Nvidia Is Buying Powered Land for a 600-Kilowatt Rack That Just Slipped to 2028
Nvidia bought into three powered land developers in eight days. The 600-kilowatt Rubin rack that justifies it just slipped to 2028, and the substation orders land at Hubbell either way.
Three deals in eight days, none of them for anything you can plug in. The substation orders land years before the chips do.
Nvidia spent the past two weeks buying stakes in companies that don't make anything you can plug in. Cloverleaf Infrastructure lines up land and grid connections. Lancium owns the power behind OpenAI's Abilene campus. SB Energy pours the shells in southern Ohio. Three developers of dirt and substations, and not one of them ships silicon.
The industry calls what they sell "powered land," which means acreage that already has a signed agreement to draw serious electricity from the grid. It's the scarce thing right now. A GPU takes months to build and a grid connection takes years, so Jensen Huang has started buying the slow half.
Three powered land deals in eight days
The Ohio one came first. On August 17, Nvidia disclosed in a securities filing that it would guarantee up to $105 billion of financing for an OpenAI campus in Pike County. It's a backstop, so Nvidia only pays if the borrower doesn't. Nvidia also put $1.5 billion of its own money into SB Energy, the SoftBank-backed developer building the site. The campus starts at 4.25 gigawatts and carries an option for another 3.75. OpenAI signed a 20-year lease on it, and capacity is expected to come online in phases starting in 2028.
Four days later Nvidia took a minority stake in Cloverleaf, a company that didn't exist three years ago. Cloverleaf launched in 2024 with $300 million from Sandbrook Capital and NGP, and it's since sold more than 7 gigawatts of projects to developers. Nvidia won't say what it paid (the Wall Street Journal put it in the several-hundred-million range, which for Nvidia is a rounding error on one week of revenue).
Then on August 24 came the Lancium deal, worth up to $3 billion. Nvidia puts in $2 billion for roughly 20% of the company, with another $1 billion tied to development milestones like landing grid interconnections. That values Lancium and its land at about $10 billion. Lancium has 4 gigawatts leased and a pipeline north of 15.
The 600-kilowatt rack explains the shopping list
A GB200 NVL72 rack, the thing everybody's been installing for the past two years, carries a nominal spec of 120 kilowatts. The GB300 that replaced it lands around 135. The Rubin Ultra NVL576 that Nvidia showed off as the Kyber rack draws 600 kilowatts.
Five times the draw in one generation. A hundred of those racks in a hall pulls 60 megawatts, which is a small city's worth of load sitting inside one building. Air cooling stopped working somewhere around 50 kilowatts, so the liquid loop is mandatory now. So is a much bigger electrical room.
The awkward part is that the 600-kilowatt rack is late. SemiAnalysis reported in July that Kyber slipped by more than a year, to 2028, blamed on yields for the 78-layer circuit board at the center of the rack, and that the NVL576 version could ship in low volume even then because its optics aren't ready. Nvidia's response to Tom's Hardware ran four words, "Our roadmap is intact," which is the kind of statement that answers a question nobody asked. Buying powered land while the rack that needs it slips a year still makes sense, because a grid connection takes longer to get than a circuit board does.
That timing gap is the whole reason Nvidia is shopping for land. Its 2027 chips sell into buildings nobody has finished designing, and the design is now specific enough that Nvidia publishes it. The DSX reference design is a set of blueprints for how much power, cooling and floor space a given block of compute needs, and every one of these three developers has agreed to build to it.
The circular financing argument, briefly
Plenty of people think this looks like Nvidia funding its own demand. Michael Burry has been saying so loudly. Huang pushed back this month with the line that the demand is real and the capital is not Nvidia revenue.
I'd separate the land deals from the customer credit. When Nvidia buys a fifth of a power developer, what it gets is the developer's place in the interconnection queue, and that place exists whether or not anyone ever orders a GPU. The Ohio arrangement is the one that earns the scrutiny. Fortune noted the guarantee landed $145 billion below what had been reported, down from talk of $250 billion for a 10-gigawatt site. Somebody negotiated that gap hard in the days before the filing.
The gear goes in before the GPUs do
Every gigawatt on that shopping list turns into purchase orders for medium-voltage electrical gear, meaning the transformers, switchgear and bus duct that step utility power down to something a server hall can use. Those orders go out two or three years ahead of the chips.
Hubbell reported the cleanest version of this in July. Data center sales rose about 65% in the quarter, and management raised full-year data center growth to around 50%. Its Utility Solutions arm grew 10% on a 1.2 book-to-bill, which just means new orders came in 20% faster than the factory could ship them. Hubbell also said quoting on transmission and substation projects has roughly doubled in two years (a quote isn't an order, but nobody quotes a substation for fun). Substations are what a powered land developer builds first.
The rest of the layer sells into the same halls. Eaton has the transformers and switchgear, Powell does custom switchgear and bus duct for mission-critical sites, and Vertiv supplies the thermal loop and busway that a 600-kilowatt rack can't run without. Our pipeline tracker has the campus-level detail on where those gigawatts are landing.
Nvidia reports Wednesday after the close. Somebody has to build a room that takes 600 kilowatts a rack, and Hubbell is already quoting it.