The Nuclear AI Trade
Oklo touched one hundred and ninety-three dollars and change in October.
NuScale touched fifty-four in November of the year before.
Both numbers, eighteen months earlier, would have read like typos.
For anyone who hasn’t been following the corner of the market that calls itself “AI infrastructure plays,” a quick orientation.
Oklo and NuScale are the two largest publicly traded developers of small modular reactors — SMRs — the new generation of compact, factory-built nuclear power plants meant to deliver electricity in 50-to-300-megawatt blocks rather than the gigawatt-scale of traditional reactors.
The pitch is simple. Artificial intelligence needs enormous amounts of electricity. Solar and wind don’t run at night. Natural gas is dirty and capacity-constrained. Nuclear is clean, baseload, and politically back in fashion.
SMRs are the modern, financeable, listed expression of that idea.
Between January 2024 and October 2025, Oklo rose more than fifteen-fold. NuScale rose more than five-fold. The two of them together became the second-derivative AI trade: if you couldn’t afford NVIDIA at four trillion, you bought the companies that would power NVIDIA’s customers.
Today, Oklo trades around sixty-two. NuScale around twelve. Each has given back more than half of its peak.
The stocks are bleeding, quietly, while the headlines still read exactly the same way they did at the top.
Open any feed on any platform this week and the explanation hasn’t moved an inch.
Artificial intelligence needs electricity.
The hyperscalers are signing with the SMRs.
The future of nuclear is small, modular, and listed.
It’s a clean pitch. It’s been repeated for eighteen months. And it has minted some of the most spectacular charts of the cycle.
There is a word, however, that you will not find anywhere in the press releases that drove these two stocks to their all-time highs.
The word is PPA.
A PPA — power purchase agreement — is the legal contract through which a utility or large customer commits, in writing, to buy electricity from a power plant at a defined price over a defined number of years. It is the document banks lend against. It is the document that turns a power project from a press release into an asset. Without one, a power plant has a customer in the same sense that a restaurant with an unsigned reservation book has diners.
The absence of that word in the SMR coverage is the entire trade.
The story
There is a perfectly reasonable bull case here, and it deserves to be laid out cleanly before anyone starts poking at it.
Hyperscaler capex in 2026 is approaching three-quarters of a trillion dollars, and a meaningful slice of it is being spent on power, not silicon. Datacenters need baseload. Solar and wind don’t provide it. Natural gas is dirty and capacity-constrained. Nuclear is clean, baseload, and after twenty years of public hostility, the political weather has flipped — the Trump administration extended a billion-dollar loan to restart Three Mile Island, the Nuclear Regulatory Commission was ordered under Executive Order 14300 to compress reactor licensing to eighteen months, and the Department of Energy now runs a Reactor Pilot Program that lets approved developers build under federal authorization rather than wait in the standard regulatory queue.
Then there are the customer announcements.
In January, Meta announced agreements with three nuclear suppliers — Vistra, Oklo, and TerraPower — for up to 6.6 gigawatts of power for its AI datacenters, on top of an existing deal with Constellation. Oklo got 1.2 gigawatts in Pike County, Ohio. NuScale has its Tennessee Valley Authority partnership and a project in Romania. X-energy is sitting on an Amazon deal. Kairos has Google.
In April, Oklo announced a three-way collaboration with NVIDIA and Los Alamos National Laboratory to develop what the press release called “nuclear-powered AI factories” — combining Oklo’s sodium fast reactor design with NVIDIA’s AI infrastructure and the lab’s nuclear fuel expertise. The stock jumped nineteen percent on the day.
Oklo broke ground at Idaho National Laboratory in September. Management is targeting criticality — the moment a reactor sustains its first nuclear chain reaction — with the Aurora powerhouse by July 2026, with commercial operations by 2028. NuScale ended the first quarter with one billion dollars of cash on the balance sheet. The DOE just signed off on fuel work for Oklo’s Pluto reactor design.
If you read only those paragraphs, you buy the SMR complex.
That is also where almost every piece of coverage on these stocks stops.
The part printed in light gray
On January ninth, Meta announced three nuclear deals on the same day. Same buyer. Same press release. Same coverage cycle.
