy from one part of the day to another, but current U.S. storage capacity is tiny compared to daily consumption, making it a supplement to generation, not a replacement.
The energy demand from AI datacenters breaks all three of these constraints. First, it requires more power generation. Fine most days, but unviable during peak demand. Second, wherever you place the datacenter, it dramatically changes transmission requirements—and high-voltage infrastructure can't be quickly relocated. What’s more, the steady demand increases the storage requirement of any system with variable output, like systems that rely on the sun shining, the wind blowing, or the river flowing. This won't work.
The optimal solution is a nearby source that reduces transmission constraints, with steady output that reduces storage needs, and dedicated capacity that reduces competing demand for the output. The guys smart enough to build AI are smart enough to recognize that there's an obvious candidate: modular nuclear reactors. These small modular reactors (SMRs) are sized for a data center or small city, offering the reliability of traditional nuclear plants but with designs that reduce risk, cut complexity, and shorten deployment timelines. Once you get one approved, your design can essentially be mass produced, requiring only site-specific review for subsequent installations. And despite lingering nuclear unease, a vestige of Cold War nuclear fear, you can be sure that when a solution solves economic, political, and geostrategic problems, the right arms will get twisted—they'll be approved.
So how does this generate free electricity? It doesn't. Considering the technical, regulatory, and implementation hurdles, it's likely that SMR startups will require hundreds of millions, if not billions, in capital to reach broad penetration, even in the relatively niche datacenter domain. But there's an obvious source for this capital—and a highly motivated, self-serving one at that—the AI companies. For them, the investment makes obvious sense because, for them, it actually could be free electricity.
Startup valuations are never "right" in any conventional sense. They have technical risk, no real track record, and project unsustainable growth. Pricing them by traditional metrics like price-to-earnings ratios generates outlandish valuations, like 100x top-line multiples. In this case, though, that's actually the benefit.
Imagine a hypothetical SMR startup. They have real engineering, plausible designs which can be reviewed and vetted, but not a lot of capital to navigate the time and complexity of regulatory approval, pilot implementation, and necessary iterations. Most importantly, they don't have any customers. This company isn't worth very much. They have no revenue, no market validation, and no survival odds without a capital injection. But if they get that capital injection, from a politically critical investor who just so happens to agree to be their anchor customer, and offers a huge upfront contract or letter of intent, suddenly this company is worth a lot. Probably a whole lot.
So an AI company invests $100M in exchange for 25% of the SMR startup. This values the startup at $400M—reasonable for a company with promising technology in a massive market, but no customers or revenue yet. Then comes the post-investment transformation: the AI company offers a $200M power purchase agreement with an as-soon-as-available timeline and becomes their rabbi through the regulatory approval process.
Now the startup has everything investors love: $200M in booked revenue, regulatory approval that creates a competitive moat, and customer validation. Startups with these credentials routinely trade at 50-100x revenue multiples—those initially ludicrous valuations you see across high-growth sectors. Apply even a 10x multiple to the $200M contract, and suddenly you have a startup with a $2B market cap.
The AI company's original $100M investment? It's now worth $500M (25% of $2B), completely offsetting the energy contract. But here's where the financial engineering adds leverage. The startup still needs hundreds of millions in additional capital ]to scale production, but with a $2B valuation, raising additional capital becomes easy. They can raise another few hundred million while only diluting existing shareholders marginally.
Meanwhile, the AI company can borrow cash using their $500M-valued stake as collateral to fund additional power purchase agreements with the SMR startup and participate in subsequent financing rounds to maintain their stake. More contracts increase the revenue base, which increases the valuation, which increases their stake value, which increases their borrowing capacity for even more contracts. They can write purchase agreements whenever they want—each new contract boosts their stake value, giving them the borrowing capacity to actually pay for it. The investment returns from this loop exceed the actual cost of electricity, effectively making power a profit center rather than a cost.
Eventually, multiples will compress and the merry-go-round will end. When that happens, the SMR startup will IPO, and the AI company will be able to rinse their profits through the liquidity of the public markets. With their investment realized, they begin repaying the borrowed money at whatever pace the public market can absorb the sales. The company keeps locked-in energy contracts, likely at reduced initial cost, completely offset by their investment gains. This tidy little trick is perfectly legal, and any challenges to the legality are a trivial cost of doing business at this scale.
There are of course, risks. It's possible that the SMR design doesn't work as intended, takes longer to refine, or faces more public outcry or regulatory delays than anticipated. There's also the economic drag of lending rates and the multiple-compression that occurs if the startup can't quickly expand to multiple customers. But at their core, these are implementation details. The kind that multi-hundred-billion dollar companies like OpenAI, Anthropic, Microsoft, Amazon, Meta, and X deal with routinely. They have armies of implementers which can be unleashed at these problems as a matter of course. In the end, free energy is a prize that solves for a lot.
Is it already happening? Sam Altman invested early in Oklo, a small modular nuclear reactor developer, and served as board chair until April 2025. He stepped down from that role amid reports of energy-related discussions between OpenAI and Oklo. So yes—it's already happening, at least indirectly. He’s also invested in Exowatt, a startup building solar-powered, heat-storage modules aimed at AI data centers, and in Helion Energy, a fusion startup, where he has served as chairman. Bill Gates’ nuclear startup, TerraPower, which is developing advanced sodium-cooled SMRs, received significant investment from NVIDIA's venture arm. Amazon has announced SMR partnerships and financing agreements with X-energy.
We're at the early stages, but we can see which way the wind’s blowing. This is the investment stage of the process, not the validation stage. There are many hurdles left before they can claim their free energy, but they look surmountable. The most successful investors and engineers in the world are not pouring cash into snake oil, they're front-running. Meanwhile, the public complaints grow about rising energy costs, setting up the perfect backdrop for the penultimate act, when the AI companies step in to collect thanks for solving the problems they created themselves. At enormous profit.
The final scene will play out as it always does, when the scheme unwinds and is offloaded onto the public. After capturing the hyper-growth phase of SMR adoption, these investors will inevitably seek a return of capital by selling their stakes in public markets. Whether or not you buy these companies, your pension managers, index funds, and mutual funds will. All of the costs built into the valuation will be baked into the price you pay, while your energy costs will continue to be borne by you. In a best case scenario, this de-risks a novel energy solution that ultimately lowers everyone's costs, improves grid reliability, and de-carbonizes the infrastructure of our nation. But even if that plays out, until SMRs are rolled out at scale, the risks and costs will increase for everyone. And yours won't be offset by their investment gains, they'll be their investment gains.