By design, datacenters are big capital machines. But to get a handle on the scale of resources that AI and HPC are sucking in out of the global economy, just consider what is happening at TeraWulf’s Lake Mariner site on the shore of Lake Ontario, just outside Buffalo, New York.

In mid 2022, TeraWulf was pleased to announce it was in the final steps of firing up 50 megawatts of Bitcoin mining capacity at the site – a former coal-powered power station – 18 months after first launching. Another 50 megawatts of capacity was scheduled to switch on in early 2023.

By 2025, following a 2 megawatt AI/GPU pilot at the site, TeraWulf had pivoted, turning itself into an HPC/AI company, expanding its footprint at the site to 157 acres and aiming to boost its capacity at the Lake Mariner site to 750 megawatts. It will still produce Bitcoin on an opportunistic basis. But it is clear the management and shareholders have decided there’s a better opportunity to make real money by providing the infrastructure to allow other people to produce AI tokens.

The Next Platform took a look around the site recently, along with Schneider Electric. The French power kit firm has supplied much of the electrical infrastructure at the site, while its Motivair subsidiary has provided much of the liquid cooling technology that is now a prerequisite for an AI-grade site.

Once it switched its attention to HPC and AI, TeraWulf’s initial effort was the CB-1 datacenter, a 20 megawatt facility, with the 50 megawatt CB-2 datacenter slated for 2025. Sovereign AI specialist Core42, which is a strong partner of Cerebras Systems and AMD as well as other AI system suppliers, was the firm’s first banner client. It has been running AMD-based systems in CB-1 for the last ten months.

AI infrastructure specialist Fluidstack – whose operations are backed by Google and who are helping Anthropic create and install its own TPU systems and which is also a backer of TeraWulf – went into production at the site’s CB-3 building a few weeks before we visited. So that was off limits. (It may be where Anthropic has parked its initial TPU systems for all we know.)

But we did see the CB-4 building. This is a 330,000 square foot, 200 megawatt beast, dwarfing the earlier operations. It encompasses four data halls, each spanning 33,000 square feet, illustrating just how much space, power, and mechanical infrastructure takes up. Construction here began in January. The building is in the final states of preparation ahead of flicking the power switch towards the end of the summer.

Once the buildings are swept clean and powered up, it will be down to the firm’s customers to install racks for whatever compute they want to run and decide on exactly how they want to manage the space.

Sean Farrell, chief operating officer for TeraWulf, said the halls were built with slab concrete floors. These are easier to build than raised floors, but more importantly can support increasingly dense, and therefore heavy, racks.

High-density AI capable racks, which are packing liquid cooling kit, and potentially 800 volt infrastructure, are getting too heavy for raised floors. The company has built with a spec of 8,000 pounds per rack in mind and could support up to 10,000 lbs.

Farrell described the non-compute mechanical rooms supporting this building as “massive.” And yes, they are pretty enormous. The data halls themselves account for less than half the total footprint. The mechanical rooms handle a closed loop cooling system, with chillers and coolers in the data hall capturing the heat generated by the GPU heavy racks, before heat exchangers exhaust it out via roof top towers. The halls have space for fan walls should customers want them.

No water is used in normal cooling operations, Farrell said, and once “charged” the fluid in the cooling system lasts ten to fifteen years.

The company has already spent $290 million with Schneider and Motivair across UPS and batteries, CDUs, in-rack manifolds, racks, rack coolers, and software and services.

The CB-5 building next door is also 330,000 square feet. As Farrell put it, the first steel in that building went in on April 1. The roof is now on, and while the sides are open, the roof top coolers are being installed. Farrell said the building should be “energized” by the end of the year. Last year, TeraWulf announced Fluidstack had contracted for a 160 megawatt lease on CB-5, bringing its total commitment on the site to 360 megawatts.

What the entire site doesn’t have is diesel generators for backup. Oil burners are traditionally the mark of a serious datacenter – and along with water-based cooling systems, and a thirst for grid electricity, they are a key reason for communities objecting to datacenter applications.

Lake Mariner benefits from dual 345 kilovolt power feeds coming in from separate grids. This means it has Tier 3 equivalence, which means it has N+1 redundancy and multiple power and cooling distribution paths. The grid power mix is 89 percent zero carbon – in large part because the New York Independent System Operator taps into the hydropower resources of Niagara Falls.

Power is non-negotiable for a datacenter development. TeraWulf overall has 3 gigawatts of capacity across its various sites, including a former aluminum plant in Kentucky, which has an active substation rated at 480 megawatts.

Connectivity is also essential – though there is a little more wriggle room when it comes to latency now, Farrell explained. Five years ago, he said, the rule of thumb was that a site had to be two hours away from a football stadium – no self-respecting sports arena is anything less than super connected. But where sites were focused on training, latency for returns was not quite so critical. And building out connectivity is easier than it was, he said. So, at the outset at least, super low latency is no longer a deal breaker

That said, the Lake Mariner site is about one hour away from the Buffalo Bills’ stadium, which delivers another benefit. The completion of a massive refurb at the Bills has freed up hundreds of electrical contractors. And it is trades, specifically electricians, which are the biggest bottle neck for datacenter projects, said Farrell.

The build, which is operating 24x7, keeps around 1,800 tradespeople occupied at any one time, of which 650 to 800 are electricians.

That’s a massive investment in hours and money, and from what we saw, hard hats and hi-viz vests. Building for bitcoin mining – with systems that rely on air cooling – costs around $500,000 per megawatt. The compute on top of that is roughly the same.

For AI/HPC, Farrell said, the build cost for the infrastructure works out around $7 million to $10 million per megawatt. But that presupposes liquid cooling, amongst other things. “There’s a lot more that goes into it,” says Farrell. Or into those massive mechanical rooms.

And, as a colocation site, it’s the customers who will be kitting out and managing the actual compute.

How much will they be paying TeraWulf for datacenter space? The average datacenter rental cost in the United States is around $140 per kilowatt per month. And contracts are typically for ten to fifteen years. What’s that all add up to? You can work it out, but best to use a calculator. GenAI is not great with math.