Two of those deals are not the same kind of contract.
The Vistra deal was filed with the SEC three days later as an 8-K. It is a twenty-year power purchase agreement — a binding, contractual commitment to deliver 2,609 megawatts of carbon-free power and capacity from Vistra’s existing nuclear plants in Ohio and Pennsylvania. Delivery begins in late 2026. Full operating delivery by year-end 2027. Existing reactors, already operating, already feeding the grid. Real electrons. Lawyered terms. Locked price.
The Oklo deal was filed with the SEC as something different. The disclosure language calls it a “Prepayment Agreement” — a mechanism for Meta to prepay for power and provide funding to advance powerhouse deployment. Oklo will use Meta’s funding, the filing says, to secure nuclear fuel. First phase targeted to come online “as early as 2030.” Full 1.2 gigawatts by 2034.
Both deals were celebrated the same morning by the same desks at the same banks.
One of them is a contract to buy electricity from a plant that already exists. The other is a check that helps a pre-revenue company buy uranium for a plant whose first phase, on a best-case schedule, comes online in four years, with full delivery in eight.
This is the sentence I keep waiting to read in the mainstream coverage and never quite do.
The market is pricing both of these as the same victory.
They are not the same victory. They are not the same product. They are not even the same risk.
That asymmetry is the first thing.
The second thing is the dilution.
When a young company without revenue needs to fund itself, it sells more shares. Each new share issued cuts the ownership stake of every existing shareholder a little smaller — that’s dilution. The cleanest version, called an “at-the-market” offering, lets the company sell shares directly into the market over time at prevailing prices. No announced placement, no marketed deal, just a steady drip.
On May thirteenth, Oklo filed an 8-K announcing a new at-the-market equity offering of up to one billion dollars. On the same day, the company terminated its previous facility — a one-and-a-half-billion-dollar program under which Oklo had already placed nearly sixteen million shares for gross proceeds of about $1.5 billion, opened only five months earlier in December.
Read that paragraph slowly.
One and a half billion dollars raised in five months. Through a company with a market capitalization of around twelve billion. With zero revenue. And the moment the program was nearly exhausted, the company opened another billion-dollar facility to keep going.
For a business with a published 2026 cash burn guidance of eighty to one hundred million dollars, that is twenty-five years of operating runway raised in five months. Which means it isn’t operating runway.
It’s pre-construction capital for actual reactors — a category of spending several orders of magnitude larger than anything currently reflected in operating expenses, and which has to come from somewhere.
In Oklo’s case, somewhere is the holders of the stock.
Strategically, this is elegant. Hyperscalers won’t write checks for an asset that doesn’t exist, so the developer raises equity to build the asset, and once the asset exists the hyperscaler pays for the power. Tesla did a version of this. Every capital-heavy growth story does. There is nothing inherently wrong with it.
But there is a difference between buying a stock where the customer pre-pays for the product, and buying a stock where the shareholders pre-pay for the customer’s product.
The market is pricing the first.
The cap table is doing the second.
The third thing is what happens at NuScale every quarter, in plain English, on a line nobody reads.
NuScale has a strategic commercialization partner called ENTRA1, which holds the rights to commercialize NuScale’s modules in certain markets — a structure not unlike a master franchise. In the third quarter of 2025, NuScale’s general and administrative expenses exploded by more than three thousand percent to $519 million, driven by a $495 million payment to ENTRA1 tied to the Tennessee Valley Authority development agreement. In the first quarter of 2026, the company made another $260 million payment to ENTRA1 as part of “Milestone Contribution 1,” contributing to $315 million of operating cash outflow on the quarter — against $565,000 of revenue.
Roughly seven hundred and fifty-five million dollars, in nine months, paid out to a commercialization partner before the company has booked a meaningful nuclear-equipment order. Management has stated that total payments across all six TVA projects could reach several billion dollars before NuScale receives any equipment orders.
This is not a hidden fact. It’s in the filings. It’s just not in the pitch.
The fourth thing is the regulator.
