On the bank of Lake Ontario, about an hour’s drive north of Buffalo, New York, symbols of both past and present energy and technological upheaval collide. A retired coal plant sits alongside a massive 1,800-acre site where the construction of an artificial intelligence data center campus is underway.
The developer, TeraWulf, started out as a cryptominer but has since expanded into AI, with plans to scale to 750 megawatts of high-capacity computing power at its campus in western New York. TeraWulf already has deals to lease space to customers such as the startups FluidStack, backed by Google, and Core42, backed by the UAE government; both will park servers on-site and sell GPUs to their own customers.
This is the world of the neoclouds, a new breed of company that’s emerged in the last few years to try and cash in on the AI training and inference boom. There are multiple ways neoclouds are approaching the opportunity. Infrastructure providers like TeraWulf focus strictly on securing power and building the physical data halls. Pure-play neoclouds are gobbling up chips, largely from Nvidia, to sell GPU compute itself. Meanwhile, players like Crusoe are vertically integrated and doing both.
Across the more than 100 neoclouds that have popped up in recent years, the vast majority are “very early-stage or GPU resellers,” rather than companies that actually own and operate data center infrastructure at scale, according to research by Sightline Climate. The majority probably won’t be able to manage AI workloads on their own because they don’t have access to power, the latest chips, or a managerial team with experience running data centers, industry analysts told Latitude Media — all of which are key to attracting major hyperscalers or AI labs as tenants.
But some can. And that group is reshaping the grid by repurposing old industrial sites or embracing the bring-your-own-power movement to sidestep lengthy wait times to interconnect with the larger system.
“These neoclouds can be ex-crypto miners or even a tennis shoe company like Allbirds that’s decided that they want to get into the data center business,” Chris Campbell, senior director of AI solutions at World Wide Technology, which builds data centers and advises neocloud companies including CoreWeave, Lambda, Nebius, and Vulture. “It’s a place where you can raise capital today.”
Campbell added that many investors want to see a power contract before providing capital to a neocloud, since securing one can be a bottleneck. But some neoclouds are doing early-stage fundraising just to go and secure land with utility agreements, particularly in Texas, where energy generation is abundant.
TeraWulf, for its part, has acquired former industrial sites in regions like western New York — where the grid mix is about 90% zero-carbon, including from both nuclear and hydroelectric power from Niagara — as well as in coal- and gas-rich Kentucky.
TeraWulf, which markets itself as prioritizing low-carbon projects, in February announced plans to turn a former aluminum processing site in Kentucky into a 480-MW data center campus; local residents promptly circulated a petition to pause the project until the impacts on the workforce, energy, and the environment are clearer. The company this week also acquired an industrial campus in the eastern part of the state, where the local utility is building a substation to support up to a gigawatt of capacity.
TeraWulf executives told a group of reporters during a tour last week that they aim to energize three gigawatts of AI data centers within the next three years. (No pictures are allowed of the highly secured site, where a guard checks IDs at the gate.)
The cost-benefit ratio is clear. Sean Farrell, TeraWulf’s COO, said the capex of building its campus north of Buffalo, known as Lake Mariner, will be between $7 million and $10 million per MW, or between $5.25 billion and $7.5 billion total. But the company could earn about $1.7 million per MW in revenue each year — or more than $19 billion total over a single 15-year contract.
A year ago, TeraWulf had 18 MW of high-performance computing capacity as a proof of concept for hosting AI workloads. Then the company acquired the infrastructure holding company Beowulf Electricity & Data in May 2025, which gave it expertise across power infrastructure, site operations, and engineering, according to an SEC filing. Then in August, it announced a $3.7-billion, 10-year hosting deal at Lake Mariner backed by Google, which took a stake in the company. This year, TeraWulf has also closed a $1 billion public stock offering to help fund its expansion into Kentucky.
In the world of neoclouds, lining up power, hyperscaler tenants with access to chips, and cash are all good signs.
Some of TeraWulf’s competitors include CoreWeave, which locked in large power contracts across multiple U.S. states years ago and counts OpenAI and Microsoft among its anchor tenants, as well as Nebius and Nscale. Nscale — a two-year-old startup — secured access to cheap Nordic hydropower early on and is expanding into the U.S., including under a deal with Microsoft in West Virginia for an off-grid, gas-powered campus. The latest contract between Nscale and Microsoft in March comes with a guarantee of 1.35 GW of Nvidia’s Vera Rubin GPUs; a trio of Sightline researchers — Olivia Wang, Declan McCarthy, and Paola Saenz — said that guarantee suggests “real supplier relationships and the financial credibility to back large orders.”
Meanwhile, Blackstone and Google last week announced their own neocloud joint venture, offering Google’s AI chips for compute-as-a-service. The Sightline team said that helps validate the neocloud model because Google and Blackstone are essentially building their own neocloud to compete in the same enterprise AI compute market that CoreWeave and Nebius pioneered.
However, Campbell said the majority of neoclouds won’t be able to manage modern, high-performance AI workloads. “A lot of ex-crypto miners or investment bankers with access to money who want in don’t understand how hard it is,” he said. “Do you know how to build and run a brand new data center with all the latest and greatest liquid cooling? That’s a question for many of these founders, and they haven’t really thought about it.”
Campbell predicted neoclouds that don’t end up delivering AI compute at scale, but have some physical assets like IT equipment or powered land, will instead become acquisition targets. In five years, he added, the winners and losers will come into view.
Editor’s note: The reporting for this story was made possible in part by Schneider Electric, which facilitated travel and site visits to TeraWulf’s campus in western New York.
