The utility filed the numbers in the spring. On paper, the average residential bill in the territory would climb 9.4% in a single rate cycle. The reason code was not "fuel." It was not "storm recovery." It was not "wildfire mitigation." It was four words buried on page 47 of a 214-page filing: "load growth, large computing customers."
The load did not show up in the voter's feed. The bill did. That is the whole story. Everything else — the AI capex narrative, the hyperscaler earnings beat, the data center REIT pitch deck — is downstream of a meter reading that a household in Ohio, Virginia, or Georgia opened on a Tuesday and did not understand.
The code spoke, but the metadata lied. The metadata said "digital economy." The meter said "your neighbor's grocery budget."
I have spent fifteen years tracing systems that promise one thing on the surface and do another underneath. I have audited token contracts where the whitepaper said "decentralized" and the admin key said "trust me." I have watched an algorithmic stablecoin unwind in 72 hours because the stake weights were concentrated in fewer hands than the marketing admitted. The data center problem is the same anatomy wearing a different skin. A promise at the top. A cost at the bottom. A gap in the middle where accountability is supposed to live — and doesn't.
This is not a piece about whether AI is real. It is. This is not a piece about whether data centers are useful. They are. This is a forensic accounting of who pays for the electricity that makes them run, and why the bill has just become a midterm liability that nobody on the buy side has properly marked to market.
Let me show you the receipts.
Context: The Load That Ate the Forecast
For two decades, the American electric grid was a flatline business. Demand grew about 1% a year, sometimes less. Utilities built rate cases around replacement, not expansion. Load forecasting was a sleepy actuarial exercise. Then, in roughly eighteen months, the flatline bent.
The bend has a name in the industry: load growth. It is the single most disruptive variable in utility planning since the shale gas boom. The cause is the collision of three demand curves that all point the same direction.
First, the AI build-out. A single rack of Nvidia's current-generation accelerators draws roughly 120 kilowatts. A decade ago, a dense enterprise rack drew 8 to 12 kilowatts. That is not a step change. That is a different physical regime. A hyperscale training campus can pull 300 megawatts to 1 gigawatt — the output of a mid-sized nuclear reactor — and it wants that power 24 hours a day, seven days a week, at a power quality that tolerates no brownout.
Second, the reshoring of manufacturing. Chip fabs, battery plants, and electrified industrial processes are all enormous, inelastic loads. They do not flex with the weather. They flex with the order book.
Third, the electrification of transport and heating. Slower, but cumulative. Every EV is a small load. A million of them is a utility's problem.
The grid was not built for any of this. It was built for a world where the biggest load on the feeder was a Walmart. Now the biggest load on the feeder is a building that hums like a jet engine and pays for its power with a contract that most ratepayers will never see.
Here is where the crypto reader should lean in. The data center build-out did not appear out of nowhere. It inherited the physical, political, and financial infrastructure that the Bitcoin mining industry built between 2017 and 2024. The same interconnection queues. The same substation engineers. The same municipal tax-abatement playbook. The same willingness to sign 15-year power purchase agreements at prices that look insane until the price of the alternative moves.
When the mining boom cooled, a lot of that capacity did not disappear. It converted. Companies that had spent years building the muscle to chase cheap power — and to negotiate with utilities as a large, sophisticated, mobile load — turned that muscle toward AI hosting. Core Scientific, Riot, TeraWulf, IREN, Hut 8. The tickers changed their story from "we mine Bitcoin" to "we host GPUs." The physical plant did not change much at all. A container full of ASICs is not that different from a container full of accelerators. Both need a substation, a cooling loop, and a utility that will pick up the phone.
That conversion is the hidden spine of this story. It is also why the Crypto Briefing coverage matters. The publication that covers miners now covers the same buildings that voters are angry about. The industry that spent a decade defending its energy use against critics is now watching a much larger, much better-funded neighbor absorb the same criticism — and discovering that the criticism does not care who owns the building.
