Why Nvidia Is Willing to Backstop $250 Billion for One Data Center

Everyone’s talking about data centers these days. How much water they use, how much land they swallow, and the eye-watering sums required to build them. OpenAI is in advanced talks to lease one whose total price tag, chips included, is expected to clear half a trillion dollars. So what exactly costs more than $500 billion to stand up, and why is Nvidia discussing a $250 billion financing backstop for it? Let me walk you through it the way I explain these things to my kids when they ask why the power bill is so high—except with fewer eye rolls.

First comes the land. SoftBank’s SB Energy is developing the PORTS Technology Campus on the old Portsmouth Gaseous Diffusion Plant site in southern Ohio, down in Pike County near Piketon. The footprint covers roughly 3,700 acres of federally owned Department of Energy land plus some neighboring private parcels under option. Lease terms remain private, but the economic value of that federal ground is likely somewhere in the low-to-mid hundreds of millions at most. In other words, the dirt itself is the cheap part of the story. As a dad who’s spent more than that on travel baseball and basketball alone—and still ends up with a perfectly good SUV that permanently smells like cleats and gym bags—I find this oddly comforting.

Power is where the real money starts. The existing grid near the site simply doesn’t have enough spare capacity for a 10-gigawatt data center. The solution is to build approximately 9.2 gigawatts of new natural-gas generation as part of a larger ~10 GW package, plus the transmission lines to move it. Japan is putting up about $33 billion under a U.S.-Japan agreement; fold in the $4.2 billion transmission upgrades and the power side lands around $37 billion. Gas turbines still carry multi-year lead times, which means the schedule is measured in patience as much as concrete.

You can’t just run a cable from the plant to the servers and call it done. Voltage has to be stepped down, and the whole operation needs backup for the inevitable outages. Diesel generators are on multi-year wait lists, so the industry has turned to modular solid-oxide fuel cells—think Bloom Energy-style blocks of about 300 kilowatts that can be stacked and delivered in months rather than years. At this scale the backup and bridging package can easily reach the low tens of billions. It’s the technological equivalent of keeping a flashlight and spare batteries in the kitchen drawer, only the flashlight costs more than most people’s houses.

Then you actually have to build the thing. That means engineers, thousands of construction workers, concrete, and steel shaped into essentially hurricane- and tornado-resistant shells designed to outlive every chip that will ever sit inside them. The working estimate for the full shell-and-fit-out at this scale is about $55 billion. These are not temporary warehouses. They are the data-center version of the solid oak table my wife’s parents handed down to us 20 years ago—still standing after two generations of kids, dogs, cats, and one unfortunate science-fair volcano.

Cooling is the part that makes the public nervous. Traditional evaporative cooling towers would pull water on the scale of the city of San Francisco. Reasonable public pressure has largely moved the industry past that approach. The modern solution is closed-loop liquid cooling: a plate sits directly on each chip, heat moves to an exchanger, and dry coolers finish the job. After the initial fill, ongoing water use drops dramatically—more like a single city block than an entire metropolis. The infrastructure to make that work at 10 gigawatts still runs $30–35 billion. Efficiency is expensive until you consider the alternative.

Networking comes next. Millions of chips are useless if they can’t talk to each other at extreme speed. Inside the racks that conversation happens over Nvidia’s NVLink copper connections running the CUDA software stack. Between racks it switches to fiber and high-speed switches. The networking bill alone is roughly $30 billion. It’s the digital equivalent of making sure every kid at the dinner table can actually hear the request to pass the salt.

Finally, the chips themselves. We’re talking the equivalent of roughly four to five million high-end GPUs, each paired with high-bandwidth memory. OpenAI and Nvidia have discussed a package valued up to about $350 billion. That number, plus a healthy share of the networking spend, flows straight to Nvidia.

Add the pieces together and you arrive at a working total of roughly $527 billion. More interesting than the sum is the motivation behind Nvidia’s $250 billion financing backstop for OpenAI’s lease payments and construction-related debt. By helping underwrite the project, Nvidia is effectively guaranteeing a path to hundreds of billions in its own hardware revenue. It’s circular financing at industrial scale—the kind of arrangement that makes perfect sense once you see the whole picture, and still feels a little like watching the same person both sell you the car and co-sign the loan.

This is the first of two posts on the subject. In the next one—What Wholesalers Can Teach Us About Nvidia’s $250 Billion Backstop—I’ll examine the real risks of this kind of circular financing through the everyday lens of wholesalers who extend credit to their distribution chains. The goal is clarity, not panic.

As always, please call or email us if you would like to discuss how large-scale AI infrastructure, the companies supplying the power and silicon, and the financing structures around them may intersect with your financial plan or how we are positioning portfolios around the firms enabling this transformation. These systems can process more data in a second than my household generates in a month of group chats, yet they still can’t settle the debate over whose turn it is to unload the dishwasher. It’s usually mine.

Your capital, our expertise, a bespoke creation.