In January 2022, the Nuclear Regulatory Commission ended its review of Oklo’s first combined license application and denied it, without prejudice. Oklo regrouped, redesigned, scaled the reactor up from 1.5 megawatts to 75, and refiled. For the Idaho project, the company is now proceeding under Department of Energy authorization through the Reactor Pilot Program rather than the standard NRC framework — using the DOE’s authority under the Atomic Energy Act to exempt reactors built “under contract with and for the account of” the department from traditional NRC licensing requirements.
This is legal. It’s also strategically correct for accelerating the first deployment. But the second, third, and fortieth deployments — the ones at Meta’s Pike County campus, at the hyperscaler buildouts that justify the valuation — those don’t run on DOE pilot exemptions. They run on commercial NRC licenses. Which Oklo has filed for, once, and been denied.
The fifth thing is the insiders.
Oklo’s co-founder and CEO Jacob DeWitte has executed 82 sales and zero purchases over the company’s life as a listed entity, disposing of roughly 2.4 million shares for an estimated $189 million. The CFO has executed 7 sales and zero purchases for about $13 million. The co-founder and COO has a net sale of roughly 3.8 million shares over eighteen months.
The sales are filed under what’s called a 10b5-1 plan, meaning they were programmed in advance and aren’t trading on inside information. That distinction matters. It’s also worth knowing that the plan governing these sales was adopted on March 31, 2025, when Oklo’s shares traded near ten dollars. Selling 100,000 shares at ninety dollars under a plan adopted at ten is not insider trading.
It is also not the behavior of people who think the stock is undervalued.
The thing nobody puts in a graph
If artificial intelligence needs electricity in 2026, in 2027, and in 2028 — and the consensus says it does, and the capex commitments say it does — the most useful question for an investor isn’t whether SMRs will eventually power AI factories. They probably will. Eventually.
The useful question is what powers the factories that are already breaking ground.
In January 2026, Crusoe and Tallgrass broke ground on the 1.8 gigawatt Cheyenne AI Factory. They chose Bloom Energy’s solid oxide fuel cells — a non-combustion power technology that runs on natural gas — for the first 900 megawatt phase, specifically because the technology could bypass the interconnection queue entirely. The first data halls go online by late 2026.
Eighteen months. Not eight years.
Simple-cycle gas turbines, the dirty fallback nobody likes to talk about, can be built in eighteen to twenty-four months. Across the PJM grid — the regional power market covering thirteen states from the Mid-Atlantic to the Midwest — sixty percent of fossil-fueled plants that had been slated for retirement have postponed or canceled. Capacity prices in one auction rose more than eight hundred percent. The grid is keeping the old fleet alive because the new fleet — gas, SMR, fusion, doesn’t matter — cannot arrive in time.
The bankable timeline for SMR power as the primary source on a datacenter financing — meaning the timeline at which a lender will accept SMR megawatts as security on a debt facility — opens around 2028, and only for projects with confirmed positions in a vendor’s orderbook today.
Which means the implicit bet inside the SMR equity trade is not “AI needs power.”
It is “AI customers, knowing that fuel cells deploy in eighteen months and gas peakers in two years, will choose to wait six to ten years for a first-of-a-kind small modular reactor that has never run commercially.”
Some of them might. A handful of hyperscalers with carbon commitments and infinite balance sheets — Meta, Google, Microsoft — can afford to sign options on every horse in the race and let one or two win. That is plausibly what the deals announced so far actually are: cheap call options on a long-dated technology, structured as prepayments and memoranda rather than PPAs, with no termination penalties if the project slips by three years.
Which is what NuScale’s Romanian project already has. Initial deployment target: 2029. Current target: 2033. Four years of slip before the first shovel breaks ground.
A call option is a perfectly reasonable thing for Meta to buy.
It is a much harder thing to value at two hundred and fifty times forward sales.
The price
NuScale generated five hundred and sixty-five thousand dollars of revenue in the first quarter of 2026, missing the $14.8 million consensus by ninety-six percent and falling 95.5 percent year over year.
Oklo generated zero.
For these two facts, the market is paying:
Oklo, at roughly sixty-two dollars, holds a market capitalization of about $10.8 billion. Against analyst-projected 2027 revenue of approximately sixteen million dollars, that is more than six hundred times projected 2027 sales. Against 2028 sales projections near forty-eight million, it is roughly two hundred and fifty-three times.
