OpenAI is in advanced talks to lease a data center campus in southern Ohio that could cost more than $500 billion to build — by broad agreement, the largest data center project ever announced. When a number that size appears, it stops meaning anything. Half a trillion dollars is an abstraction. So this piece does what a good engineer does with an incomprehensible system: it takes the machine apart, names each component and its cost, and then explains the one thing that matters most — the financing structure underneath it.
The facts here are verified against Wall Street Journal, Bloomberg, New York Times, CNBC, and Department of Energy reporting. The site is a former uranium-enrichment facility in Pike County, Ohio, about 50 miles south of Columbus, developed by SoftBank’s SB Energy on federal land. The target is 10 gigawatts of power — roughly the annual consumption of 8 million US households, more than three times the current largest data center campus on earth. Here is what half a trillion dollars actually buys.
The land: ~$100-300 million
It starts with 3,700 acres of federal land — the decommissioned PORTS uranium-enrichment site, already owned by the US government and to be leased rather than bought. Against everything that follows, the land is a rounding error: somewhere between $100 and $300 million. The most expensive machine ever built sits on the cheapest line item in its own budget. That inversion — land trivial, everything on top of it colossal — is the first tell of what this project actually is. It is not a real-estate play. It is a power-and-silicon play.
The power: ~$37 billion
America’s grid largely dates to the 1970s, and there is nowhere near enough spare power near this site to feed a 10-gigawatt load. So they are not connecting to the grid. They are building their own power plant on site — targeting 9.2 gigawatts of natural gas generation, which would rank among the largest gas-fired plants on earth. Japan agreed to invest roughly $33 billion in this generation as part of a broader US trade deal, in exchange for lower tariffs.
The constraint here is not money. It is physics and supply chains. Gas turbines of this class carry a roughly five-year waiting list. New transmission lines — reported at over $4 billion in upgrades — must be built to move the power from plant to campus. All in, the power layer runs about $37 billion. And the deepest lesson is buried in that turbine wait time: the binding constraint on AI is no longer chips or code. It is the physical ability to generate and move electricity.
The bridge power: ~$20 billion
You cannot simply plug a data center into a power plant. Voltage steps down through substations, and power fails intermittently, so you need backup. Traditional backup is diesel generators — which now carry their own three-year waiting list. To bridge that gap, the project turns to fuel cells.
The scale is worth feeling. A single ceramic fuel cell produces about 25 watts — enough for a couple of light bulbs. A stack of them powers a house. Stacks combine into 300-kilowatt blocks, and companies like Bloom Energy sell arrays of those blocks that deliver megawatts within months rather than years. That speed — megawatts in months instead of turbine-years — is the whole reason this layer exists, and it costs about $20 billion. The bridge power exists to buy time against the supply chains that the main power layer cannot beat.
The building: ~$55 billion
Only now do you build the actual structure. It requires engineers and thousands of construction workers, plus enormous quantities of concrete and steel, to create what is effectively a hurricane- and tornado-proof box — a powered concrete shell designed to outlive every piece of technology inside it. That shell runs about $55 billion.
Sit with that comparison. The concrete box costs more than the entire power-generation layer. And the box is deliberately built to outlast its contents — the shell endures for decades while the silicon inside is obsolete in three to five years. The most permanent thing in the project is the empty building; the most valuable thing is the chips that will be landfill before the concrete has finished curing its reputation.
The cooling: ~$35 billion
Cooling is the source of the fiercest public controversy, because of water. The old method pumped heat from the servers into external cooling towers, where evaporation carried it away and fresh cold water was pumped back in. It is enormously wasteful — at this scale, the evaporative approach would consume roughly as much water as the entire city of San Francisco.
So this project uses a closed loop, more like a car radiator: the water is filled once and recirculated. A cooling plate attaches directly to each chip and connects to a heat exchanger. Racks face each other across a cold alley; pipes pull heat off the chips, carry the hot water past fans that cool it, and return cold water through the alley. Instead of a whole city’s worth of water, the campus uses roughly as much as a single city block. The closed-loop system costs about $35 billion — and it is the difference between a project a community will tolerate and one it will fight.