Now layer on the politics. The United States enters a midterm cycle with electricity prices as a live kitchen-table issue for the first time in a generation. Inflation cooled from its 2022 peak, but the last mile of disinflation is the stickiest, and electricity is one of the stickiest components. A voter who has stopped noticing the price of eggs has not stopped noticing the price of power, because the power bill is monthly, visible, and rising for a reason that feels abstract and unfair.
The reason feels abstract because the benefit is abstract: "AI leadership." The cost is concrete: dollars per kilowatt-hour. When a benefit is abstract and a cost is concrete, politics resolves the asymmetry in one direction. The abstract benefit loses.
Core: A Forensic Teardown of the Cost-Allocation Machine
Let me open the machine and look at the gears. There are five, and each one moves money from a household to a balance sheet. None of them are hidden by malice. All of them are hidden by structure.
Gear One: The Interconnection Queue Is a Queue to Nowhere
Start at the physical bottleneck. To build a data center, you need generation, transmission, and a grid operator willing to connect you. The connective tissue is the interconnection queue.
As of the most recent national accounting, more than 2,600 gigawatts of generation and storage capacity sits in interconnection queues across the United States — waiting for studies, upgrades, and approval. That is roughly twice the installed capacity of the entire American grid. The median wait time has stretched past five years in the largest markets. Some projects have been in queue since 2016.
This is not a data center problem. It is a grid governance problem that predates the AI boom. But the AI boom poured gasoline on it. A data center developer that needs 500 megawatts of firm capacity does not just wait in the queue. It leaps the queue by paying for its own upgrades, signing a dedicated generation contract, or building behind the meter.
Each of those moves has a cost-allocation consequence. When a large load "brings its own power," it avoids paying into the shared network upgrade pool that traditional loads fund through retail rates. When it instead connects to the shared network, it triggers upgrades whose costs are socialized across all ratepayers — including the household that will never use a single GPU-hour.
The queue is where the fiction of "the market will handle it" dies. The market cannot handle a five-year lead time on a 24/7 load that the customer needs next quarter. So the load finds a shortcut. The shortcut is not free. Someone pays. The question is whether that someone can afford it.
Gear Two: The PPA Is a Hedge That Becomes a Subsidy
Power purchase agreements are the industry's favorite word. They sound like private contracts. In practice, they are the mechanism by which the cost of a data center's electricity gets split between the customer and everyone else.
Here is the structure. A hyperscaler signs a long-term PPA with a generator — often a nuclear plant, a solar farm, or a wind project. The PPA price is fixed, the term is long (10 to 20 years), and the headline is clean: "100% renewable," "carbon-free power," "green by 2030."
Now the forensic part. A PPA for a specific generator does not necessarily deliver electricity to the data center at the time the data center uses it. The electrons flow on the shared grid. What the PPA actually does is transfer the financial attributes of a generator — the environmental credits, the fixed price — to the buyer. The physical power still comes from whatever is on the grid at 3 a.m. on a still, dark night.
This matters because the grid has to be built to serve the physical load, not the paper contract. When the sun does not shine and the wind does not blow, the data center still draws. The system has to have capacity ready for that moment. That capacity — peakers, storage, transmission — costs money. If the data center's paper contract does not pay for it, the cost lands in the capacity market, and the capacity market clears into retail rates.
I have seen this exact pattern in tokenized infrastructure. An asset claims to be "backed" by something off-chain. The on-chain representation looks fully collateralized. Then you reconcile the on-chain claim against the off-chain reality and find a gap the size of a warehouse. The data center PPA is the same reconciliation problem. The contract says "clean and firm." The physics says "sometimes neither." The gap is not a fraud. It is a structural mismatch that the accounting does not surface.
Gear Three: The Capacity Market Is Where the Bill Actually Lands
If you want to find the true cost of AI load growth, stop looking at the retail rate. Look at the capacity auction.
The capacity market is how a grid operator buys the guarantee that power will be available at peak. Prices in these auctions have historically been boring — a few dollars per megawatt-day. Then the load forecasts changed.