NuScale, at twelve dollars, trades at a market cap of about $3.9 billion — roughly nineteen times its 2027 sales projection.
For context, the company we wrote about last week — AMD, the stock everyone agrees is expensive, the stock trading seventy-three dollars above its average analyst target — trades at roughly sixty times its next twelve months of earnings. The company generating ten billion dollars of revenue every ninety days, with a contracted deal book that already extends into 2030, is roughly five times less richly valued than the company that has not yet sold a single kilowatt-hour.
The analyst dispersion on Oklo is also worth pausing on. The average price target is ninety-one dollars. The low is fourteen. The high is one hundred and forty.
Ten-to-one ranges in price targets are not the sign of a misunderstood compounder. They are the sign of an asset that nobody on the sell side actually knows how to model — because the asset has no operating data, no commercial license, no contracted revenue, and no comparable peer that has ever crossed from pre-revenue to scale in this technology. It is the dispersion you see in early-stage biotech, where one model gives you ninety percent terminal-value compounders and another gives you zero, depending on whether a single regulatory decision lands.
That is what holding Oklo at sixty-two dollars actually is.
It is a long-dated regulatory call option, traded with the volatility of a stock, and the valuation of a software company.
What is actually working
The bull case isn’t nothing, and it deserves more than dismissal.
The NRC reform is real. The DOE pilot program is real. The cultural pivot — hyperscaler CEOs talking publicly about restarting nuclear, the political winds, Three Mile Island coming back online — is real. Aurora-INL has broken ground. Criticality is targeted for July of this year, eight weeks from now, and if it hits on schedule it would be the first new commercial fission reactor designed and delivered in the United States in a generation. The Meta prepayment brought in real cash. The Romanian project, slipped though it is, has not been canceled. The TVA partnership has structure. NuScale, alone among western SMR companies, has a fully NRC-design-certified module.
If you believe the AI capex cycle runs through 2032 and beyond, and you believe nuclear baseload eventually wins a structural share of datacenter power, the SMR complex gives you exposure that almost nothing else does — at a fraction of the market cap of the incumbent power names.
That case is intact.
What is actually risky
A market cap of $10.8 billion for a company with no revenue. A forward multiple of 253 times on 2028 sales. A $1.5 billion equity program exhausted in five months and replaced with another billion the same day. A $755 million cash outflow to a single commercialization partner over nine months with no equipment orders booked. A first NRC license application denied in 2022. A first commercial deployment now running through the DOE rather than the NRC because of that denial. A first-of-a-kind reactor whose criticality milestone is eight weeks away. A 10b5-1 plan adopted at ten dollars that has sold $189 million of CEO stock at higher prices. A customer deal structure that, on the same day from the same buyer, paid Vistra in a binding PPA and Oklo in a prepayment.
One of those flashing yellow is digestible.
Two at once is uncomfortable.
Six at once is the kind of moment retrospectives are written about.
So — buy, hold, or short?
The honest answer, as last week, doesn’t fit on a bumper sticker.
It depends on which name. Oklo at sixty-two and NuScale at twelve are not the same trade. One has no revenue and a 253 times multiple. The other has a real, if tiny, revenue line and trades at nineteen times sales — a number that, in any other sector, would not be called cheap, but in this one is the only thing close to a fundamental anchor.
It depends on whether you separate the SMR equities from the broader AI-nuclear thesis. Vistra, Constellation, Talen — the existing-reactor names that signed the actual PPAs — trade at single-digit and low-teen earnings multiples, generate cash today, and supply electrons that hyperscalers are paying for now, not in 2030.
It depends on what your view is of July fourth.
That’s Oklo’s criticality target. If it hits, the entire risk framework I just laid out re-rates. If it slips, the entire framework gets harder.
In the premium section, I lay out the cash runway math for each SMR name down to the month, the specific NRC and DOE catalysts that will land between now and year-end, the price levels at which I become a buyer, a seller, or a shorter on each, the pair trade I think the market is mispricing between the SMR complex and the existing-reactor utilities, and — at the very end — the position I’m taking…