The network: ~$30 billion
There will be millions of chips in this building, and they are worthless unless they can talk to each other fast enough to act as one machine. Inside each rack, chips communicate through NVIDIA’s NVLink — on the order of two miles of copper cabling per rack — speaking CUDA, NVIDIA’s proprietary software language. To link racks into one coherent supercomputer, the project needs fiber-optic cabling and switches moving data at the speed of light. The networking layer runs about $30 billion, and much of it, too, flows to NVIDIA.
Note what has quietly happened by this layer. The copper standard is NVIDIA’s. The software language is NVIDIA’s. The interconnect is NVIDIA’s. Before we even reach the chips, NVIDIA’s technology defines how the machine is wired and how it thinks. The lock-in is structural, not incidental.
The chips: ~$350 billion
And here is where the half-trillion goes. Each unit in a rack is a high-bandwidth memory system with a logic chip fused onto it. Multiply that by roughly four and a half million units, and you reach about $350 billion — two-thirds of the entire project cost, in silicon, nearly all of it from NVIDIA.
Sum every layer — land, power, bridge power, building, cooling, networking, chips — and the project comes to roughly $527 billion. But the single number that explains the whole enterprise is this: between the chips and the networking, the great majority of that half-trillion flows to one company. NVIDIA is not a supplier to this project. NVIDIA is the project’s primary destination.
The financing structure: why NVIDIA backstops $250 billion
This is the part every leader needs to understand, because it is the part that reveals what kind of bet the whole AI infrastructure boom actually is.
OpenAI is not profitable and cannot obtain an investment-grade credit rating on its own. A project of this size cannot be financed by an unprofitable tenant. So NVIDIA is in talks to backstop roughly $250 billion of the project’s lease and construction debt — to stand behind the loans, using its own $5-trillion-market-cap, 75%-margin balance sheet, so that lenders will fund a facility whose main tenant could not otherwise qualify. As one analyst put it, it is like co-signing a mortgage for someone who lacks the credit score, except the house is a half-trillion-dollar AI factory.
Now follow the money in a circle. NVIDIA guarantees the debt that lets OpenAI lease the building. The building fills with roughly $350 billion of NVIDIA chips plus much of the $30 billion of NVIDIA networking. So NVIDIA is guaranteeing the loans that pay for the purchase of NVIDIA’s own products. By backstopping ~$250 billion of debt, NVIDIA effectively secures itself on the order of $400 billion of revenue. The guarantee is not charity or even ordinary vendor financing. It is a mechanism for converting NVIDIA’s balance-sheet strength directly into NVIDIA’s own future sales.
The circular financing question
This structure has a name that critics use and defenders dispute: circular financing. The concern, articulated by analysts including Bernstein’s Stacy Rasgon and academics like Boston College’s Aleksandar Tomic, is straightforward: when a company finances its customers’ purchases of its own products, it can make end demand look stronger than it organically is. If NVIDIA is underwriting the demand for NVIDIA chips, how much of the AI boom’s apparent demand is real, and how much is manufactured by the seller’s own balance sheet?
The concern is not fringe. When the Ohio backstop was reported, NVIDIA shares fell around 4.5% and its credit-default-swap spreads widened to their biggest single-day gain on record — the market pricing in the systemic risk of an increasingly interlinked web of NVIDIA, OpenAI, SoftBank, SK Group, CoreWeave, and others, tied together by leverage and mutual dependence.
In fairness, defenders make a real case. Vendor financing is common in capital-intensive industries. NVIDIA’s guarantees, while enormous in absolute terms, are set against a company generating hundreds of billions in revenue and cash. And the model works fine as long as the end demand is real — as long as the AI capacity gets used and generates cash flow. The structure is only fragile if end demand weakens. Whether it does is the half-trillion-dollar question, and no one honestly knows the answer. This is not a prediction of collapse. It is a map of where the load-bearing assumption sits.
