Knoxville, Tennessee Correspondent

The Tennessee Valley Authority was built in the 1930s to bring power to a region that had none, and ninety years later the largest single category of new load it has been asked to serve is a set of buildings that nobody lives in, built by companies headquartered in other time zones, to run calculations that most of the people paying the electric bill will never use. I work in a county public-works shop in the TVA’s home territory, and I want to begin there because the question asks about power demands and the view from the valley is not an abstraction. The data centers have come. They have come to the counties along the interstate that have flat land and a substation, they have come with nondisclosure agreements and tax abatements negotiated in executive session, and they have come with a load request that in several cases exceeds the peak demand of the city they are sitting next to. The utility has told them yes, because a utility that says no to the largest customer it has ever been offered is a utility whose board does not last. And the utility has told the rest of us, in the rate case, that the generation to serve the new load will be built with gas, on an accelerated schedule, and that the cost will be shared, which is the word a utility uses when it means the residential customer will pay part of a bill that the data center caused. That is the power question as it reaches a county engineer, and it is not whether the sector is sustainable. It is who is paying for the substation.

Now the numbers, because the question deserves them and the valley deserves to know what it is hosting. The five largest technology companies guided about six hundred billion dollars of capital spending this year, three-quarters of it for this, and the fourteen largest operators are tracking closer to seven hundred and fifty, against under four hundred and fifty last year. That is more than the Interstate Highway System cost in today’s dollars, spent in a single year, by five companies, on a product whose revenue model is still being demonstrated. The industry’s own analysts say that a third to a half of the capacity planned for this year will slip to 2028 or later because the power is not there, and the companies have responded by announcing fifty gigawatts of their own gas generation, which is more than the entire generating capacity of this utility, to be built on their own sites so they do not have to wait in an interconnection queue that runs five years. A county engineer looks at that and sees a private utility system being built in parallel to the public one, by companies that can afford to, which will leave the public one with the stranded cost of the upgrades it made to serve them before they decided to serve themselves.

On debt, which the question raises and which the valley understands from its own history. The companies borrowed about a hundred and ninety-four billion dollars by August against a hundred and eight for all of last year, and the credit analysts say capital spending is approaching operating cash flow, which is the point at which a company stops funding growth from profit and starts funding it from lenders. More than a hundred and twenty billion dollars of the data-center debt has been moved into special vehicles financed by private-credit firms, which is the structure a company uses when it does not want the debt on its balance sheet, and which is the structure the valley’s coal companies used when they wanted to build a mine and not own the liability. The Fed raised rates on Tuesday for the first time in three years and signaled another, and the ten-year is near four point eight, which means that every dollar of that debt costs more than it did when the plans were drawn. I have watched projects in this county that were financed at four percent and built at six and never recovered the difference, and this is the largest project in American history being financed at the moment the rate turned.

The thing I want to say about sustainability is the thing a public-works engineer says about every boom, which is that the infrastructure outlasts the reason for it. The TVA’s dams are still there; the aluminum smelters they were built to power are mostly gone. The interstate is still there; the industries it was built to connect have moved. The data centers will still be there in 2040, and the gas plants built to serve them, and the transmission lines, and the question is whether the thing they were built to run will still need them, and nobody knows, because the thing is three years old. The companies say demand is constrained, not inventory, and that the rents on deployed chips have risen this year, and I take that seriously, because a company spending its cash flow and borrowing on top of it is a company that believes its own forecast. But a forecast is not a load. A load is a thing that shows up on the meter, and the valley has been promised loads before, by smelters and by chip plants and by an automaker whose battery factory is a field, and the utility built for the promise, and the ratepayer paid for the field.

On systemic risk, I will say what the structure tells me rather than what the market does. The risk is not that the technology fails. It is that the financing structure has concentrated a trillion dollars of commitments in five companies, a handful of chip suppliers, a set of private-credit vehicles, and a group of leveraged cloud providers whose leases the big companies signed to the tune of a hundred billion dollars in six months, and that all of them are betting on the same forecast at the same time with borrowed money at a rising rate, and that the utilities and the counties and the ratepayers have been enlisted into the bet without being asked. A county that abated the taxes, a utility that built the plant, and a ratepayer who is paying the shared cost are all long the forecast, and none of them chose to be, and if the forecast is wrong they hold the infrastructure and the companies hold the write-off, and the write-off is the easier thing to hold.

So my assessment is that the sector is not heading toward a sustainable phase; it is heading toward the phase every infrastructure boom reaches, in which the building outruns the demand, the financing outruns the cash flow, and the public is left holding the physical plant, and that the specific new risk is the one the valley has seen before and is seeing again: a private industry has persuaded a public utility to build for a forecast, on terms negotiated in executive session, with the cost shared and the benefit not. Whether the technology pays off is above my pay grade. Whether the substation is paid for is not, and the answer, in the rate case, is that it is being paid for by the people who do not use it, which is the oldest arrangement in the valley, and the dams are still here to prove it, and the smelters are not.

