The second-quarter numbers arrive like a bear trap. TeraWulf's Bitcoin mining revenue collapsed 73% year-over-year to $12.8 million. Net loss hit $940.8 million in a single quarter. Year-to-date losses stack past $1.4 billion. A surface read says liquidation.
Then reverse the stack. HPC/AI leasing contributed $31.9 million — 71% of total revenue. Of that $940.8 million "loss," $755.7 million is a non-cash warrant liability remeasurement: an accounting artifact tied to prior equity-linked financing, not cash burning. Total quarterly revenue of $44.8 million is down roughly 6% — contraction, yes, but nowhere near the carnage the headline loss implies.
Reversing the stack to find the original intent: TeraWulf is no longer a Bitcoin mining company. It is a data center landlord in gestation, holding a $19 billion Anthropic lease that can expand to $33 billion, with a 20-year term and a five-year renewal option. The market, however, still prices it as a crypto proxy. That disconnect is the story.
TeraWulf operates two core physical assets. Lake Mariner in upstate New York is the anchor: 102MW of critical IT capacity already operational and generating revenue, with portions of the facility actively being converted from mining to high-performance compute. Another 336MW is under construction across the portfolio. The Kentucky campus, codenamed Justified, is designated for the Anthropic buildout — a planned 401MW of dedicated capacity.
The operating logic is simple if you understand where AI infrastructure actually bottlenecks. It is not chip supply. It is not software talent. It is electricity. CEO Paul Prager frames it directly: control of power infrastructure is where value accrues. Data centers take two to four years to build. Grid interconnection rights take almost as long and are increasingly the binding constraint in every major US market. TeraWulf's real assets are not GPUs or ASICs. They are grid access, land parcels, and power purchase agreements.
The counterparties confirm the shift. Fluidstack, an AI cloud intermediary, signed on and activated $600 million in Google credit support — effectively a Big Tech validation of TeraWulf's delivery capability. Anthropic then committed to the largest lease in the company's history. This is not Bitcoin price exposure. It is contracted AI infrastructure revenue. Management's language has shifted accordingly: the transition from "platform building" to "scaled execution" is a tell. The land-acquisition and power-securing phase is being treated as complete. The new game is delivery.
Prager's background is energy, not crypto. That matters. His emphasis on "realizing value where appropriate and redeploying capital" reads like an investor in power assets, not a mining maximalist. A capital allocator with that framing sells or retires non-core assets to fund the AI buildout. Expect asset sales and project-level financing structures in the coming quarters. The company has signaled it will selectively pursue only projects with confirmed power and confirmed demand — a maturity that separates it from the speculative expansion stories still dominant in the mining sector.
The Revenue Switch Is Real. The Cash Flow Switch Is Not.
The revenue composition says the transition is complete on paper: 71% HPC/AI leasing, 29% mining — and the mining share is actively shrinking as Lake Mariner is converted away from ASICs. But cash-flow timing tells a different story. The Anthropic lease begins contributing rent only in the second half of 2027. The 336MW under construction requires capital before that revenue stream turns on. Between now and late 2027, TeraWulf must fund construction, interest, and operations from 102MW of AI capacity plus whatever mining margin survives.
This is the single most important structural risk in the equity. The company will need fresh capital before the big cash begins flowing. The $755.7 million warrant remeasurement signals that equity-linked instruments already occupy a meaningful slice of the capital stack. If the next financing comes as equity, shareholders absorb dilution ahead of the revenue ramp. If it comes as project debt collateralized by the Anthropic lease, the cost of capital depends entirely on a rate environment that remains stubbornly high. Based on my audit experience, I have seen this exact pattern in infrastructure businesses: a signed contract extends visibility while the funding gap quietly compounds. Truth is not consensus; truth is verifiable code. The code here is the cap table — and it is not fully visible.
The implied operating loss — strip the $755.7 million warrant charge from the $940.8 million net loss — lands near $185 million for the quarter. That is still real burn, but the trajectory is a construction-phase company funding growth, not a treasury draining. Adjusted EBITDA and operating cash flow would tell the true story, but management has not disclosed them in the parsed release. The absence is notable. Until those numbers are published, the gap between GAAP loss and cash position remains an open variable that short-sellers can exploit and long-term holders must accept.
The Retrofit Engineering Risk Is Underpriced
Here is where my background pushes me to be blunt. I have spent years auditing infrastructure and protocol layers where the gap between "signed contract" and "delivered capability" is the whole game. Converting a Bitcoin mine into an AI data center is not a cosmetic upgrade.