Mapped to the Mutation Readiness framework
This anatomy maps onto three dimensions of the Mutation Readiness diagnostic — the operational instrument of the Mutation transformation practice we run for enterprise leaders.
Signal Sensitivity — the breakdown surfaces the weak signals the headline number hides. That the binding constraint is power, not chips. That turbine and generator wait times, not capital, gate the build. That the circular-financing structure means a large share of visible AI demand may be underwritten by the seller. A leader with Signal Sensitivity reads these signals inside the $527 billion figure; one without it sees only an impressive number and a confident narrative.
Ambidextrous Capital — the entire NVIDIA backstop is a capital-structure question. Is guaranteeing your customers’ debt to secure your own revenue a brilliant exploit of balance-sheet strength, or an over-extension that manufactures fragile demand? Ambidextrous Capital is the discipline of telling durable, cash-generating investment apart from circular arrangements that inflate the appearance of demand — the exact judgment this deal demands of anyone exposed to it.
Structural Flexibility — the physical reality of the project (five-year turbine waits, three-year generator waits, a concrete shell built to outlast its silicon) is a lesson in what can and cannot be reshaped quickly. The organizations that win the AI-infrastructure era will be the ones that build structural flexibility around the true constraints — power, cooling, supply chains — rather than assuming, as the headline number implies, that money alone moves the machine.
The signals your organization is missing right now
The master signal is your organization’s hidden exposure to the AI infrastructure bet the $527 billion machine represents. Whether or not you ever build a data center, your AI strategy depends on the vendors, the power, and the financing structures this project exemplifies — and most organizations have never mapped that dependence.
Look for the specific tells. Does your AI strategy assume compute will remain cheap and available, when the real constraint is power infrastructure with multi-year lead times? How exposed are you to a single vendor whose technology defines your interconnect, your software language, and your chips — the way NVIDIA defines this build? And is your AI planning resilient to a scenario where the circular-financing structure tightens and compute financing gets harder? Each is a signal the headline AI-boom narrative is trained not to show you.
Three practical questions
One: is your AI strategy bottlenecked on compute or on power? The Ohio breakdown shows the real constraint is electricity with five-year lead times. If your plans assume compute scales with money, you are missing the physical constraint that actually governs the timeline — and so are many of your vendors’ promises.
Two: how concentrated is your exposure to a single AI vendor? In this build, one company supplies the chips, the interconnect, the software language, and the financing. That is extraordinary concentration. Map your own AI stack and ask how much of it depends on one vendor’s technology and one vendor’s balance sheet, and what happens to you if either tightens.
Three: does your AI plan survive a compute-financing correction? The circular-financing structure works while end demand is real and fragile if it weakens. You do not need to predict which way it breaks. You need a plan that is resilient either way — that does not assume the current, seller-underwritten abundance of cheap compute continues indefinitely.
The closing thought
Broken into its parts, the largest machine ever built is strangely comprehensible. A few hundred million dollars of borrowed federal land. Thirty-seven billion to generate power the grid cannot supply. Twenty billion in fuel cells to bridge supply-chain gaps money cannot shorten. Fifty-five billion of concrete built to outlive everything inside it. Thirty-five billion to cool it with a city block’s water instead of a city’s. Thirty billion to make the chips speak to each other. And three hundred fifty billion of silicon that will be obsolete before the concrete weathers. Five hundred twenty-seven billion dollars, assembled into one coherent bet.
And underneath it, the structure that explains everything: the chip company guaranteeing the debt that buys its chips, converting $250 billion of balance-sheet strength into $400 billion of its own revenue. Whether that is the shrewdest industrial financing of the decade or the circular engine of a bubble is the question the whole AI economy now turns on — and the honest answer is that it depends entirely on whether the demand for all that compute turns out to be real.
Every leader is now exposed to that question, whether they build data centers or not. The ones who understand the machine — its true constraints, its concentration, its financing — will navigate the AI infrastructure era with their eyes open. The ones who see only the headline number will be navigating on a narrative built, in part, by the company that profits most from it.
The world has changed. The leaders who notice will be the ones the next decade is built around.