Cedar Rapids, Iowa Correspondent

Iowa has more data centers per capita than almost any state, because we have wind, flat land, cheap power, a legislature that will abate anything, and a location that is nowhere near anybody who matters, which turns out to be exactly what a server wants. I kept the books for a plant for thirty years and I keep a side book now on the local bond issues and the tax abatements, and the data-center abatements are the largest entries in it, by a factor that would embarrass the people who voted for them if they read their own packets. So I want to assess the question as a plant accountant assesses any capital program, which is by asking three things: what is the return on the investment, how is it financed, and what happens to the assets if the return does not arrive. The industry has answered the first question with a forecast, the second with borrowing, and the third with silence, and a man who audited a plant’s capital budget for three decades knows that the silence is the answer.

On the return. The five largest companies are spending six hundred billion dollars this year, three-quarters on this, and the fourteen largest operators closer to seven hundred and fifty, against under four hundred and fifty last year, and the analysts have raised their estimates for next year by half in six months. That is a capital program growing at a rate no plant I ever audited could have sustained for two years, and it is being justified by revenue that the companies describe as demand-constrained, meaning they could sell more if they could build more, and I take that at face value for the chips already deployed, because the rents on them have risen. What I do not take at face value is the extrapolation, because a capital budget that doubles every year on the assumption that demand will keep doubling is the capital budget of every boom I have watched, and the plant I kept the books for doubled its capacity in 1998 on the same assumption and ran at sixty percent for a decade. The return on a data center is the rent on the chips, the chips depreciate on a schedule the companies set, and one investor has argued the schedule is too long by enough to overstate combined profits by a hundred and seventy-six billion dollars through 2028. I do not know if he is right. I know that an accountant who gets to choose the depreciation life of his largest asset will choose the life that makes the quarter, and that the chips in a data center built in 2024 are already two generations old.

On financing, which is where the book gets interesting. The companies borrowed about a hundred and ninety-four billion dollars by August against a hundred and eight for all of last year, with AI-related debt near five hundred billion, and the credit analysts say capital spending is approaching operating cash flow, which is the line an accountant watches above all others, because below it a company is growing on its own money and above it the company is growing on the lender’s. More than a hundred and twenty billion dollars has been moved into special-purpose vehicles backed by private-credit firms, which is to say that the debt is off the balance sheet of the company that will use the asset and on the balance sheet of a vehicle whose only asset is the lease, and I have read that structure before, in 2001, when it was called a synthetic lease, and in 2007, when it was called a conduit, and the structure was the same and the lesson was the same, which is that debt off the balance sheet is still debt and it comes home when the lease is not renewed. The Fed raised rates on Tuesday and signaled another, and the ten-year is at four point eight, and every one of those vehicles was priced at a lower rate than the one it will refinance at.

On the assets, which is the Iowa question. A data center is a building with a substation and a cooling system and a set of chips that are worth less every quarter, and if the forecast fails the building is a building and the chips are scrap and the substation belongs to the utility, which built it for the load and will recover the cost from the ratepayers, who in Iowa are farmers and small towns and a few plants like the one I kept the books for. The counties that abated the property taxes for twenty years get the abatement back in the form of an empty building that pays nothing. The industry’s own analysts say a third to a half of the capacity planned for this year will not come online until 2028 for lack of power, and the companies have announced fifty gigawatts of their own gas generation to get around the wait, which in Iowa means gas plants next to wind farms, built by technology companies, to run buildings that the technology companies may not need by the time the plants are finished. I have watched a county build a road to a plant that closed before the road was paved. This is that, at a scale the county cannot imagine, and the county has already voted.

On systemic risk, the auditor’s finding. The risk is concentration: five companies, a few chip suppliers, a handful of private-credit firms, and a set of leveraged cloud providers that the five companies have leased a hundred billion dollars of capacity from in six months, all funded by the same lenders, all depending on the same forecast, all rolling the same debt at the same rising rate. That is not a diversified portfolio. It is one position held by many hands, and when one position is held by many hands the hands all let go at once, and the thing that is left is the physical plant, in Iowa, with a tax abatement and a substation and a cooling tower, and the ratepayers and the counties holding it. The companies will write down the chips. The write-down is a number. The substation is a thing, and it stays.

So my assessment is that the sector is not heading toward a sustainable phase but toward the phase in which the capital budget exceeds the cash flow, the financing migrates off the balance sheet, and the assets outlive the forecast, and that the specific new risk is that the public has been enlisted as a silent partner in the bet, through utilities that built for the load and counties that abated the taxes, without any share of the upside and all of the downside that is bolted to the ground. I vote like someone who has watched a payroll disappear, and I will tell you what the payroll looked like the year before it disappeared: it looked like growth, and the capital budget looked like confidence, and the lenders looked like partners, and the depreciation schedule looked reasonable, and the plant doubled, and then it did not need to be double. Write that in the book. Do not net it against the forecast. The forecast is theirs. The substation is ours.

