What Bring Your Own Generation Actually Means

America has one market that has always guaranteed its product to any customer who asks for it. Energy. That is about to change, and people should care.

The guarantee is called the obligation to serve, and it is close to a hundred and fifty years old. A utility receives a legal monopoly over a territory. In exchange it must serve all customers, keep service safe and reliable, and submit to rate limits. If demand increases, the utility is required to build for it. If serving one customer costs more than serving another, the difference is spread across everybody's bill.

That is why a farmhouse at the end of a ten mile line pays roughly what an apartment in a dense city pays. It is why rural America has electricity at all. Almost no other business in the country operates under a rule like this. A grocer can turn you away. A bank can decline you. Your utility cannot.

In the summer of 2026, federal regulators began writing rules that end that guarantee for one class of customer. The mechanism is a shift called bring your own generation (BYOG). Instead of the system taking responsibility for procuring power, the largest new customers will have to arrive with their own supply already secured, their own collateral already posted, and their grid fees paid whether they draw from the grid or not. A potential outcome based on the filings with the regulator is a stop to buying power on behalf of customers.

The current system relies on the cost-causation principle, meaning costs should be paid by the customers who cause the utility to incur those expenses, but shared grid benefits, social policy goals, and other limits result in those costs being spread across ratepayers.

Why this reaches past the companies involved

The outcome decides who pays. Someone has to fund the next wave of power plants. The rules being written this summer will determine whether the cost lands on the new customers driving the demand or gets spread across every household bill in the region, the way it always has been.

It decides whether the lights stay on. Reserve margins in parts of the country are already below levels grid operators say is safe. Shifting the obligation to procure power does not create power. If nobody builds, the shortage is still there.

It sets a precedent. Once the duty to serve becomes conditional for one class of customer, the principle that made American electricity universal will have changed. That is a decision worth making deliberately rather than in a docket that is generally read and followed by those with a vested interest.

The decision also has consequences the industry has largely overlooked. The same rules will determine whether America builds a domestic nuclear manufacturing capability over the next decade, or strands one at a cost nobody can afford. That is the argument this piece makes, and it takes some building.

America Cannot Produce the Power It Has Already Promised

Start with the shortage because everything else follows from it.

The new customers in question are very large single sites, mostly computing facilities. One can consume as much electricity as a small city, continuously, day and night, with no seasonal lull and no tolerance for interruption. Dozens are queued up to connect, and they want to be running by 2027 and 2030. Whether that projected demand is as solid as it looks is a real question, and Section VI takes it seriously.

In the mid-Atlantic and Midwest region alone, roughly 30 gigawatts of new demand is projected by 2030. That region's spare capacity margin is already below the level its own operator says is safe.

Now look at what it takes to serve that.

• Gas. You need a turbine. Three companies make most of the world's large turbines. Their order books already run several times what their factories can build in a year, prices have roughly tripled in three years, and expanding a turbine factory takes three to five years. Order today and you are realistically waiting until the early 2030s. Make no mistake though, the orders are being placed.
• Nuclear. The earliest even semi-credible date for a new small modular reactor (SMR) is 2030, and that is the optimistic end of a range that has slipped before.
• Batteries. Not available at anything close to the scale required.

So the demand lands between 2027 and 2030, and the supply lands in the 2030s. The arithmetic simply does not work. Not because anyone made a mistake, but because electricity infrastructure takes about a decade to build and these requests appeared in about two years.

That leaves regulators with a question America has spent a century refusing to ask: When there isn't enough power, who gets cut off?

The System's Answer Is to Stop Promising

For a hundred years the deal was simple. If you wanted electricity, you connected, and the grid was obliged to find the power. The cost of building the plants was spread across everyone's bill. In the summer of 2026, that promise began to be withdrawn.

On June 18, the federal regulator told all six of America's regional grid operators that their rules for connecting very large customers may be unfair, and ordered them to justify or fix them. On July 20, the operators filed their answers to a direct question. How, exactly, will you make sure enough power exists? What came back was mostly a list of future meetings.

One operator, New England, told the regulator plainly that it intends to stop buying power on behalf of these customers altogether.

The emerging replacement has a name, bring your own generation. You connect only if you show up with your own power secured, your own collateral posted, in some deals roughly a decade's worth of expected energy costs, and your grid fees paid on your full power appetite whether you draw from the grid or not.

Electricity has stopped being something you buy. It has become something you build, and prove you can afford, before you earn a dollar.

So how does this impact nuclear as baseload power?

Read that as a nuclear story and it looks like a gift. For the first time, buying reliable long-term power isn't optional. It is a condition of being allowed to switch on. At first glance, that is exactly the guaranteed demand the nuclear industry has always needed; however, the opposite is far more likely, for reasons that have nothing to do with whether anyone wants nuclear power.