In one recent cycle, the capacity price in the largest American market cleared at roughly nine times the prior year's level. The following auction cleared at the administrative cap, which the regulator had already raised. These are not rounding errors. They are regime changes. And capacity costs flow directly into the retail rate that a household pays.
Here is the part that should make a forensic analyst's hair stand up. The load forecast that drove those auctions included data center demand that was, in many cases, speculative — projects announced but not built, pipeline that may or may not materialize. A utility that over-forecasts load buys too much capacity. Ratepayers pay for it. A utility that under-forecasts load risks blackouts. Either way, the household absorbs the uncertainty, and the hyperscaler signs a fixed contract.
I have audited systems where the oracle fed the contract a number that no one independently verified, and the contract settled on the bad number anyway. Capacity markets have the same oracle problem. The forecast is the oracle. If the forecast is padded, the settlement is padded. Nobody goes to jail. The bill just goes up.
Gear Four: The Tariff Is a Clawback That Arrives Too Late
Regulators are not asleep. In several states, commissions have approved special tariff classes for very large loads. The logic is sound: make the new load pay for the system upgrades it causes, so existing ratepayers are held harmless.
The best-known example requires a large computing customer to commit to a minimum share of its projected demand for a decade or more, with penalties for underuse. It is a clawback. It is the regulator saying: you do not get to build the load, soak up the capacity, and leave when the model changes.
It is also, structurally, a confession. The existence of the tariff proves that without it, the cost would have been socialized. The tariff is a patch on a leak, not a redesign of the pipe. And patches have boundaries. The tariff applies to new loads above a threshold. It does not apply to the loads already connected. It does not apply to the transmission upgrades already triggered. It does not apply to the capacity already procured on a forecast that included the load before the tariff existed.
So the household pays for the past, and the hyperscaler negotiates the future. The clawback catches the next project. It does not claw back the last one.
Gear Five: The Tax Abatement Is the Invisible Transfer
Now the deepest gear, and the one least visible to the voter.
Data centers are aggressively courted by local governments. The courtship usually takes the form of property tax abatements, sales tax exemptions on equipment, and infrastructure commitments — roads, water, substations — funded by the public. The pitch is standard: jobs, investment, tax base.
The forensic reality is more textured. A data center's permanent employment is small. A large campus might employ 50 to 200 people once operational, most of them facilities and security roles. The construction jobs are real but temporary. The equipment is largely imported and often tax-exempt. The property tax abatement removes the single largest recurring revenue stream a locality would otherwise collect.
What the locality gets in return is a fraction of the promised tax base, a modest permanent payroll, and a very large electricity load that the local grid must serve. If the load drives up system costs and those costs are socialized, the locality has effectively subsidized a global technology company with its residents' utility bills. That is a transfer. It is not on any ballot. It is on every meter.
This is the crypto parallel that the industry keeps trying to bury. The Bitcoin mining boom promised the same package to rural counties: jobs, tax revenue, revitalized infrastructure. Some of those promises held. Many did not. When the hash price collapsed, the miners left, and the counties were left with the buildings and the bad taste. The data center pitch is the mining pitch at a larger scale, with a better credit rating and a longer runway. The structural question — who captures the upside and who absorbs the downside — is identical.
The DePIN Illusion and the Tokenized Data Center
Let me address the crypto-native angle directly, because this is where a lot of retail money is currently flowing and where the forensic gap is widest.
There is a growing narrative that decentralized physical infrastructure networks — DePIN — will solve the compute and energy bottleneck by coordinating distributed resources on-chain. The pitch: instead of one gigawatt campus, thousands of small nodes, incentivized by tokens, sharing compute and power.
I have audited enough of these to say the following without flinching. The coordination layer can be decentralized. The physics cannot. A token does not move an electron. A smart contract does not cool a GPU. The hard constraints — interconnection, transmission capacity, cooling, latency — do not care about your governance token. They care about copper, water, and megawatts.
That does not make DePIN worthless. It makes it mispriced. The market is paying for the narrative of decentralization while the actual delivery depends on the same centralized infrastructure everyone else uses. The token is the metadata. The megawatt is the reality. And when you reconcile the two, the metadata is doing a lot of work.