Bismarck, North Dakota Correspondent

I am going to make the case that the build-out is sustainable, because somebody on this forum should and because North Dakota is one of the places it is being built, and I am going to make it as a woman who works in energy compliance and has watched two booms in this state end and the state survive both. Straight talk: the question asks whether the capital spending and the power demand are creating new systemic risks, and the honest answer from the Plains is that the power demand is the most solvable problem in the American economy and the capital spending is the first time in a generation that private money has been willing to pay for the solution. For twenty years the utilities in this country have been told that load was flat, that no new baseload would be needed, that the grid could be run down and the gas plants retired and the transmission deferred. The data centers have ended that. The companies have announced fifty gigawatts of their own gas generation, they are negotiating to pay directly for substations and transmission, and in this state they have come with checkbooks to a place that has gas it cannot sell, wind it cannot export, and a grid that was built for an oil boom that has already plateaued. A compliance officer who has spent years explaining to developers why the interconnection queue is five years long is not going to call it a systemic risk when the developers finally decide to build their own.

Let me take the numbers, because the skeptics have them and I want to show what they look like from a state that produces the inputs. Six hundred billion dollars from the five largest companies this year, three-quarters on this; seven hundred and fifty from the fourteen largest operators; borrowing at a hundred and ninety-four billion by August against a hundred and eight for all of last year; capital spending approaching operating cash flow; a hundred and twenty billion in special vehicles. Those are large numbers and they are being spent by the five most profitable companies in the history of the world, on a product they say is demand-constrained, with chips they say are renting for more this year than last. The comparison the skeptics reach for is the telecom build-out of 2000, and the comparison fails on the one fact that matters: the telecom companies were spending borrowed money with no cash flow, and these companies are spending cash flow first and borrowing second, and even the borrowing is against balance sheets that could carry it twice over. A company that is approaching the point where capital spending equals cash flow is a company that is spending a great deal. It is not a company that is insolvent, and the difference is the whole difference.

Now the power, which is my beat and the part of the question I will defend most firmly. The industry’s own analysts say a third to a half of the capacity planned for this year will slip to 2028 for lack of power. I would say that is the system working, not failing: the capacity that cannot be powered is not being built, which is what a constraint is supposed to do, and the companies have responded by becoming energy developers, which is what a constraint is supposed to cause. Fifty gigawatts of new gas generation is more than the Plains states have added in a generation, and it is being built by parties that can pay for it, in places like this one where the gas is flared because there is no pipeline to carry it. The Iran war raised gas prices since March and the skeptics say that squeezes the economics; a producer state says it raises the value of every molecule the data centers will burn, and the data centers will burn it here, where it is cheapest, which is why they are coming. The structures the utilities are negotiating, in which the large customer pays directly for the generation and the transmission, are the structures this state has wanted for twenty years, because they put the cost of new load on the party that caused it rather than on the farmer at the end of the line, and I will note that the critics who complain about ratepayers subsidizing data centers are complaining about the arrangement the data centers are now offering to replace.

On the debt, I will concede what a compliance officer concedes. The special-purpose vehicles are a structure that has failed before, and the Fed raised rates on Tuesday and signaled another, and the ten-year is at four point eight, and the private-credit firms backing the vehicles are not the lenders I would choose for a thirty-year asset. If the forecast is wrong, the vehicles fail first, and the leveraged cloud providers that have signed a hundred billion dollars of leases fail second, and the big companies write off the chips and keep the buildings. That is a real risk and it is a risk to the lenders and the providers and the chip makers, which is to say to the people who chose to take it. It is not a risk to the farmer in Mandan unless the utility built for the load on his bill, and in this state the utilities are increasingly declining to do that, because the data centers are paying, and the states where the ratepayer is exposed are the states that let the utility build on a promise, which is a regulatory choice and not a feature of the technology.

The thing I want to say about systemic risk is the thing this state learned from the Bakken. A boom is not a bubble because it ends; every boom ends. A boom is a bubble when the thing that was built has no use when the price falls, and the Bakken’s wells still produce at sixty dollars and the roads the counties built still carry trucks and the man camps became apartments. The data-center build-out is leaving behind gas plants, transmission lines, substations, and buildings with cooling systems, in places that needed all of those things and could not get anyone to pay for them, and if the chips inside are scrap in five years the plants and the lines are not, and the next use will find them, as the next use found the Bakken’s roads. That is the difference between a bubble and a capital cycle, and the skeptics who count the capital spending and not the capital stock are counting the wrong column.

So my assessment is that the sector is heading toward a sustainable phase in the only sense that matters, which is that the constraint has forced the companies to build the energy infrastructure the country needed anyway and to pay for it themselves; that the financial risk is real and sits with the lenders, the leveraged providers, and the chip makers who chose it; that the ratepayer risk is a regulatory choice that this state is making correctly and others are not; and that the physical plant will outlast the forecast and be useful when it does. Straight talk cuts both ways, so I will say the other way too: if the five companies are wrong about demand, it will be the largest capital write-off in history, and the lenders who financed it off balance sheet will discover what the valley and the Plains both know, which is that the thing you did not put on the books is still in the ground. But the thing in the ground will be a gas plant and a substation in a state that needed both, and I will take that boom over the last one, because the last one left us wells, and this one is leaving us the grid we could never get anyone in Washington to pay for.

