TeraWulf Inc. WULF

Key data

The business

TeraWulf runs Lake Mariner, a data center campus on a former power-plant site in western New York, wired to grid power that is cheap and largely carbon-free. Until 2025 it did one thing: run rows of specialized computers around the clock guessing numbers to earn bitcoin, a business whose revenue swings with the coin price and gets harder every time the network grows. The pivot is to rent that same power and land to companies that need somewhere to plug in Nvidia chips for AI. TeraWulf signed multi-year hosting deals, the marquee one with Fluidstack carrying a payment guarantee from Google, and is racing to pour concrete and hang transformers before the tenants arrive. The moat, if there is one, is the power: interconnected capacity at a low rate is scarce, and a signed tenant with a Google backstop is hard to dislodge. What you hold today is mostly a construction site with a bitcoin mine attached.

The numbers

Two things are true at once: the top line is stalling, and the bottom line is a horror show that mostly is not real.

Revenue in the first half of 2026 was $78.8M, below the $82.0M of the same half in 2025, because bitcoin economics tightened and the AI rent has barely started flowing. The huge net losses, $940M last quarter, are not operating losses: the campus lost about $140M running, and the rest is a paper mark on the value of the convertible debt and warrants issued to fund the buildout. The mining tells the real operating story, gross margin at 57% in Q3 2025 collapsing to 20% by Q2 2026 as difficulty and power costs bite.

Here is the whole wager in one line of arithmetic. Revenue ran at about $168M in 2025. Analysts model $1.6 billion by 2028, roughly ten times more, and the first small profit that same year. To get there, TeraWulf turned zero long-term debt into $3.1 billion in late 2025 and is spending it at a billion-a-year clip on buildings that sit empty until tenants arrive.

Cash was still $2.62B at June. What this memo believes that the price does not: the gap between an empty building and a paying tenant is measured in quarters of cash burn, and the only thing that closes it is contracted hosting revenue landing in the revenue line, not another deal announcement.

Management

The record here is one signal, loud. Over the last year insiders bought about $150,000 across four small purchases and sold about $25 million across fourteen, and the seller is overwhelmingly founder Paul Prager, who sold more than $10 million in three lots in May and June alone. Plan status is not disclosed in the filings, so whether these were pre-scheduled is unknown. Either way, the founder took real money off the table into the same buildout he is asking shareholders to finance. Stock-based compensation runs at 95% of trailing revenue, and share count keeps climbing as the company issues equity and convertibles. The capital allocation is all-in on one construction bet, funded by debt and dilution, while the man running it sells.

How it fails or surprises you

The tenants slip or the money runs short. The buildings earn nothing until customers move in, and the buildout is debt-funded. If a marquee tenant delays its ramp two or three quarters, or bitcoin weakens while $3.1 billion of debt accrues interest, the $2.62 billion cash pile drains toward a raise on bad terms. Watch cash: another $600M quarterly drop like Q1 into Q2 is the tell.

The profit the model needs has never existed here. This company lost money every year in the record, $661M in 2025, and its mining margin just fell to 20%. The case that 2028 brings $1.6 billion of revenue and a profit rests entirely on contracts converting cleanly. Nothing in eight years of filings shows this team earning a sustained operating dollar. If hosting revenue is not clearly in the P&L by mid-2027, the read is wrong.

The Google-backed AI rent arrives bigger and sooner (right tail). If Lake Mariner fills faster than modeled and the Fluidstack-type deals convert to recurring revenue at the contracted rates, TeraWulf re-rates from a bitcoin miner priced on hope to a power-advantaged AI landlord priced on cash flow. The market pays for none of this today because none of it is in the revenue line yet. The first two clean quarters of hosting revenue would reveal it.

Closing thoughts

One number settles most of this: contracted HPC hosting revenue actually appearing in the income statement, at the rates the signed deals imply. An ambiguous print, another deal announced with no revenue booked and another quarter of cash draining, leaves the reader waiting and watching the cash balance. A clean print, hosting revenue ramping toward the run-rate the 2028 model needs, converts this from faith to arithmetic. The fatter tail is genuinely uncertain: the power and the Google-backed contract are real assets, but the founder is selling, the debt is large, and the margin on what it does today is thin and falling. What is at risk if the tenants slip is a dilutive raise into a weak stock. What the upside is worth is a re-rating that makes 2028 numbers look cheap.

The bet is still that TeraWulf becomes a landlord for artificial-intelligence computers, renting buildings full of chips next to cheap power in upstate New York, and that the giant customers it has signed, one backed by Google, pay for years. It breaks if the buildings stay empty longer than the balance sheet can wait. The one pair of numbers that tells you first: hosting revenue landing in the top line against the cash balance falling, quarter by quarter. Until contracted AI revenue is a real line item and the cash has stopped bleeding, you are financing a construction site next to a founder heading for the exit.