The economics of a small modular reactor work like this. The first unit produces electricity at $80 to $150 per megawatt-hour, which loses to solar, wind, and gas by a wide margin. The more attractive number, $50 to $80, only shows up around reactor twenty, after the same design has been built enough times to overcome the learning curve.

The cost of a reactor is not set by how much power the world wants. It is set by how fast orders arrive.

That single sentence is the whole risk. A reactor company can be sitting on record announced demand and still be finished because if orders merely slow down, the factory never gets built, the cost never falls, and every unit stays stranded at a price nobody will pay. This isn't hypothetical. It already killed the most advanced SMR project in America, when twenty-three of thirty-five utility customers walked away from NuScale over cost.

Now put the two halves together. Bring your own generation makes each power contract more binding, but it also raises the cost of building AI capacity. Nuclear cannot meet the required timeline, six regions are each writing their own rulebook, while forcing everyone to spend their available money on gas turbines first.

It makes each contract harder to escape and each contract less likely to be signed. Which is why the number that matters here is not how much electricity America needs, and not whether the demand forecasts turn out to be right, but rather how fast orders are accumulating, and unlike almost everything else in the AI trade, that number is set by policy, not by markets. It is being set right now, in filings due in August. Which brings us to why this pattern should look familiar.

This Has Happened Before

We have watched an industry destroy itself while demand for its product was still rising. That history offers a useful comparison because the mechanism is identical.

In the summer of 2006, everything looked fine. House prices were at record highs and still climbing. And subprime borrowers had already started to default.

Those mortgages were never meant to be paid off. They were cheap for two or three years, then the rate jumped. The plan was that by then your house would be worth more, so you'd borrow against the gain and refinance into another cheap loan. A treadmill, powered by rising prices.

So the loans didn't need prices to fall. They needed prices to stop rising quickly. Around 2005 and 2006, price growth slowed from roughly 15% a year to something much smaller. Still positive. Still gains. Just slower.

Slower wasn't enough. People couldn't refinance. Defaults began at the peak of the market.

The mortgage was priced off the rate of growth. The reactor is priced off the rate of orders. Same structure, different industry. Which gives the rule the rest of this piece rests on.

If something can only stay afloat by borrowing against growth, it doesn't need growth to stop. It just needs growth to slow down. And there's a cruel gap in the middle. Between the moment growth starts slowing and the moment it turns negative, everything still looks excellent: order books at record highs, every headline number up. But the business is already broken. It simply hasn't been reflected yet.

There are roughly 130 small reactor designs in development today sitting in exactly that gap.

Levels, Speed, and Acceleration

Three different things, and people constantly confuse them.

Think about driving a car.

• The level is where you are, meaning how many miles you have covered.
• The speed is how fast you're going.
• The acceleration is whether you're pressing the gas or easing off.

Wall Street watches the level. Sometimes it watches the speed. It almost never watches the acceleration. But acceleration is where the trouble starts, because you can be easing off the gas and still moving forward at 60 miles an hour. Everything feels fine. The change has already happened, and nothing has visibly gone wrong yet.

That's the gap. That's where 2008 lived. And that's where the nuclear industry is sitting right now.

What Just Happened in Washington

Some background on who's in charge of the electricity grid. The US grid is run in big regional chunks by organizations with clunky names. PJM covers the mid-Atlantic and Midwest, ISO New England covers New England, CAISO covers California, and so on. They're overseen by a federal agency called FERC.

For about a century, the deal has worked like this. If you want electricity, you connect, and the system is responsible for making sure enough power plants exist to serve you. The cost of building those plants gets spread across everybody's bill.

Now AI data centers have shown up. They're enormous. A single one can use as much power as a small city, and they want to connect within a few years. The system cannot keep up.

So on June 18 2026, FERC told all six grid operators that their current rules for connecting these giant customers may be unfair and unreasonable, and that they need to fix them.

Two details matter more than the headline.

FERC had been about to write one national rule. It had collected 3,500 pages of feedback from 175 organizations. Then, right before the meeting, it pulled that plan and opened six separate regional cases instead. Six regions, six rulebooks, potentially six different answers.

FERC asked a direct question. How, exactly, are you going to make sure there's enough power for these new customers?

The answers came back on July 20. They were mostly lists of meetings, working groups and future deadlines. Translated, that means nobody has an answer.

But a direction is emerging anyway, and it has a name, bring your own generation, or BYOG.

Under the old deal, you connect and the grid finds the power. Under BYOG, you connect only if you bring your own power.

ISO New England told FERC flatly that it plans to stop buying power on behalf of these giant customers. PJM is looking for BYOG to solve the problem, despite recent auctions where generation didn't show up.