Novi, Michigan Correspondent

The automotive industry spent the first half of this decade on the largest capital program in its history, building battery plants and electric-vehicle lines on a forecast that every executive, every analyst, and every government in the developed world agreed on, and it has spent the past two years writing a great deal of it off, and I want to begin there because I manage a supply chain in that industry and I have watched the write-offs arrive one plant at a time, and the AI build-out looks, from a loading dock in Michigan, like the same movie with larger numbers. The forecast was demand-constrained. The capital was borrowed against balance sheets that could carry it. The suppliers tooled for the volume. The governments abated the taxes. And then the customer did not show up at the rate the forecast said, and the plants that were built for a million units a year ran at a third, and the suppliers who had tooled for the million were the ones who went under, because the automaker could idle a line and the supplier could not idle a loan. I say this not to predict the same outcome but to name the structure, because the structure is identical and the structure is what fails, not the technology, and the people who say this time is different are saying the thing everyone says about the technology and nothing about the structure.

So let me describe the structure as a supply-chain manager reads it. Five companies are spending six hundred billion dollars this year, three-quarters on this, and the fourteen largest operators closer to seven hundred and fifty, against under four hundred and fifty last year, and the analysts have raised next year’s estimate by half in six months. That is a capital program growing at a rate the auto industry never approached even in 2022, and it is being spent through a supply chain that is as concentrated as anything I have ever seen: one chip designer with most of the market, one foundry that makes the chips, one or two makers of the memory, a few builders of the servers, and a set of leveraged cloud providers who lease the capacity to the five companies, a hundred billion dollars of leases in six months. In my industry a supply chain that concentrated would be flagged by every risk officer in the building, because a single disruption at any node stops every line downstream, and the AI chain has three or four nodes that are single points of failure, and the five companies have signed contracts that assume none of them fails. The auto industry learned about single points of failure in 2021, with a chip shortage, and the lesson cost it millions of vehicles. The AI chain has not had its lesson yet.

On financing, the part the write-offs taught me to read first. The companies borrowed a hundred and ninety-four billion dollars by August against a hundred and eight for all of last year, and the analysts say capital spending is approaching operating cash flow, which in my industry is the line where the CFO starts attending the capital-committee meetings. A hundred and twenty billion dollars has been moved into special-purpose vehicles financed by private credit, which is the structure the auto industry used for its battery joint ventures, in which the plant is owned by a vehicle that owns nothing else and the automaker’s exposure is a lease and an offtake agreement, and I have watched three of those vehicles fail in two years when the offtake was not taken, and the lenders discovered that a lease on a plant that is not needed is a lease on a building. The Fed raised rates on Tuesday and signaled another, and the ten-year is near four point eight, which means that every vehicle priced at last year’s rate will refinance at this year’s, and the leveraged cloud providers whose leases the five companies signed are the ones who will refinance first and worst.

On the power, which reaches a Michigan supplier two ways. The first is the grid: the data centers have arrived in this state with load requests that the utilities are building gas plants to serve, and the rate cases are allocating the cost, and a plant like ours that runs three shifts on industrial power is watching its rate go up to pay for a customer that has not yet plugged in, and the utilities’ answer, that the data centers will pay directly, is an answer the auto industry has heard from every large customer that ever asked for a dedicated substation and then moved. The second is the supply chain itself: the industry’s own analysts say a third to a half of the capacity planned for this year will slip to 2028 for lack of power, and the companies have announced fifty gigawatts of their own gas generation to get around it, and the turbines and transformers and switchgear for fifty gigawatts are made by the same handful of manufacturers that make them for everyone else, and the lead times have gone from a year to three, and a supplier in Michigan that needs a transformer for an expansion is now behind a data center in the queue. The Iran war raised the gas price since March and the transformer price with it. The build-out is not only consuming power; it is consuming the equipment that makes power, and every other industry is paying the lead time.

I want to say what I think is actually sustainable, because the question asks and the auto industry’s experience has an answer. The electric-vehicle forecast was wrong in rate and right in direction: the vehicles are being built, more slowly, by the companies that did not overbuild, on the plants that survived the write-offs. The AI build-out will likely be the same. The technology will be used; the demand will grow; the companies that spent cash flow will keep their buildings and write off their chips; the companies that spent borrowed money through vehicles will lose the buildings; the leveraged providers will be absorbed or fail; the chip maker will have a very bad year after several very good ones; and the gas plants and substations will be there, serving whoever comes next, with the cost allocated in a rate case to people who did not choose it. That is not a collapse. It is a correction, and it is the ordinary end of every capital program I have ever watched, and the only question is whether it arrives in 2027 or 2029, and the Fed’s rate decision on Tuesday moved it closer.

