Lynn Doan of Bloomberg asked what one AI data center costs. The answer, at the high end, is about $30 billion for a single gigawatt in the Abilene area of Texas. A gigawatt is enough power, at one moment, for roughly 750,000 homes. The building that uses it is not a warehouse with a few servers. It is a power plant, a chip order, a cooling plant, and a construction job, stacked into one address.
Here is the only idea that matters for an investor. That $30 billion is two checks, and they fail in different ways. The larger check is chips and servers. On the day the story ran, Doan put that line near $17.1 billion. The smaller check is the real world: power gear, cooling, construction, steel, cement, copper, land, and the contingency you add because the first number was a wish. The chip check is a bet that AI keeps earning. The building check is a bet that the build does not stop. Own one, and you do not own the other. A rising copper price does not make a chip profitable. A clever model does not pour a foundation if the pipeline is late.
This is not advice to buy a miner, a chip maker, a bond, or a data-center landlord. It is a way to read a bill.
The four companies, then the one building
Bloomberg starts with a wider number, and it is the one that makes people stop. Meta, Amazon, Microsoft, and Alphabet, which owns Google, are expected to devote roughly $750 billion to capital spending this year. Bloomberg says that is more than the whole economy of Ireland or of Belgium. It is more than half of what every company in the S&P 500 spent on capital projects, taken together, last year. Spread as a flat check, it is about $5,500 for every household in the United States.
John McWilliams, in the research group at Cushman & Wakefield, told Bloomberg he once would have found that sum impossible. Four firms, one year, mostly data centers. "We've never seen it," he said. Then his team priced the shell. Bloomberg Intelligence priced the hardware. At the high end of both, the theoretical Abilene gigawatt lands near $30 billion. McWilliams's second line is the useful one. In that context, he said, the big number "doesn't seem as crazy as the number might suggest." The facilities are extremely expensive. The $750 billion is what you get when many extremely expensive facilities are ordered at once.
Abilene is not a random dot. Bloomberg notes that some of the world's largest computing sites are going up there. A later account of the same story said an OpenAI campus in that area was expected to draw about 1.2 gigawatts when finished. The $30 billion site in the article is still a model, not a receipt from that campus. A model is a clean way to see a mess. It is not a bid.
Where the $30 billion sits
Doan walked through the split the same day on Bloomberg's own broadcast. About $17.1 billion, she said, is chips and servers. About $3 billion is power. About $2.4 billion is construction. About $1.7 billion is cooling. About $2.8 billion is contingency. The rest of a bill that Bloomberg's article rounds to roughly $30 billion is the smaller lines: network gear, engineering, land, commissioning, and the items that never make the headline. Those spoken figures are a high-end sketch from Cushman & Wakefield's shell costs and Bloomberg Intelligence's hardware costs. They are not a contract. They are enough to see the shape.
The shape is the surprise. People hear "data center construction" and picture steel and jobs. Steel and jobs are in the bill. They are not most of the bill. More than half of this theoretical gigawatt is the computers that go inside. A mining investor, a turbine maker, and a cement firm are arguing over the smaller half. The larger half is a semiconductor order. If you came to this story for copper, you are in the room. You are not holding the biggest pen.
That does not make the smaller half small. Three billion dollars of power equipment for one gigawatt is a shock to a supply chain that was not built for it. Bloomberg says demand for gear that generates and carries electricity is so strong that supplier backlogs stretch past 2030. Cooling at about $1.7 billion is a factory of its own. Construction at $2.4 billion is the shell people can see from the road. Contingency at $2.8 billion is an admission. The planners do not believe the first pass. They have left a pile of money for the pass that goes wrong.
Why every line is getting worse
Bloomberg's account of the inflation is plain. Steel and cement have jumped because of trade fights, fuel costs, and wars. The AI boom itself has pushed up chips and copper. Short-term memory used in these centers, Bloomberg Intelligence's data show, costs more than five times what it cost a year ago. That is not a gentle cycle. That is a part that repriced in twelve months because too many buyers wanted the same shelf.
Doan, describing the work with Cushman and Bloomberg Intelligence, said the bottleneck runs across raw goods, not just chips. Natural gas turbines. Power transformers. Concrete and cement. Copper. Steel. A site can be fully financed in a spreadsheet and still wait on a machine that will not be delivered before the next decade. The wait is a cost. Interest runs while the turbine is a line on someone else's order book. So does the chance that the plan gets redrawn into something dearer.