If you're funding AI data centers, electricity has just stopped being something you buy. It's become something you build.

BYOG Is Not the Loophole People Think It Is

A lot of the industry heard "bring your own generation" and got excited. Build your own power plant next door, plug straight into it, skip the queue, skip the grid fees. Freedom. That's not what the June orders say.

FERC specifically went after the trick where a data center subtracts its on-site power from its total usage to shrink its grid bill. It called that unfair.

It went further. Even if you're running on your own plant and drawing nothing from the grid, you should still pay for certain grid services because you're relying on the grid being there. If your own plant hiccups, the grid catches you. That's a service, and FERC says you pay for it based on your full power appetite, not on what you actually pulled.

Meaning, BYOG isn't an escape hatch. It's an extra bill. You pay to build your own power plant, and you keep paying for the grid.

There's more. FERC has already approved a handful of individual deals between utilities and data center developers. One was with Amazon in late 2025, and five more followed in Illinois in early 2026. Those deals come with real teeth. Deposits, credit requirements, penalties if you don't show up, exit fees. In some cases the security posted has been in the range of ten years' worth of expected electricity costs.

Put it all together and this is what it now takes to switch on a large data center.

• Prove you've secured your own reliable power
• Put up collateral worth roughly a decade of energy bills
• Pay grid fees on your full power appetite, whether you use the grid or not
• Do all of it before the data center exists and earns a single dollar

That's not an energy rule anymore. That's a credit check.

What's Actually Being Taken Apart Here

To see what's ending, you have to see what it's replacing, because the American electricity system has been quietly socialist for a hundred years.

Here's the old bargain. A utility gets a legal monopoly over an area. In exchange, it takes on three duties.

• Serve everyone who asks
• Charge fair rates
• Ensure the grid is reliable

In return, it's allowed to earn a steady, regulated profit.

That first duty is genuinely remarkable. Almost no other business is required to accept every customer, run wires out to reach them, and build extra capacity for demand it never asked for and can't turn away. That obligation is what makes electricity a shared public thing rather than just a product.

And the sharing happens in how bills are calculated. Most of the cost of electricity is fixed. Poles, wires, substations, power plants. It costs roughly the same whether you use a lot or a little. But you're billed by the unit you consume, at the same rate as everyone else in your customer class. So a farmhouse at the end of a ten-mile line and an apartment in a dense city pay the same rate, even though the farmhouse costs far more to serve.

That's a subsidy, and it wasn't an accident. It's how a country this big got electricity to everybody.

Now here's the part people miss: a promise to serve everyone, at every hour, only works if you have power plants that can actually deliver at 3am in February. Not when it's windy. Not when it's sunny. Always.

For most of the last century that meant coal, oil, gas, hydro and nuclear. The promise to serve everyone was only credible because those plants existed. They were the collateral behind the promise.

The Places That Shared the Cost Removed the Collateral

Now let's look at the highly worrisome pattern. The places most committed to treating electricity as a shared public good, where the cost is spread and everybody chips in, are in many cases the same places that shut down the one technology that made a shared, always-on promise physically possible. They kept the sharing without fully understanding the underlying physics and engineering that created the foundation for reliability.

First, New England. Vermont spent years working to close the Vermont Yankee nuclear plant; its Senate voted 26 to 4 in 2010 to block the license. The plant closed in December 2014. It had been supplying about 5% of the region's electricity and over 70% of Vermont's own power. Along with the loss of that plant came the loss of an entire community.

The Pilgrim plant followed in 2019. Add in closed coal and oil units, and the region lost roughly 4,600 megawatts, more than a tenth of its total capacity.

Natural gas filled the hole. It went from about 15% of New England's power in 2001 to around half by the mid-2010s. Emissions went up after Vermont Yankee closed. On the coldest winter days, when the gas pipelines are full, the region burns fuel oil to keep the lights on.

Germany ran the same experiment nationally. It agreed to phase out nuclear in the early 2000s and shut its last three reactors in April 2023, in the middle of an energy crisis, while burning coal and unwinding its dependence on Russian gas.

Germany paid for the transition with a surcharge on electricity bills. And that surcharge is worth pausing on, because it exposes something uncomfortable. A fee charged per unit of electricity hits poorer households harder. Your electricity use doesn't scale with your income. A rich household doesn't use ten times the power of a poor one.

Meanwhile, the way to benefit from the transition was to put solar panels on your roof. Which requires owning a roof, having credit, and having cash.

The costs were shared. The assets were not. Nobody designed it that way. It's just what the arithmetic does. And it's the same arithmetic running through every behind-the-meter installation in America today.