So my assessment is that the sector is heading toward a correction rather than a sustainable phase, that the correction will be in the financing and the supply chain rather than in the technology, that the systemic risk is concentration at a handful of nodes financed off balance sheet at a rising rate, and that the power demand is real and is consuming not only electricity but the equipment and the lead times that every other industry needs. I prefer paper maps because they do not change while you are reading them. The capital-spending map for this industry has been redrawn every six months for two years, each time larger, and I have seen a map redrawn like that before, in 2021, with charging stations where the battery plants were going to be, and the map was wrong in rate and right in direction, and the suppliers who believed the rate are the ones who are not here to read the new one.

Dublin, Ireland Correspondent

Ireland got to this question before the Americans did, which is unusual for us, and the way we got there is the thing I want to describe, because it is the question’s answer in miniature. Dublin hosts the European operations of most of the American technology companies, and in the 2010s it hosted their data centers too, on the outskirts of the city, in grey sheds the size of several football pitches that nobody noticed until the grid operator announced that they consumed a fifth of the country’s electricity and that it could not connect any more of them without risking the lights in the capital. The state then did a thing the Americans have not yet done, which is to say no: a moratorium on new connections in the Dublin region, a requirement that new data centers bring their own generation, and a national argument about whether a small country should give a fifth of its power to foreign companies’ computers. The companies responded by building elsewhere, by promising to build their own gas plants, and by reminding the government how much corporation tax they paid, and the argument continues. I say this at the outset because the American debate is conducted as if the power constraint were a surprise, and it is not; it is the ordinary limit of a physical system, and Ireland reached it in 2021 and the Americans are reaching it now, at a scale a hundred times larger.

Now the scale, which is the thing a compliance analyst who sells into American financial buyers watches most closely, because our buyers are the people lending the money. Five companies guided six hundred billion dollars this year, three-quarters for this, and the largest fourteen operators are tracking seven hundred and fifty, against under four hundred and fifty last year. They borrowed a hundred and ninety-four billion by August against a hundred and eight for all of last year; the strategists say capital spending is approaching operating cash flow; a hundred and twenty billion has gone into special-purpose vehicles backed by private-credit firms whose names my buyers know well; and the five companies have signed a hundred billion dollars of leases in six months with cloud providers that are themselves leveraged. A compliance desk reads that as a chain: the big companies’ credit supports the providers’ leases, the providers’ leases support the vehicles’ debt, the vehicles’ debt is held by private-credit funds that have raised money from pensions and insurers on the promise of the big companies’ credit. Every link is sound if the first one is. The first one is a forecast of demand for a product three years old, and the Fed raised rates on Tuesday with another to come, and the ten-year is at four point eight, and the chain was priced at a lower rate than it will roll at.

Let me say what the European experience adds, because it is the part the Americans skip. European regulators, having watched the 2008 chain of mortgages and conduits and funds, built rules that require the exposures in such a chain to be reported and capitalised, and the private-credit vehicles that now hold American data-center debt are, by design, outside those rules on both sides of the Atlantic, which is why the money went there. The European Central Bank has said in its financial-stability reviews for two years that it cannot see inside private credit and does not like what it cannot see. My buyers in American banks say the same thing privately and something more cheerful in public. The systemic question is therefore not whether the technology works but whether the exposure can be seen, and the answer at present is that a hundred and twenty billion dollars of it cannot, and that is the exact structure that produced the last crisis, and the Europeans wrote the rules to prevent it and the Americans wrote the vehicles to avoid them.

On the power, the comparative lesson is short. The American industry’s own analysts say a third to a half of the capacity planned for this year will slip to 2028 for lack of connection, which is the Irish moratorium arriving by physics rather than by regulation, and the companies have answered with fifty gigawatts of their own gas generation, which is the Irish compromise arriving by necessity. The Iran war has raised the gas price since March on both sides of the Atlantic, and the Europeans, who learned in 2022 what gas dependence costs, are watching the Americans build a new industry on it with the fascination of someone watching a neighbour install a boiler of the type that exploded in their own house. The difference is that the Americans have the gas and we did not. The similarity is that the price is set by a strait neither of us controls.

I want to say the plain thing about sustainability as the Irish understand it. Our data centers are still here. The companies that built them are still paying tax. The moratorium did not kill the industry; it moved it, and made it pay for its own power, and the sheds on the outskirts of Dublin are a permanent feature of the landscape and the economy, consuming their fifth of the grid, and nobody now imagines they will go away. That is what a sustainable phase looks like: not growth at the rate of the forecast, but a settled presence at the rate the physical system allows, paid for by the parties who use it. The Americans will get there. The question is whether they get there by regulation, as we did, or by a correction in the chain I described, and the chain is the thing that will decide it, because a physical constraint produces a slower build and a financial constraint produces a sudden one.