The same logic applies to tokenized data center assets — RWA plays that put a data center's cash flows on-chain as a yield-bearing instrument. I have watched this movie. In 2021 it was tokenized real estate. In 2022 it was tokenized treasuries. In 2026 it is tokenized compute. The pattern is constant: the yield is real until the underlying contract renegotiates, and then the on-chain holder discovers that their "passive income" was a residual claim on a counterparty they never met.
Garbage in, permanence out. Put a speculative cash flow on an immutable ledger and you have not made it safer. You have made it permanently wrong.
The Mining-to-AI Conversion: A Balance Sheet Autopsy
The most under-covered financial story in this cycle is the conversion of Bitcoin miners into AI hosting companies. Let me do the autopsy.
A Bitcoin miner owns three things that an AI host needs: a large power interconnection, a building shell designed for high-density compute, and a workforce that understands 24/7 operations under thermal stress. What a miner does not own is a GPU supply chain, a hyperscaler relationship, or the balance sheet to buy accelerators at scale.
So the conversion takes the form of hosting deals. The miner provides the shell and the power. The AI company provides the chips and the customer. The miner collects a hosting fee, which is more stable than the mining margin and, crucially, easier to finance.
The forensic question is whether the hosting fee actually covers the true cost of the power, or whether the miner is underwriting the deal with the same optimism that killed the mining model in the last down cycle. Here is the trap. The miner signed its utility contract when it was mining. Mining is interruptible. Mining can curtail. AI hosting is not interruptible. A training run cannot pause because the spot price spiked.
So the miner converts from a flexible load to a firm load — and the utility reprices accordingly. The hosting contract may or may not capture that repricing. If it does not, the miner is eating a cost it did not model, and the next earnings call will reveal it. If it does, the AI customer is paying a premium it may not have disclosed to its own investors.
Either way, someone is holding a mismatch. That is the definition of a hidden liability. And hidden liabilities are what I get paid to find.
The Nuclear PPA and the Co-Location Fight
No discussion of this topic is complete without the nuclear angle, because it is where the cost-allocation war has become explicit.
The cleanest-sounding solution to the AI power problem is nuclear. It is carbon-free, it is firm, and it runs 24/7. Hyperscalers have signed long-term agreements to buy power from existing and restarted nuclear plants. On the surface, this is the ideal outcome: AI pays for clean firm power, the grid gets a new revenue stream, and emissions fall.
The forensic problem is co-location. If a data center sits physically behind the meter at a nuclear plant, it takes that plant's output directly. The grid loses that generation. The capacity that the grid was counting on to serve everyone else is now dedicated to one customer. The transmission system still has to be built for the peak. The cost of that build does not disappear because the electrons went behind a fence.
This is why the co-location debate has gone to the federal regulator. The question is not whether the deal is legal. The question is who pays for the grid that the deal bypasses. If the answer is "everyone else," then the cleanest-sounding solution is also a cost shift. The nuclear PPA is a beautiful contract wrapped around an accounting problem.
The Transformer Bottleneck Nobody Priced
There is a physical constraint that makes all of the above worse, and it is not a chip. It is a transformer.
A large power transformer is a bespoke, hand-built machine. Lead times have stretched from a year to three or four years. The supply chain is concentrated. There is no spot market that clears the shortage. If you cannot get a transformer, you cannot energize a data center, no matter how many GPUs you have bought.
This is the infrastructure fragility that the AI narrative systematically underweights. The bottleneck is not compute. The bottleneck is the copper and steel that moves the compute's fuel. And unlike a GPU, you cannot fab a transformer in a quarter. You cannot spin one up in a cloud region. You need a factory, a skilled workforce, and years.
The forensic implication is brutal. Every data center project currently in the pipeline assumes a grid connection that depends on equipment that may not exist on schedule. The pipeline is therefore overstated. And if the pipeline is overstated, the load forecast is overstated, and the capacity procured against that forecast is overstated. The ratepayer is paying for a queue of projects that will not all get built.