Bitcoin miners are remarkably tolerant machines. They handle latency without penalty, tolerate mediocre PUE ratios, and have modest power density demands relative to modern AI racks. AI workloads are the opposite: 100kW+ per rack, liquid cooling loops, N+1 redundancy, ultra-low-latency fiber meshes. The electrical distribution systems built for ASIC rows — medium-voltage switchgear, transformer configurations, cooling loops — were never designed for that load profile.
Power density is the metric to track. A mining hall running 20-30kW per rack cannot be repurposed without re-engineering the entire electrical spine. Rack re-spacing, transformer re-specification, and cooling plant replacement are not capital-light projects. They are full-scale construction, disguised as "conversion."
The counterintuitive conclusion: retrofitting can be slower and more expensive than greenfield construction. The market assumes "we already have the power hookup, just swap the machines." That assumption fails at the level of electrical engineering. TeraWulf's 102MW of delivered critical IT capacity is a real proof point, but it is small proof. The true test is whether the Lake Mariner conversion sustains the SLA standards Anthropic and Fluidstack will demand — and whether the 336MW expansion hits its schedule. The market has never watched this management team operate at hyperscale AI standards.
The Valuation Framework Hasn't Caught Up
Traditional miner valuation models multiply Bitcoin price, hashrate, and efficiency margins. That framework is obsolete for TeraWulf. The AI transition demands a real-estate lens: discount long-dated lease cash flows, add land and interconnection residual value, and price the whole thing like a data center REIT or a contracted utility. EV/EBITDA and adjusted Funds From Operations replace the old hashprice math.
But the market's anchor is still "miner." GAAP losses will look ugly for multiple quarters because of warrant remeasurements and accelerated depreciation on converted facilities. An investor anchored to the old framework reads those losses as distress. An investor running the REIT framework reads the contracted backlog, the Google credit support, and the 20-year duration as visibility no traditional miner can match. That disagreement — not the earnings number — is the actual trade.
Competitive context sharpens it. Core Scientific is further along in total contracted AI wattage. CoreWeave is the purer AI infrastructure operator. But among US-listed former Bitcoin miners, TeraWulf's 71% AI revenue share is near the top of the sector. IREN is still hunting flagship tenants; Cipher's pipeline remains in negotiation. TeraWulf has already secured the two-tier endorsement: an AI cloud intermediary and a frontier lab. Execution gaps will be punished, but the commercial proof is on the table.
The Contrarian Angle: This Is Now Single-Tenant Real Estate
The bullish narrative is "miner pivots to AI and re-rates upward." The contrarian angle cuts the other way: this is now a single-tenant real estate company, and real estate's worst risk is tenant concentration.
Anthropic accounts for the overwhelming majority of forward capacity — 400+MW planned and dedicated. One 20-year lease, with a five-year renewal option. If AI capex cycles down — and I have watched compute-efficiency improvements kill infrastructure demand forecasts in a single model refresh — TeraWulf carries a revenue vacuum without a warm pipeline to fill it. Management's own target of 250-500MW in new signings per year is an admission that concentration must be reduced, and fast.
There is a second trap embedded in the long-term lockup. Lockups are two-way obligations. Anthropic is bound to pay for a specific technology basis. But TeraWulf is equally bound to that same basis for two decades. If AI compute migrates to custom silicon, or if inference efficiency doubles repeatedly through 2029, the power-hungry, GPU-centric campus model could face mid-lease repricing pressure. Abstraction layers hide complexity, but not error. The 20-year lease is the abstraction. The buried assumption is that AI workloads stay GPU-heavy and centralized — an assumption that has historically decayed faster than real estate can amortize.
The market's pricing error works in both directions. If the story holds, the current crypto-proxy multiple undervalues contracted cash flows. If the conversion fails, the REIT premium the bulls want never materializes, and the stock trades back as a low-multiple hard-asset play with shrinking mining revenue. That asymmetric outcome — big upside on execution, slow bleed on failure — is why the next two delivery cycles are the only numbers that matter.
The Watchlist
The next twelve months matter more than the next twelve earnings calls. Watch three signals. First, the refinancing event — its terms will expose the true cost of capital. Second, Lake Mariner's conversion milestones, and whether delivered capacity actually hits SLA commitments. Third, new customer signings that dilute Anthropic concentration.
If TeraWulf delivers on conversion while signing another 250MW+ of capacity, the market will eventually be forced to reprice it as AI infrastructure rather than a mining stock. If engineering slips, the $19 billion lease becomes a footnote on a stranded conversion story. This is no longer a Bitcoin trade. The question is whether the market recognizes that before the delivery schedule forces the issue.