So my assessment is that the sector is heading toward a sustainable phase by way of an unsustainable one, that the power constraint is real and self-correcting and the Irish have the proof, that the financial structure is the risk and is deliberately placed where regulators cannot see it, and that the systemic exposure is not the technology companies, who will write off their chips and keep their buildings, but the private-credit chain beneath them, which holds pension money against a forecast at a rate that has just gone up. Ireland said no to new connections in 2021 and the industry survived it. Nobody in America has said no to anything yet, and the physics is beginning to say it for them, and the finance will say it next, and the finance says it faster. I walk the canal past the sheds. They hum. They will hum in 2035. The question is who will own them, and the answer, as with every Irish asset that was financed abroad, will be decided in a court, after the music stops, by lenders who had not read the lease.

Tyler, Texas Correspondent

Texas has its own grid, which means that when the data centers come to Texas, and they have come in numbers that would embarrass California, the question of who pays for the power is not a federal question or an interstate question but a question for a legislature in Austin that I have watched for thirty years, and the legislature answered it last session with a law that lets the grid operator shed the data centers first in an emergency and makes them pay more of their own interconnection cost, which tells you two things: that the people who run this state’s grid believe the load is real enough to threaten the lights, and that they do not trust the companies to pay for it without being made to. I write commercial insurance, and I have been asked this year to quote coverage on two data-center projects in East Texas, neither of which I wrote, because the carriers I place with do not yet know how to price a building whose entire value is in equipment that is worth less every quarter and whose entire function depends on a power supply that the grid operator has reserved the right to cut. That is the underwriter’s view of the question, and it is the view I will give, because an underwriter is the one party in this story who is paid to ask what happens when the thing goes wrong and is not paid to hope it will not.

Who pays? Let me go through it. The five companies are spending six hundred billion dollars this year, three-quarters on this; the largest fourteen operators are near seven hundred and fifty, against under four hundred and fifty last year; and the money comes, in order, from the companies’ cash flow, which is running out, from their borrowing, a hundred and ninety-four billion by August against a hundred and eight for all of last year, and from special-purpose vehicles financed by private credit, a hundred and twenty billion so far, which is where the money comes from when the first two sources are spoken for. The private credit comes from pension funds and insurance companies, which is to say from the retirements of teachers and the reserves that back the policies I write, and the pension funds and the insurers bought the vehicles because they yielded more than Treasuries at a time when Treasuries yielded little, and the Fed raised rates on Tuesday and signaled another, and the ten-year is at four point eight, and the yield that looked generous at three percent looks ordinary at five, and the vehicle’s only asset is a lease to a cloud provider that is itself borrowed to the hilt. So the first answer to who pays, if the forecast fails, is the teachers, through a vehicle they never heard of, and the policyholders, through a reserve that was invested in a building whose chips are scrap.

The second answer is the ratepayer, and Texas has at least tried to do something about it. The industry’s own analysts say a third to a half of the capacity planned this year will slip to 2028 for lack of power, and the companies have announced fifty gigawatts of their own gas generation to get around it, and in this state that means gas plants built next to data centers in counties that abated the property tax for twenty years on the theory that the jobs would come, and the jobs are a few dozen technicians and the tax base is the equipment, which depreciates on a schedule the company sets. The legislature’s law makes the companies pay more of the interconnection and lets the grid cut them first, which is more than most states have done, and it does not change the fact that the gas plants are being built on an accelerated schedule by a utility system that will recover the cost from everyone if the data center that caused it is not there to pay. The Iran war raised the gas price since March, which raised the cost of every one of those plants and the value of every molecule they will burn, and the two cancel for a producer state and do not cancel for the ratepayer in a town that does not have a well.

The third answer is the one the underwriter sees and nobody else is pricing, which is the single point of failure. The chips come from one designer and one foundry. The transformers and the turbines for fifty gigawatts come from a handful of manufacturers whose lead times have gone from one year to three. The leases, a hundred billion dollars in six months, run through a set of cloud providers that are leveraged and that the big companies have chosen to lease from rather than build themselves, which tells you the big companies also wanted the risk off their balance sheets. An underwriter does not insure a risk with one supplier, one customer, and one lender without a premium that reflects it, and the market has not charged that premium, because the market is pricing the forecast and not the structure, and the structure is what fails. I have paid claims on a lot of businesses that had a perfectly good product and one supplier.

On sustainability, I will give the planning-commission answer, because it is the one East Texas knows. Growth and capacity are not the same word. The growth is real: the companies say demand is constrained, the rents on the chips have risen, the buildings are full. The capacity is the question: the power, the transformers, the water, the debt, the pension money, the ratepayer, the county’s abated tax base. Every boom I have watched in this state outran its capacity and then found it, and the finding was always paid for by the parties who had not been asked, and the law in Austin is the first time I have seen a legislature try to ask them in advance, and I would rate it as half a solution, because it protects the grid and not the pension fund, and the pension fund is where the money came from.

