TeraWulf's $940mn Loss Is Mostly an Illusion, Its Real Bet Is in Kentucky

The company reported $45mn in quarterly revenue, with HPC (high-performance computing) lease revenue rising 52% quarter-on-quarter to $32mn, or 71% of the total. Bitcoin mining, once the core business, contributed less than one-third as TeraWulf continued converting its power portfolio towards longer-term AI contracts. TeraWulf now has 102 MW of revenue-generating computing capacity operating at its Lake Mariner campus in New York, with another 336 MW under construction across two data halls, CB-4 and CB-5, targeted to begin generating lease revenue in late September and very early January respectively.

Power becomes the product

TeraWulf's strategy increasingly rests on controlling scarce grid connections rather than owning mining machines. Years spent securing low-cost electricity for Bitcoin have left the company with sites that can be repurposed for the energy-intensive demands of AI, the same dynamic that separated winners from losers across the mining sector this earnings season: it was reported this morning that investors are now pricing Bitcoin miners on their AI and power infrastructure rather than mining output alone. TeraWulf's quarter is a third data point in that same repricing.

CFO Patrick Fleury put a number on the scale of the problem TeraWulf is selling into: Texas alone has a grid connection backlog of more than 400 GW against roughly 60 GW actually cleared for the state's "batch zero" fast-track process. "You basically have to build another California in the Texas grid to meet all of the batch zero projects," he told analysts, noting combined-cycle gas plants take three years each to build, and the industry needs hundreds of them. That scarcity is the entire premise behind TeraWulf's model: whoever already controls power gets to set terms with whoever needs it.

What the Anthropic deal is actually worth, and what it isn't

That model expanded beyond New York during the quarter. TeraWulf acquired the Muskie campus in Eastern Kentucky, which has agreements supporting up to 1 GW of power, and signed a 20-year lease with Claude AI model-maker Anthropic covering approximately 401 MW at the Justified Data Campus in Hawesville, Kentucky. Fleury told analysts the deal represents a 20% internal rate of return relative to what TeraWulf spent developing the site, calling it justified proof "that we can take a legacy industrial power site and convert it into a large-scale operating high-power compute campus." The agreement represents roughly $19bn in contracted revenue before extension options, or as much as $33bn if both five-year options are exercised. Deliveries are not expected to begin until the second half of 2027, meaning the value still depends on construction, financing and timely customer acceptance.

Fleury was explicit on the call about why TeraWulf structures these deals differently from some competitors: "If your WAC [weighted average cost of capital: a blended rate combining what a company pays for its debt and what it needs to return to equity investors, weighted by how much of each it uses to fund a project]is not below your yield on cost, by definition, there is no equity value... I've seen some folks in our market signing deals that appear where their WAC is actually in excess of the yield they're earning on the lease. That doesn't work. That's kind of called real estate, where you buy something and pray that in 10 or 20 years it's worth more. That's not the game we're in."

The loss the headline number hides

TeraWulf's GAAP net loss for the quarter was $940mn, against $18.4mn a year earlier. That figure is dramatic enough to deserve its own explanation, and the mechanism behind it is almost entirely disconnected from how the business is actually performing: $756mn of the loss came from revaluing warrants issued to Google as part of its credit-support arrangement backing tenant Fluidstack's lease obligations, a non-cash charge that grows larger as TeraWulf's own share price rises. Fleury confirmed the change "had no impact on our liquidity." The stock is up roughly 48% this year, which means the AI-pivot rally that has excited investors is the same rally mechanically inflating the loss that will lead most headlines. On an adjusted basis, which strips out the warrant charge along with stock compensation and other non-cash items, EBITDA was negative $18mn, a much smaller swing from the positive $15mn TeraWulf posted in the same quarter last year.

The margin number that also needs unpacking

A second figure in the release looks worse than it is for a related reason. Reported HPC leasing segment profit margin was approximately 28% for the quarter, well below TeraWulf's long-term target of roughly 85%. But that figure includes one-off tenant fit-out costs, pre-revenue operating expenses at sites not yet generating rent, and development spending across the uncontracted portfolio; stripped of those, Fleury said, the underlying margin was closer to 80%, in line with where the business is meant to land once capacity matures. The gap between 28% and 80% is almost entirely a function of TeraWulf building three data halls at once rather than one at a time.

Part of that one-off spending stems from a July agreement with tenant Fluidstack to amend the Lake Mariner leases: TeraWulf will contribute roughly $150mn towards tenant fit-out costs, in exchange for more than $300mn of incremental rent over the leases' remaining terms. Combined with a separate capacity increase across CB-4 and CB-5, Fleury said the amendments are expected to generate over $500mn of additional lease revenue in total, at what he described as a mid-teens return, funded with equity rather than debt by design.

Betting the model works twice, and eyeing a third continent

Chairman and CEO Paul Prager framed the Kentucky expansion as proof the Lake Mariner playbook travels, and offered an unusually vivid explanation for why TeraWulf refuses to concentrate its buildout in one place. "The Russians would put all their major equipment and machinery in one part of the boat, so if you hit that part of the boat, they were dead in the water," he said, recalling his Navy background. "American submarine design was to put major machinery throughout the boat, so we had the ability to sustain ourselves even if we took hits in one area." He said customers are increasingly sensitive to the same logic: nobody wants a data centre wholly dependent on one grid, one regulator or one fuel source.

That same call carried two pieces of regulatory news not in the press release. New York Governor Kathy Hochul issued an executive order affecting data centre development in the state; Prager said TeraWulf does not expect it to disrupt timing at Lake Mariner or Lake Hawkeye, since the former is already permitted and the latter remains in early-stage development. Separately, in response to an analyst question, Prager confirmed one of the three to five sites TeraWulf currently has in active diligence would mark the company's first international data centre, with Northern Europe under consideration.

Execution replaces mining risk

TeraWulf is also selling its 50.1% stake in the Abernathy joint venture for approximately $530mn, a transaction Fleury said represents a 20% internal rate of return on the original investment. The proceeds arrive in three tranches rather than all at once: $250mn received in July, roughly $150mn expected by the end of this year, and about $130mn by April 2027. "Our objective is not simply to accumulate megawatts," Prager said. "It is to build a durable, capital-efficient platform that compounds value for shareholders."

The investment case has nevertheless changed. TeraWulf is becoming less exposed to Bitcoin's price, and more exposed to whether it can deliver complex data centres on schedule, in a labour market where the company says it had roughly 1,000 electricians on site at peak and views skilled-labour availability, not capital, as the binding constraint on how fast it can grow. Its next test is not how efficiently it mines coins. It is whether a large pipeline of power and leases becomes operating infrastructure on the timeline it has promised.