Now let's close the loop. Having removed the reliable plants that made "we'll serve everyone" a real promise, the most cost-sharing grid in America is the first one telling FERC it wants to stop finding power for large customers. New England is leading the charge on bring your own generation. The region that shared the most is the region handing the bill back first.

One clarification, because it matters. The problem isn't that left-leaning politics is anti-nuclear. France disproves that completely. It built 56 reactors under the most state-directed energy system in the West and got the cleanest large grid in Europe out of it.

The point is simpler. If you believe electricity is a public good, you need something that can actually deliver at 3am in February. Take that away and you don't get a greener version of the same social contract. You get gas, then you get price spikes, then you get a regulator quietly writing rules that hand the duty to serve back to whoever can afford to pay for it privately.

That's not a prediction. That's the July 20 filings.

Everything from here depends on a question this piece should state plainly rather than assume, because the whole argument turns on it. Are these new loads real, durable, and important enough to justify rewriting a century-old bargain?

Reasonable people disagree, and the honest answer is that nobody knows yet. Both cases are worth setting out.

The case that the demand is real and urgent

The loads are computing facilities, mostly for artificial intelligence, and the capability of an AI system currently scales with how much computing power is thrown at it. That makes it a race with a short clock, because whoever has capacity running soonest trains the better model and funds the next round.

Enormous sums have already been committed on that logic, arguably the largest private construction program in modern American history, and much of it was announced before the power to run it existed. Computing is also portable in a way that a steel mill is not. If a developer cannot get power in one state by 2028, it can build in another state or another country, so the facilities get built either way and the only question is who collects the investment, the tax base and the jobs. On this reading, a state that cannot deliver power is not protecting its ratepayers. It is exporting an industry.

The case for skepticism is equally serious

Utility load forecasts have a long history of being wrong in the same direction, and the interconnection queues these projections rest on are widely understood to contain duplicate and speculative applications, because a developer loses very little by filing in several places at once and waiting to see which one clears first.

That inflates the headline number. Efficiency has repeatedly undercut this kind of forecast before, and there is no strong reason to assume the computing cost of a given capability stays fixed. The economic benefits are also contested. Data centers consume a great deal of power and land while employing relatively few people once built, and they are often granted substantial tax concessions to arrive at all. And the race framing is exactly the sort of argument that gets made when someone wants a regulator to move quickly.

There is a third possibility that fits the evidence better than either. The aggregate demand may be genuine while a large share of the individual projects are not, in which case the grid is being asked to plan around a number that is directionally right and specifically wrong.

This piece does not need to resolve that, and readers should be suspicious of anyone who claims to have. The argument that follows works either way, and here is why. If the demand is real, the rules being written now determine whether America can serve it, and whether it builds the domestic capability to keep serving it after 2030. If the demand is inflated, the rules determine how much stranded infrastructure gets built at everybody else's expense before that becomes obvious.

Either way, the uncertainty itself is the problem. Nobody can forecast this demand reliably, and as the next section shows, one industry has quietly built its entire cost structure on the assumption that somebody can.

Now to the part of the industry that's most exposed.

Small modular reactors (SMRs) are the hot idea in nuclear. Instead of one enormous custom-built plant, you build small standardized reactors in a factory and ship them out.

Here's the problem with the economics.

The first SMR costs somewhere between $80 and $150 per megawatt-hour of electricity. For comparison, solar is $30 to $50, wind $25 to $50, gas $40 to $75. The first SMR loses to everything.

The attractive number, $50 to $80, only shows up at reactor number 20 or so. And it only gets there through one mechanism, which is building the same thing over and over until you get good at it. Factory production. Repeat suppliers. Crews who've done it before. Regulators who've already approved the design. Every bit of that requires volume.

So the value of an SMR company doesn't rest on how many orders it has. It rests on orders arriving fast enough, for long enough, to justify a factory and get down the cost curve before the first reactors are even finished. Which means that if orders slow down, the business doesn't shrink a bit. It breaks. Every reactor stays stuck at first-of-a-kind cost forever, a price at which nobody wants to buy it. For a normal company, slower growth means a lower valuation on the same economics. Here, slower growth means the good economics never arrive at all.

And it's a trap. Nobody can justify buying reactor number one at reactor-one prices. But reactor-twenty prices only exist if somebody buys reactor number one.

We've already watched this play out. NuScale had the only SMR design the US regulator had ever certified, with $1.4 billion of federal support behind it. As the project developed, the expected price of its power rose from $58 to $89 per megawatt-hour, and the project cost went from $5.3 billion to $9.2 billion. Twenty-three of its thirty-five utility customers walked away.