So my assessment is that the sector is not heading toward a sustainable phase; it is heading toward the moment when the three parties that have been enlisted without being asked, the ratepayer, the pension fund, and the county, discover that the forecast was the companies’ and the capacity was theirs. The technology will survive the discovery, as technologies do. The vehicles will not, the leveraged providers will not, and the chips will be written off by companies that can afford it, and the gas plants and the substations will be in East Texas, serving someone, with the cost in a rate case. Who pays? I have named them. The one party I cannot find in the list is the forecast, which is the thing everyone is lending against, and a forecast has never paid a claim, and I have been in this business long enough to know that when the claim comes the party who wrote the forecast will have a very good lawyer and the party who held the lease will have a building full of scrap in a county that already spent the tax.

Las Vegas, Nevada Correspondent

This city was built on a bet that a hundred economists in 1950 would have called a bubble, and the bet was that enough people would drive into a desert to lose money at games they knew were rigged that a city could be built on the proceeds, and the bet paid, and the people who called it a bubble were right about every individual casino that went broke and wrong about the town. I want to begin there because the question asks whether the AI build-out is heading toward a sustainable phase or a systemic risk, and the honest answer from a town that has watched a dozen booms is that it is both, that the individual bets will go broke in large numbers and the thing they are betting on will be here in twenty years, and that the trick is to tell which casino you are standing in. The money is real. The five largest companies are spending six hundred billion dollars this year, three-quarters on this, and the fourteen largest operators are near seven hundred and fifty, and in this valley that has meant a string of data centers along the interstate, built in the heat, cooled with water the valley does not have, powered by a utility that is building gas to serve them, and staffed, after construction, by almost nobody. The floor looks finished. The quiet hours after it looks finished are when the operations man sees what it costs to run.

So let me take the floor as a floor man takes it, which is by asking who is the house, who is the player, and who is the marker. The house is the chip maker, which sells to every player at a margin the casinos here would envy and does not care who wins. The players are the five companies and the leveraged cloud providers they have leased a hundred billion dollars of capacity from in six months, and the players are, by the credit analysts’ own account, approaching the point where capital spending equals operating cash flow, which in this town is the point where a player stops betting his winnings and starts betting his credit line. The marker is the debt: a hundred and ninety-four billion borrowed by August against a hundred and eight for all of last year, and a hundred and twenty billion moved into special-purpose vehicles backed by private-credit firms, which is the casino’s way of extending credit without the credit showing on its own books, and the vehicles are the part of the floor where the dealers do not look up. The Fed raised rates on Tuesday and signaled another, and the ten-year is at four point eight, and in this town when the cost of the marker goes up the players who were betting the credit line are the ones who get walked to the cage.

Now the thing a vendor knows that the analysts skip, which is the cost of keeping the floor running. A data center in this valley burns power to compute and burns more to cool, and the cooling is the part the desert makes expensive, and the utility has told the regulator that the new load requires new gas plants and that the cost will be allocated, and the allocation is the fight, because the resorts on the Strip are the utility’s largest existing customers and they have discovered that the data centers are being charged a rate the resorts would like to have, and the residential customers have discovered that both are being charged a rate the residential customer is subsidizing. The industry’s analysts say a third to a half of the capacity planned this year will slip to 2028 for lack of power, and the companies have announced fifty gigawatts of their own gas, and in a desert that means gas plants with cooling of their own, and water, and the water is the thing nobody in the build-out has an answer for, because the Colorado River does not care what the forecast says and the valley’s allocation is already cut. The Iran war raised the gas price since March. The desert raised the water price long before that, and it raises it every summer, and the data centers were built as if it did not.

On sustainability, I will say what the town’s history says, because it is the one thing the town is expert in. The booms that lasted here were the ones where the thing being built had a use beyond the bet: the hotels were hotels, the convention center was a convention center, and when a casino went broke the building became a different casino. The booms that did not last were the ones where the building was the bet, where a tower was built to be flipped and not to be slept in, and 2008 left a dozen of those as concrete shells on the Strip for a decade. A data center is a building whose value is in its chips, and its chips are worth less every quarter, and the building without the chips is a warehouse with a very good cooling system in a place with no water, and the question the town would ask is what the second use is, and I have not heard an answer, and the people who built the shells in 2008 did not have one either.

On systemic risk, the floor man’s reading. The risk is not the five companies, who can write off the chips and keep the buildings and have done it before. It is the chain beneath them: the leveraged providers, the private-credit vehicles, the pension money inside the vehicles, the utilities that built the gas plants for the load, and the counties that abated the taxes, all betting on the same forecast at the same table with money that is partly theirs and partly the marker’s. When one player at a table of correlated bets loses, the table does not lose one player. It loses the table, and the house walks away with the chips it already sold, and the cage is left to collect from the markers. I have watched a table go that way. The house was fine. The players were not. The waiters who served the table were paid for the night and not for the week.