That is the deepest version of the problem. The bill is not just paying for AI. The bill is paying for the possibility of AI — including the parts of it that never happen.
Contrarian: What the Bulls Actually Got Right
I have spent several thousand words dismantling the cost narrative. Now let me do the harder thing and steelman the other side, because a one-sided teardown is just a different kind of marketing.
The bulls are right about three things, and each one is load-bearing.
First, load growth is not automatically a bad thing for ratepayers. A grid that grows can spread fixed costs over more kilowatt-hours, which can lower unit costs over time. Utilities that were flat for two decades were also under-investing in resilience. A growth cycle forces capital into a system that badly needed it. If the data center load is priced correctly and the upgrades are shared fairly, the household can end up with a more reliable grid at a similar or lower long-run cost. The problem is not growth. The problem is mispriced growth.
Second, the technology is genuinely productive. This is not a token with no use case. AI is doing real work — in drug discovery, in materials science, in logistics, in code generation. The productivity dividend is real, even if it is unevenly distributed and slower to arrive than the hype implies. I have watched enough crypto cycles to be allergic to narratives without cash flows. AI has cash flows. The bulls are not wrong that something real is being built.
Third, the energy transition and the AI boom can be complementary, not adversarial. A data center that signs a firm nuclear PPA is a data center that funds the continued operation of a zero-carbon baseload plant. That is a genuine emissions benefit. The co-location fight is about who pays for the grid, not about whether the power is clean. If the cost-allocation problem is solved — and it can be — the AI boom becomes a tailwind for decarbonization rather than a drag on it.
Where the bulls go wrong is in the time horizon. They are pricing a decade of productivity gains against a quarter of political pain. The voter does not experience the decade. The voter experiences the quarter. And the quarter is what the midterm decides.
This is the blind spot. The market is long AI on a ten-year view and short electricity on a one-month view. Those two positions cannot both be right in the same portfolio. One of them is a hedge that has not been recognized as a hedge.
I have seen this exact mismatch in DeFi. A protocol prices its yield on a multi-year emissions schedule and its risk on a single block. When the block breaks, the multi-year schedule is irrelevant. The data center trade has the same structure. The multi-year AI thesis is real. The single-meter political shock is also real. The portfolio that holds the thesis without hedging the shock is not diversified. It is exposed.
Takeaway: The Meter Is the Oracle, and the Oracle Is Unreliable
Here is where I land, and I want to be precise, because precision is the only thing that survives a cycle.
The data center electricity problem is not a technology problem. It is a distribution problem wearing a technology costume. The AI trade has been priced as a story about compute. It is actually a story about who pays for the copper, the capacity, and the queue. Those costs are real, they are arriving, and they are landing on the least politically protected balance sheet in the system: the household.
The midterm cycle is the mechanism that converts a distribution problem into a policy problem. Voters who cannot name a hyperscaler can read a power bill. Candidates who cannot explain load growth can promise to make someone else pay. The policy that emerges will not be a ban on data centers. It will be a repricing — special tariffs, capacity charges, efficiency mandates, clawbacks on abatements, and a slow, grinding renegotiation of the social contract between a global technology company and a local utility territory.
For the crypto reader, the lesson is the one the industry keeps refusing to learn. The energy narrative that powered Bitcoin mining was never about energy. It was about cost allocation, and it was always going to be tested by the next, larger load in the queue. That load has arrived. It is bigger, better capitalized, and more politically connected than mining ever was. And it is losing the same argument, in the same counties, for the same reason.
The meter said yes. The voter said no. And in a system where the bill is monthly and the benefit is a decade away, the voter gets the last word.
The only question worth asking now is not whether AI is worth the power. It is whether anyone has actually done the accounting — or whether the whole industry is running on a forecast that nobody verified, a contract that nobody reconciled, and a grid that nobody can build in time.
I have audited that pattern before. It does not end with a correction. It ends with a bill.
Check the diff, not the deck. The diff is where the money went. And the diff, this cycle, is written in kilowatt-hours.