The not-so-subtle irony is that this cost is highly competitive in today's market. The head of Idaho National Laboratory told Congress the project didn't fail on technical grounds. It failed on cost. Nobody stopped wanting clean, reliable power in 2023. Customers just stopped signing up fast enough for the cost curve that needed them to keep signing up faster. The level of demand was fine. The rate was not.

The One Thing BYOG Genuinely Fixes

Now the honest case in favor, because there is one. About thirteen deals have been announced between big tech companies and nuclear developers, covering roughly ten gigawatts. Meta has signed up for as much as 6.6 gigawatts across four partners. Microsoft signed a twenty-year deal worth around $16 billion to restart Three Mile Island. Google committed to up to 500 megawatts. Amazon put $700 million into X-energy. Oklo and Switch signed something covering twelve gigawatts.

Now sort that pile by how binding it actually is. A twenty-year contract to buy power from a licensed plant that's being restarted is a real contract. Money will change hands. An "agreement contemplating twelve gigawatts" of a reactor design that has never produced a single watt of commercial electricity is not a contract. It's a free option on a product that doesn't exist yet.

In between sit letters of intent, memoranda, frameworks and partnerships, all announced in gigawatts, as if a gigawatt on paper were the same as a gigawatt under contract.

Under the old system, that softness made sense. A tech company could sign a 2035 nuclear framework as a hedge, a headline, and a cheap option, because ultimately the grid was on the hook for keeping the lights on. It was the easiest line in the budget to quietly abandon. No announcement, no write-down. It just stops getting mentioned.

BYOG changes that, and this is the strongest argument for it. If proving you have reliable power becomes a condition of getting connected, then buying power stops being optional and becomes mandatory. It moves into the same column as land, transformers and water. These are the things you cannot skip.

That's exactly the hardening the nuclear order book has always needed, and exactly what private markets could never provide on their own. If BYOG turns soft frameworks into binding, bankable, long-term contracts, it does more for nuclear than any subsidy currently on the books. That's the case. But here is why it most likely won't work out that way.

Six Regions, Six Rulebooks

Remember that FERC scrapped the single national rule and opened six regional cases. For most purposes, reasonable people can argue about whether that's good federalism. For nuclear specifically, it's close to fatal, and almost nobody in the industry is saying so.

The entire cost case for SMRs rests on repetition. The reason France built 56 reactors at manageable cost from a handful of standard designs, while America built a nearly one-off fleet, isn't physics. It's repetition. And repetition is a regulatory question before it's an engineering one.

Now think about what six independently written rulebooks produce.

• Six definitions of what counts as "bringing your own generation"
• Six ways of measuring how much reliable power a given reactor counts for
• Six sets of connection standards
• Six collateral formulas
• Six sets of grid fees

A developer can standardize the reactor completely and still face a different fight in every region it sells into. That alone can break the cost curve, because the learning curve isn't only in the steel. It's in the license application, the financing template, the power contract, the grid study, the insurance. Those are exactly the parts that get rebuilt from scratch when the rules differ region to region.

You can perfect the hardware and still pay first-of-a-kind paperwork costs forever. So the procedural fight matters far more than the industry seems to realize. If the six regions each write their own fix, nuclear gets six markets. If FERC writes one, nuclear gets something close to one market. For an industry whose economics depend on doing the same thing repeatedly, one rulebook is worth more than most of the subsidies currently on offer.

Before It's a Nuclear Mandate, It's a Gas Mandate

The second problem is timing, and it's simple arithmetic. BYOG requires you to have reliable power when you connect. The data centers in the queue now are targeting 2027 to 2030. The earliest even semi-credible date for a commercial SMR is 2030, but given the reality of what is required to build a grid-scale nuclear power plant, this date is becoming less realistic. The first nuclear power actually dedicated to an AI data center will come from restarting an old plant, not building a new one.

This is one of the central problems of our time and it is a bipartisan issue. The U.S. shut down multiple perfectly operating nuclear reactors under both Democrats and Republicans.

But more importantly, nuclear physically cannot serve the window BYOG creates.

What can? Gas. Massive volumes of batteries, which don't yet exist at anything close to the scale needed. And paying customers to use less power at peak times, which brings us back to reliability.

So a rule that looks like the biggest gift nuclear has ever received functions, in the near term, as an order to go buy gas turbines. And it lands in the tightest turbine market in living memory.

Three companies make most of the world's large gas turbines. One of them saw its order backlog jump from 62 to 83 gigawatts in a single quarter at the end of 2025, and is targeting at least 110 gigawatts by the end of 2026. Another is carrying its largest backlog ever. Turbine prices are reportedly up roughly threefold in three years. The order book is several times what the factories can produce annually, and expanding a turbine factory takes three to five years.

