An Amazon Web Services data center towers above a suburban residential neighborhood in Northern Virginia.
Getty Images
vinegarStartup Cloverleaf Infrastructure last week announced a massive initial round of funding of $300 million from Sandbrook Capital and NGP.
The company didn’t disclose how much stock it gave up in exchange for the growth capital. Cloverleaf co-founder Brian Janos said the company’s goal is to provide a site on which to begin building hyperscale data centers specialized in processing AI queries.
But this isn’t just about selling land to house artificial intelligence. Before companies like Google, Meta, Amazon, and Microsoft invest $10 billion in gigawatt-scale data centers, Cloverleaf believes they’ll need some help securing the land, preparing the permits, and arranging for adequate power supply and transmission. “The biggest challenge for this industry isn’t access to GPUs. It’s power,” says Janous, who was Microsoft’s first vice president of energy.
Cloverleaf co-founder Brian Janos.
Coatest Clover Leaf
In the early days, large data centers used 10 megawatts of power, enough for about 8,000 homes. “We grew up in an era when utilities were cutting back on increased power loads from things like LED lighting.” Then cloud computing came along, and suddenly demand for 200-megawatt centers rose. “Utilities weren’t prepared for the growth of the cloud.” According to EPRI, data center electricity use doubled from 2017 to 2021.
Then the rise of AI changed everything. Demand is exploding. That’s because AI models are power hungry: one chatGPT request uses 2.9 watt-hours, a Google search uses 0.3 watt-hours. Data centers are a growth business that can’t be overlooked, says Janous, who spent 12 years at Microsoft, coordinating 15 gigawatts of power to support cloud systems like Azure, before leaving a year ago. “Microsoft wants every gigawatt, as many GPUs as they can get. They want to scale brute force calculations,” says Janous.
Data centers currently consume 4% of U.S. electricity, but that’s expected to rise to 9% by 2030. Grid operators aren’t allowed to redirect electricity from suburban homes to data center customers who pay higher rates, so they must be convinced that enough additional power is being generated at the same time before grid regulators will allow new data centers to be built and run.
“This is a complex power orchestration problem,” Janos says. “It has nothing to do with money, but with the difficulty of physically operating the grid and building transmission lines.” Can they build enough power to keep up? For example, to provide 1,000 megawatts, or 1 gigawatt, of power 24/7, you’d need about 1,000 giant wind turbines on 100,000 acres, or 15,000 acres covered in 20 million solar panels. Plus, you’d need enough battery backup to make up for the power shortfall at night or when the wind dies down.
Utilities may have to upgrade 500-kilovolt transmission lines, which would involve getting regulatory approval and choosing among other worthy projects. “It’s about responding to consumer demand. The problem is that utilities have to choose between megaprojects: data centers versus factories,” said Ken Ryan, who studied business at Riverstone and Goldman Sachs.
Cloverleaf has extensive transmission expertise from CEO and co-founder David Berry, who previously developed renewable energy projects with ConnectGen and Clean Line Energy Partners.
As for Cloverleaf’s potential market, analysts at TD Cowen estimate that the U.S. will build roughly 40 gigawatts of hyperscale data centers by 2028. Rob West, an analyst at research firm Sundar Said Energy, estimates that the initial capital costs for these centers will average about $10 million per megawatt of computing power, which would mean an initial investment of $400 billion in data centers across the country.
Ryan marvels that the profit margins from converting electricity into AI data could be huge. “There’s no better business in the world than the energy conversion business,” he says. “Customers will pay a premium. Power and land are a small fraction of the total cost.” Cloverleaf aims to locate projects in places where green power is plentiful, such as Wisconsin, Ohio, and Indiana.
Some long-term thinkers are already laying the groundwork for new zero-carbon, small-footprint power sources to power AI: OpenAI co-founder Sam Altman is a big investor in Helion, a nuclear fusion engine startup that pre-sold 50 megawatts of its future green power generation to Microsoft last year, while Bill Gates is investing $1 billion in a new sodium reactor currently under construction in Wyoming. Amazon Web Services’ Jeff Bezos has invested $130 million in General Fusion.
So what does longtime energy systems observer Janos see as the unexpected loser of the AI data center boom? “It’s going to be tough for hydrogen,” he says. Despite its newfound hype, hydrogen will be in direct competition with AI for the green electricity needed to run electrolyzers that split water molecules. “How can you justify converting clean electrons into hydrogen when you have Google and Microsoft looking to buy that electricity for their data centers?”