So my assessment is that the build-out is sustainable in the sense that the city is sustainable, which is that the thing being bet on will be here long after most of the people betting on it have been walked to the cage, and unsustainable in the sense that the current players are betting the marker at a rate that just went up on a floor whose second use nobody has named, in a desert whose water nobody has priced. The house is the chip maker. The marker is the private credit. The players are approaching their cash flow. The town has seen this hand before, and the town is still here, and so are the shells, and the shells are the part the players never thought about on the night the floor looked finished, which it does, right now, at every exit on the interstate, humming, in the dark, with the lights of the Strip behind it and the river’s water in a lake that is lower every year.

Billings, Montana Correspondent

The coal plant east of here that was going to close has been told it may not, because the data centers want the power, and the town that depended on the plant has been told that its future is a set of grey buildings that will employ forty people and consume what the plant makes, and I want to begin there because it is the question in a single county: a thing that was dying has been kept alive by a forecast, and the people who live next to it are being asked whether to believe the forecast, and they have no way to evaluate it, and they have voted yes, because the alternative was the plant closing, and a county does not choose between a forecast and nothing. It chooses the forecast. I adjust claims and write a civic column, and I have sat in the meetings where the data-center companies came with their nondisclosure agreements and their tax requests and their promises about water, and I have watched a county commission that could not get a straight answer about the water vote for the abatement anyway, because the plant was the county and the plant was closing.

So let me assess the question the way the county should have and did not, which is by separating the thing being built from the reason for building it. The thing being built is power: the companies have announced fifty gigawatts of their own gas generation, they are keeping coal plants open that were scheduled to close, they are negotiating to pay directly for substations and transmission, and the industry’s own analysts say a third to a half of the capacity planned this year will slip to 2028 because the power is not there. In Montana that is the largest investment in the grid since the dams, and I will say plainly that the state needed it, that the federal government was never going to pay for it, and that a coal plant that runs another fifteen years to serve a data center is a coal plant that keeps a town alive for fifteen years, and the people in the town will take it, and I would not tell them not to. The reason for building it is a forecast of demand for a product three years old, made by five companies spending six hundred billion dollars this year on it, three-quarters of their capital budget, with borrowing at a hundred and ninety-four billion by August against a hundred and eight for all of last year, and a hundred and twenty billion moved into vehicles financed by private credit, at a rate the Fed raised on Tuesday. The thing is durable. The reason is not. The county has bet the first on the second.

Now the water, because it is the part the companies will not discuss and the part Montana cannot ignore. A data center cools with water or with power, and in a dry state it uses both, and the river that the plant east of here draws from is the river the ranches draw from and the river the tribes have senior rights to, and the companies have promised closed-loop systems and have declined to say how much the loop loses to evaporation in July, and the county signed anyway. I fix fences badly and I know a fence that will not hold stock when I see one, and a water agreement that does not state a number is that fence. The sustainability question, in a state where the limiting resource is not capital but water, is whether the thing built on the forecast will have water when the forecast is tested, and the answer is that nobody asked, because the companies would not say and the county would not insist, and that is how every boom in this state has started, with a promise about water that was not written down.

On the debt and the systemic risk, I will take the adjuster’s view, which is that the loss falls where the policy says and the policy here was not read. The vehicles that hold a hundred and twenty billion dollars of data-center debt are financed by private-credit firms that raised the money from pensions and insurers, and the leases inside the vehicles run to cloud providers that are leveraged, and the big companies have signed a hundred billion dollars of leases with those providers in six months, which tells me the big companies wanted the risk somewhere other than their own books. An adjuster recognizes that chain. It is a loss that has been transferred three times, and every transfer was made by a party who understood the risk better than the party who accepted it, and the last party in the chain is a pension fund that bought a yield, and when the loss arrives it will arrive there, and the people whose pensions those are will not have been asked either, and in Montana those people are the state’s teachers and the state’s retirees, and they are the county’s neighbors.

I want to say the thing about the forecast that a man who adjusts hail claims knows. A forecast is not a lie. The companies believe it; they are spending their own cash flow on it, and cash flow is the most honest thing a company has. But a forecast made by the party that profits from it, financed by lenders who do not see the asset, built in counties that cannot evaluate it, on water that is not counted, at a rate that just went up, is a forecast that has been protected from every check that would ordinarily test it, and a thing protected from every check is the thing that fails hardest when the check finally comes. The Bakken had a price to check it. The telecom boom had a market. This has a forecast and a set of counties that voted yes because the alternative was nothing.

So my assessment is that the build-out is leaving behind things Montana needed and could not pay for, power plants and transmission and a reason for a town to exist, and that it is doing so on the strength of a forecast that no party in the chain has an interest in questioning and no county has the means to, financed by debt that has been transferred away from everyone who understands it toward everyone who does not, and cooled by water that was never counted. Whether that is sustainable depends on whether the forecast is right, and I do not know, and neither does the county, and the county voted anyway, as counties do. The plant is still running. The town is still there. The grey buildings are going up by the river. The water agreement does not have a number in it. I have adjusted a lot of claims that began with an agreement that did not have a number in it, and the claim always came, and the number always turned out to be larger than anyone had been willing to write.