BYOG joins that queue. It doesn't shorten it. And there's a knock-on effect. Money spent buying turbine slots, posting collateral and building expensive gas plants in 2027 is money not available to order reactors in 2029. The urgent obligation crowds out the important one. Which is precisely backwards from what nuclear needs, which is orders placed now for delivery later.

There's also a much worse version of this. If everyone meets their BYOG obligation with gas, then by the time SMRs are ready, the reliable power has already been built. The shortage that justified paying a premium for nuclear is gone. And the reactor arrives into a market that no longer needs it at that price. Things that take a decade to build always arrive after the moment has passed. BYOG speeds up the arrival of the substitute.

BYOG Is a Credit Filter

The third problem is who can actually comply. Add it up. Secure your own power plants, post ten years of collateral, pay grid fees on your full appetite, guarantee minimum revenue. All before you earn a dollar. That's a balance sheet test, and maybe four or five companies on earth pass it comfortably.

The smaller AI cloud providers can't. Their whole business is the spread between money they borrowed and computing they rent out. There's no balance sheet to absorb this. The AI labs can't either. Their obligations are paid out of the next funding round, not out of profits. Independent developers will need a big tech company to stand behind them, which means it's the tech company's credit being tested, not theirs.

So BYOG does to the data center industry what going public does to an unprofitable company. It forces the weakest players to prove, upfront and in cash, that they can fund their own promises.

Many can't.

And here's the consequence people miss. This concentrates the entire buildout onto a handful of balance sheets. The nuclear order book becomes less diversified, not more. In 2008, the hidden thing connecting thousands of unrelated mortgages across the country was one number, national house prices. Here, the hidden connecting number is AI spending. BYOG tightens that connection instead of loosening it. When everything depends on one variable, you don't notice until that variable moves.

Why This Pulls the Brake Forward

Now the uncomfortable bit. Right now, every big tech company spends enormous sums on AI because the market rewards spending and punishes whoever flinches first. Nobody can stop unilaterally. That whole arrangement rests on a shared belief about what the next dollar of spending means. And beliefs change. The turning point comes the first time a big tech company announces a spending cut and its share price goes up instead of down.

Because everyone is watching everyone else, that flip won't be gradual. Now, BYOG shifts the cost of power onto the data center's own books. That raises the cost of building each unit of computing capacity and lowers the return on it. It doesn't reduce demand for AI at all. It raises the cost of serving that demand, and therefore lowers the payoff.

Which means BYOG pulls forward the day a finance chief gets rewarded for restraint. That's the paradox at the center of this. The reform that makes nuclear contracts more binding also worsens the economics that produce nuclear contracts in the first place. Each contract becomes harder to walk away from, and less likely to be signed at all. Which effect wins is genuinely unclear, and depends on details nobody has written yet.

Four questions, four answers

Will it happen? Yes. No regulator wants to explain to ordinary households why their bills went up to pay for data centers. FERC has already said the current rules are unfair, and the operators' own filings admit they don't have a solution. Some version of BYOG is coming to most regions.

Will it protect ordinary customers? Mostly yes, and this is its real purpose. Whoever causes the cost should pay the cost. That's coherent and probably correct.

Will it solve the actual power shortage? No. And this is the crucial distinction.

BYOG changes who is responsible for finding power. It does nothing about whether power plants actually get financed and built. Those are different problems. Speeding up the connection process helps if paperwork is the bottleneck. It does nothing if the bottleneck is that plants aren't being financed, or that you can't get a turbine until 2031. Moving an obligation onto someone who also can't get a turbine doesn't produce a turbine. It produces a data center that doesn't get built. That's technically a solution to the power shortage, but not the one anyone is claiming.

Will it save nuclear? Not as currently designed, and most possibly the opposite.

It hardens contracts, which is the single best thing that could happen to SMR economics. But it does so on a timeline nuclear can't meet, in six different regulatory flavors that undermine the repetition the cost curve depends on, while burning near-term money on gas and pulling forward the day the AI spending boom ends.

BYOG is best understood not as a fix for scarcity but as a way of rationing it. It decides which data centers get built. It doesn't decide how much power exists.

How It Breaks

The sequence, in order. AI spending growth flattens out. It doesn't fall, it just stops accelerating. BYOG raises the cost per unit of computing capacity and squeezes returns. The belief flips, and one big tech company gets rewarded for spending less.

The non-binding nuclear frameworks quietly don't get converted into real contracts, starting with the most distant and least developed designs. Announced pipeline gigawatts, the number every developer's valuation rests on, stop growing.

Orders no longer justify a production line. The factory never gets built. The learning curve never starts. Every reactor stays stuck at first-of-a-kind cost. Developer shares, which were never valued on actual revenue, get repriced from "exciting option" to "what are the assets worth."