Local fights and permits add a second kind of cost, and Bloomberg gives it a name. Last month, Oracle issued a force majeure notice on a giant data center in New Mexico after a gas pipeline the project needs ran into regulatory delays. Force majeure is the clause that says a party could not perform because of an event outside its deal. A pipeline hearing is not a chip shortage. It can still stop the site. Developers then look for costlier power. The gigawatt does not care that the model was elegant. It cares that the gas arrived.
Put those together and McWilliams's line holds. Virtually every major category in the build is ballooning. The $30 billion is not one spike. It is many spikes, added.
Sixty gigawatts is not sixty times a sure thing
Bloomberg says that over the next four years, more than 60 gigawatts of this kind of capacity is expected to be added in the United States alone. That would triple the computing power of the country that already has the most. The same build is underway in Brazil, China, Britain, and the rest of Europe. That is how estimates of a boom measured in trillions start to look like addition instead of hype.
Do the multiplication once, and then refuse to worship it. Sixty gigawatts times $30 billion is $1.8 trillion. That product is not a forecast. It is what you get if every new American gigawatt cost what this theoretical Abilene site costs at the high end. Many will cost less. Some will cost more, once delays and redraws land. Some will be announced and not finished. The product is a ceiling check. It tells you the unit is large enough that a national program is a macro event. It does not tell you the program will be completed at that unit price.
The ripple Bloomberg describes follows from the unit, not from the slogan. Turbines, transformers, construction, real estate, hardware, and network gear all get a bid while the sites are being ordered. Private credit, bonds, and stocks have all been flooded with the financing. Jobs are tied to it. So is a stack of debt. The bid is real while the order book is real. It is not a claim that the computers will earn their keep after the ribbon is cut.
The rent is the part that is not in the invoice
Bloomberg states the wager without softening it. These centers are being built on the assumption that demand for AI systems will keep growing, and that the chips inside will pay back for years. Some of the most powerful builders of AI models have not turned a profit. The demand outlook, the article says, is far from certain.
Michael Burry, known for the bet against the housing bubble, has warned of an AI bubble, and he has pointed at the circular deals. Hardware makers, Nvidia among them, invest in and partner with big customers, OpenAI among them. The supplier helps fund the buyer. The buyer then buys the supplier's goods. Burry's line, as Bloomberg quotes it, is that all the capital spent and lent by the richest companies on earth will not be enough to stop the bubble from bursting. Ray Dalio of Bridgewater has warned of a bubble too. A warning is not a date. A circular deal is not proof of a fraud. It is proof that some of the demand is being financed by the same boom it is supposed to justify. An investor should see that loop and not pretend it is a customer walking in off the street.
This is where the two checks separate for good. The chip check, $17 billion of a $30 billion site, is paid up front to the companies that make the machines. Those companies can book the sale when the box ships. The earning that is supposed to justify the box comes later, from use. If use disappoints, the chip maker may already have been paid, and the owner of the building may be the one holding a room of machines that do not earn the interest on $30 billion. If the build is cancelled before the box ships, the chip order shrinks, and the construction invoice shrinks with it. Same boom. Two clocks.
What a mining investor is actually holding
Copper is in this story because Bloomberg put it there. The boom has driven up the cost of chips and of copper. Transformers, cable, and the power hall are copper-heavy. Turbines and the steel around them are in the same sentence Doan used. Cement is in it. None of that is a reason to treat a copper mine as a claim on AI profits.
A miner, a royalty, or a turbine firm gets paid if someone still writes the building check. They get paid while the holes are drilled, the wire is pulled, and the machine is delivered. They do not get a second pay when a model finally makes money. If the four companies keep spending near this year's pace, the invoice for metal and machines stays large even if the models are still losing money. If the four companies cut, because the rent looks weak or because a pipeline notice becomes a habit, the invoice shrinks even if the technology is real. The metal follows the crane. It does not follow the demo.