Turbine factories finish expanding into a softer market. The premium for reliable power collapses. The going rate for firm power falls below what a first-of-a-kind reactor needs to pay its debts.

The projects with binding contracts get finished anyway, at costs their owners can no longer defend. Then well-capitalized utilities, industrial buyers and government programs buy those finished reactors cheaply and run them profitably for sixty years.

Notice where that lands. The reactors mostly get built. They just don't end up belonging to the people who paid for them. That's not a market working efficiently. That's a coordination failure destroying something the country needs, and then rebuilding it a decade later at far greater cost. Which is the argument for stepping in while the rules are still being written, and they're being written right now.

The Problem Isn't Money

There is already a lot of public money going into nuclear. A 2025 executive order set up a pilot program aiming for at least three advanced designs to go live by July 2026, with eleven projects selected. A fuel program followed. The 2026 budget put $1.785 billion into the Office of Nuclear Energy and redirected $3.1 billion toward advanced reactor cost-sharing. The Department of Energy awarded $2.7 billion in January 2026 for domestic uranium enrichment. Loans went out for restarting old plants, up to $1.52 billion for one and $1 billion for another. In June 2026 it committed up to $17.5 billion for ten large reactors.

That's not a funding shortage. That's serious capital, deployed with real urgency. The problem is how it's being spent.

Almost all of it funds demonstrations. But demonstrations aren't the bottleneck. Eleven pilot projects across eleven different designs produces eleven first-of-a-kind reactors and zero learning curves. It pays for the most expensive reactor each of eleven companies will ever build, and then stops precisely where the economics were supposed to start improving.

The cost savings don't live in reactor one. They live in reactors two through twenty.

The $17.5 billion commitment is the exception that proves the point. Ten reactors, up to five sites, one standard design. That's not a demonstration. That's an order book, and it's the right instrument. The government has already figured out how to fund a learning curve for large reactors. It just hasn't applied that logic to the small ones, where the learning curve is the entire investment case.

So here's the reframe. What the state actually needs to supply isn't money. It's reliable long-term demand, and one consistent rulebook. Private investors can't underwrite reactors five through twenty, because that requires a confident view of electricity demand in 2035 that no board will sign off on. Governments can, because they have a longer horizon and an interest in the capability existing regardless of what AI does.

What Solar and Wind Actually Taught Us

Solar panels didn't get cheap because someone invented a better photon. They got cheap because policy guaranteed enough volume, for long enough, that manufacturers had time to get good at it.

Three lessons. The industry keeps citing the least useful one.

De-risk the first few, then step back. In 2011 the government guaranteed loans for the first five solar farms over 100 megawatts ever built in America. Private lenders wouldn't touch them because nobody had ever financed one. Once those five were running, seventeen more got financed entirely privately with no federal help. The government bought a track record, and the track record was the product. One demonstration doesn't create a track record. Five do.

The failures were factory bets; the successes were project bets. Solyndra is the famous political disaster at $535 million, and Abound Solar failed too. Both were manufacturers. Manufacturing was only about 8% of that loan portfolio; actual power projects were about 74%, and the projects worked. By 2016 the program was in the black overall, with interest received exceeding total losses. Back the plant, not the champion.

Wind is the cleanest proof of the whole thesis. The wind tax credit was created in 1992, extended about a dozen times, and allowed to lapse six times. In the years after each lapse, installations fell between 76% and 93%. In 2013, after the credit expired and was restored two days late, the industry installed about 8% of the previous year's total. In 2010, the mere threat of expiry cut installations 43%, without the credit actually lapsing.

Read that against everything above. Demand for wind power did not fall 90% in 2000, or 2002, or 2004, or 2013. What slowed was the policy signal, and the order book collapsed anyway, because the industry's costs depended on orders piling up steadily, not on how much people wanted wind power.

The wind credit is this whole argument, proven four separate times, at public expense. Which gives us the design rule. Reliability beats generosity. A big subsidy with a two-year horizon is worth less than a modest one with a twenty-year horizon, because building a reactor takes six to ten years and the learning curve needs maybe twenty units. Support that expires before your second reactor is even licensed is a rounding error with a press release attached.

Six Things That Would Actually Help

Roughly in order of impact. The first is urgent because of the August calendar.