The share of the bill matters, and it should humble the pitch. On this high-end sketch, chips and servers are more than half. Power is about a tenth. Construction is less than a tenth. Copper is a slice of the power hall, the cable, and the cooling plant, not a slice of the $17 billion chip line. A shortage of transformers can delay a gigawatt. It cannot turn a $3 billion power budget into the whole story. Investors who buy a copper name because "AI needs metal" have bought a true sentence and a false size. The sentence is that the build uses copper. The size is that copper is not where most of the $30 billion goes.
The other humble fact is time. Backlogs past 2030 mean the physical world cannot absorb $750 billion of intent on a software schedule. Some of the capital will sit in queues. Some will be redesigned around whatever power can be had. Oracle's New Mexico notice is the small version. A hearing delayed the gas. The data center did not get to ignore the hearing. Multiply that by a national program and the smooth GW charts will not all become buildings. The ones that do will have paid up for scarce gear. That payment is the miner's window. It is also the developer's problem.
What would make this reading wrong
The reading is wrong if the high-end Abilene model is a scare price, and typical sites come in far below it. Bloomberg said "roughly $30 billion" at the high end, not the average. If the median gigawatt is half that, the $1.8 trillion ceiling check was a ceiling, and the panic about the unit was half panic. The two-check split can still be true at a lower total. You would rewrite the dollars. You would not automatically erase the split.
The reading is wrong if the chip line is not $17.1 billion in the graphic Bloomberg published, and the spoken walk-through rounded a range into a point. Treat $17.1 billion, $3 billion, $2.4 billion, $1.7 billion, and $2.8 billion as the author's high-end description, not as invoices. If the published bars differ, use the bars. The article's own settled phrase is the one to keep: roughly $30 billion to build and house the theoretical site.
The reading is wrong if AI demand does arrive, the chips earn for years, and the bubble warnings were early in the way many bubble warnings are early. Then the rent shows up, the building check and the chip check both look wise, and Burry's loop looks like a financing method that worked. You would have been right to refuse, in October 2026, to count the rent before it existed. You change the count when the profits exist. You do not change it because the spending is large. Large spending is what made the bill. It is not what pays it back.
The reading is also wrong if the build stops for a reason that is not demand. A permit wave, a power crunch, or a memory price that breaks the budget can cut the crane count while the models are still wanted. In that world the building check fails first. Copper, turbines, and cement feel it before the chip maker's long-term story does. The investor who owned "AI" as one trade will not know which check bounced. The investor who split them will.
The idea, once
Four companies may spend about $750 billion on capital projects this year, mostly on the AI build. One theoretical gigawatt near Abilene prices, at the high end, near $30 billion, and more than half of that is chips and servers. Power, cooling, construction, and contingency are the rest, and they are already dear enough to push turbines and transformers past 2030 and to lift copper, steel, and cement. The United States is expected to add more than 60 gigawatts in four years. That is a building boom if the orders survive. It is a stranded boom if they do not. The rent those buildings are supposed to earn is not in the invoice. Separate the chip check from the building check. One is a bet that AI pays. The other is a bet that the crane does not stop. An investor who mixes them will not know what he owns when one of them fails.
A note on sources and limits
The $750 billion capital-spending figure, the comparisons to Ireland, Belgium, S&P 500 capital spending, and a $5,500 household check, the quotes from John McWilliams, the memory price at more than five times a year ago, the equipment backlogs past 2030, the Oracle force majeure on a New Mexico data center, the roughly $30 billion high-end cost of a theoretical one-gigawatt site near Abilene, the 750,000-home comparison, the expectation of more than 60 gigawatts of U.S. additions over four years, the global build, the warnings from Michael Burry and Ray Dalio, and the point that some leading AI model makers have not turned a profit are from Lynn Doan's Bloomberg feature of October 12, 2026. Cushman & Wakefield supplied the shell costs. Bloomberg Intelligence supplied the hardware estimates. The line items near $17.1 billion for chips and servers, $3 billion for power, $2.4 billion for construction, $1.7 billion for cooling, and $2.8 billion for contingency are from Doan's description of that same analysis on Bloomberg's broadcast the day the story ran. They are high-end descriptions, not audited invoices, and they are not a full table. The $1.8 trillion figure is 60 gigawatts times $30 billion, labeled here as arithmetic, not as Bloomberg's forecast. Nothing here is investment advice, a solicitation, or a personal recommendation. Builds get cancelled. Prices fall. You can lose money. Do your own work.