• Make BYOG mean the same thing everywhere. This is the highest-leverage move available in 2026, and it costs nothing. The six regions should share one definition of what counts as qualifying reliable power, one method for measuring how much a given reactor counts for, and one process under which a certified standard design is pre-approved everywhere. Flexibility on paperwork is fine. Flexibility on what counts as reliable power is what turns a standardized reactor back into a one-off project. If the industry files nothing else, it should file this.
• Concentrate the nuclear money on a few designs. Replace scattered demonstration grants with standing multi-reactor orders on two or three chosen designs, covering roughly units one through twelve at published prices, tied to performance rather than to the annual budget fight. Eleven pilots across eleven designs guarantees no learning curve anywhere.
• Write the cost curve into the contract. Agree a declining price schedule upfront, with a first-of-a-kind price for unit one and a near-mature price by unit twelve, fixed and non-negotiable. Beat the curve, the developer keeps the difference. Miss it, the developer eats it. Public support then shrinks automatically as volume grows, with a built-in end date that requires no political decision.
• Put the duration in law, not in an executive order. Twenty years minimum, with the wind-down written on day one. Right now the support rests on executive orders and annual budgets. An executive order can be undone by the next one. That's not a horizon anyone builds a nuclear facility against.
• Fund the shared bottlenecks nobody can justify alone. Uranium enrichment, heavy forging capacity, qualified welders and inspectors, and regulator throughput. No single developer's pipeline justifies building a forge, so no forge gets built, and everyone pays inflated component prices forever. The $2.7 billion enrichment award is the right instinct; extend it across the supply chain. On licensing, the best-value item is standardized design approval with capped fees and legal deadlines.
• Have the federal government be the anchor customer, especially when the market turns. Federal buildings, energy department sites and military bases represent large, permanent, price-insensitive demand for reliable local power that has nothing to do with AI. Turn that into a standing government order that rises when private orders fall. This is the one that speaks directly to the core argument. It doesn't just subsidize the order book, it removes the uncertainty that makes the order book unfinanceable. Put a floor under the order rate, and the learning curve becomes a schedule instead of a hope.

Notice what all of these have in common. None of them require anyone to correctly predict AI demand. That's the whole point. The industry's fragility comes from building its cost structure on a demand forecast nobody can make. The answer isn't a better forecast. It's a floor and a single rulebook.

What Survives

Restarts and upgrades of existing plants are the most durable category, and that's exactly where utilities are putting money today. Existing sites, existing licenses, known costs, near-term delivery, and customers who need power now rather than in 2035. There's no learning curve to climb because there are no first-of-a-kind costs. This is where the first real electrons come from, and it's presumably why federal loans went there first.

Next, designs with a customer that isn't AI. A government program, a regulated utility, an industrial heat customer, a defense requirement. Anything whose order book doesn't depend on the same single variable as everything else.

Next, designs with a genuine industrial parent big enough to absorb first-of-a-kind costs itself, and a site that's already licensed.

What doesn't survive is the enormous crowd in the middle. There are roughly 130 SMR designs in development, depending on which database you trust. The world can support maybe a handful of production lines, and fewer in the West.

The number of designs isn't a measure of innovation, any more than record spending is a measure of demand. It's a measure of how many people wrote a business plan. Each design is a free option on an order book that mostly won't exist, and options expire.

Consolidation is coming either way. Policy decides only whether it happens as an orderly concentration onto a few designs that then get cheap, or as a disorderly collapse that leaves an entire generation of reactors stranded at first-of-a-kind cost and sets the sector back another decade.

The Number Nobody Watches

The market is treating new nuclear as a technology story. It isn't. It's a credit and property story, built on hard assets, heavy borrowing, long-term contracts that are leases in all but name, and construction times that guarantee supply shows up late.

Everyone watches the level, which is announced gigawatts. Occasionally someone watches the speed, which is deals per quarter. But this structure breaks on the acceleration, because the cost of the product depends on how fast orders pile up. And the market is treating a handful of buyers as if they were a broad, diversified energy market, at the exact moment a regulation is about to concentrate them further.

The law hasn't changed. If something can only stay afloat by borrowing against growth, it doesn't need growth to stop. It just needs growth to slow down. Here's what makes this different from the rest of the AI trade. For a chip lease or a cloud arbitrage, that slowdown is weather. Nobody can do anything about it. For a first-of-a-kind reactor, that slowdown is a policy choice.

Solar and wind came down their cost curves because policy guaranteed volume long enough for manufacturers to get good. Where that guarantee wobbled, which happened six times in the case of wind, installations fell by up to 93% in a single year while demand for the underlying product never moved at all.

Bring your own generation is about to make the same kind of choice, on a compressed schedule, in six places at once. It will decide whether buying power becomes a binding, standardized, bankable obligation you can build a factory against, or a fragmented compliance headache that hardens contracts nobody can afford to sign.

Two dates to watch: August 3, for who asks for more time. August 17, for whether the fix gets written by the six regions or by Washington. Where you build is now a regulatory question as much as a real estate one. So is the cost curve.

The market remembers 2008 backwards. The defaults didn't start when prices fell. They started when prices stopped rising faster. The nuclear order book is written on that same number, and it's the one number the government can actually set.