Download the research and models (ZIP) ↓

Suppose Anthropic stops paying while one data hall in a Kentucky building is still unfinished. Should the developer put another dollar into the job? At TeraWulf’s Hawesville campus there is an unusually good reason to consider it. Finishing the hall can give the landlord something it did not have yesterday: a contractual route to Google if the default persists. Walk away one hall short and that route stays shut. The most valuable room in the building may be the last one delivered.

The parties designed it that way. Anthropic owes rent as each hall enters service. Google LLC’s default remedies take effect when the last hall under that particular building’s lease reaches its commencement date. Each of the three buildings has its own agreement. One can have paying tenants and an effective backstop while the next is still pouring concrete. The roughly $19 billion in TeraWulf’s headline is contracted lease revenue over the initial term. The construction lender’s exposure changes several times on the way to collecting it. November 4 filing, item 1.01 · July lease announcement.

Three building targets and the rent earned before each lease’s last hall starts. Two equal delays leave seven units of rent each, but one versus three units fall before activation.

The first lender already came and went

In March, TeraWulf arranged a $500 million secured bridge that it could use to develop the campus. It drew $100 million, at term SOFR plus 275 basis points. On April 16 it paid the principal back, paid another $0.4 million of accrued interest and agency fees, and terminated the facility. Upfront financing costs had been $7.3 million; writing off the unamortized balance produced a $7.1 million accounting loss. Adding the loss to the upfront costs would count much of the same expense twice. The useful distinction is economic: TeraWulf paid to have $500 million available even though it used only a fifth. September 10-Q, “Bridge Credit Facility”.

That loan was gone before the July lease announcement and the November Google agreement. An investor building a sources-and-uses schedule today should leave the old bridge out of available capacity. Fluor’s contract now covers engineering through testing and commissioning, with milestones, open-book reporting and procurement oversight. TeraWulf said on November 4 that guaranteed maximum pricing was being established as the work advanced. For the next lender, two things therefore matter together: the support available once the building qualifies, and the cost still to be funded to get it there. Fluor agreement announcement.

Two delays, the same rent, a different credit

TeraWulf targets 137.6 MW in Q4 2027 for CB-1, 137.6 MW in Q1 2028 for CB-2 and 125.6 MW in Q2 2028 for CB-3. Those are building delivery targets; the last hall’s actual commencement date is the contractual milestone. The quarterly filing adds a useful wrinkle: the phases of each lease expire together, 20 years after the final phase starts. A delayed rent receipt in the next five months and a dollar lost over the lifetime of the lease are different calculations. Investor presentation, slide 8 · 10-Q, note 8.

Take one hypothetical building with two equal halls. Each earns one unit per month, starting at the beginning of its delivery month. Anthropic pays in full. In the base case the halls start in months one and three. Change just one delivery date, hold everything else fixed, and tally the first five months:

Which hall slips? First / last start Rent due over five months Of which due before Google’s remedies activate
Neither M1 / M3 8 2
First hall slips a month M2 / M3 7 1
Last hall slips a month M1 / M4 7 3

The two delay cases produce exactly the same rent. Yet in one, three units come due while the landlord is still waiting for Google’s remedies; in the other, just one does. Google pays zero in all three cases because the tenant performs. The last column measures the timing of exposure, not a claim or an expected loss. With five halls already earning rent, each extra month waiting for the sixth adds a month of exposure on all five. A portfolio dashboard counting only delivered megawatts misses the trade. Download the inputs and monthly results.

Why Google cares which chips are installed

If a qualifying default happens early, Raylan, the landlord, retains the right to finish the premises. Google can check that commencement conditions have been satisfied. If the qualifying default is still continuing when the last hall starts, Raylan can then serve notice. This is a reason to preserve completion funding even after a customer weakens: putting fresh money into an unfinished site can change the recovery on the money already sunk. Whether to advance it turns on remaining cost, time, the chance of qualifying and the value after completion. Credit support agreement, §4(a).

After activation, Google can step into the lease and cover arrears. For a qualifying rent default it can also postpone its decision while funding the unpaid rent. The other route, ending the lease for a scheduled fee, is normally barred for the first six years. The exception turns on the installed TPU pods: could they, when operating, consume less than half of the building’s contractual critical power? A site running lightly today can have exactly the same installed capability tomorrow. What matters to this test is what the installed machines could draw. Our reading: Google has accepted a longer tie to a building well equipped for its own computing platform; the lender has a reason to examine the equipment schedule alongside the lease. Credit support agreement, §§4(c)–(h).

The termination fee points to the intended beneficiary. Its declining schedule was designed around anticipated project debt and can be revised downward to reflect actual borrowing; the lender arrangement routes payment to a designated account. If Google pays the fee it can elect a stake in Raylan tied to the fee relative to development cost. The public form omits the dollar cap. For a recovery model, the right inputs are the fee schedule and outstanding principal on the same date. Twenty years of gross rent belongs on a different line. Contract, §§4(d)–(e).

What we would finance first

A sponsor choosing between a nearly finished building and the first hall of the next one faces a real trade. Opening the new hall starts another rent stream. Finishing the old one can start rent and change the credit behind its neighbors. If expected rent and construction cost are otherwise similar and the sponsor faces a binding equity budget, we would favor finishing the first building. If the remaining work depends on an uncertain power upgrade, or the next hall earns materially more and can open quickly, the order can reverse.

The lender’s version of that decision is to reserve enough money to finish a particular building, then revisit the advance rate when the last hall qualifies. For these three buildings we would ask for the cost to reach each commencement date, the rent already running ahead of that date, and evidence of the installed TPU capacity. Those three rows tell a credit committee more about the next dollar than a campus-wide percentage-complete figure. The next financing disclosure worth reading is the one that shows whether the people writing the checks made the same distinction.

Who writes the last $545,455?

The landlord owns the right to finish. A new lender still has to like the price. Consider a hypothetical workout, in millions of dollars. The unfinished property would sell for $63m today. Spend another $10m and, one year from now, it distributes $100m if the building qualifies or $4m if it falls short. Give completion an 80% probability. At a common 10% hurdle, the expected $80.8m is worth $73.455m today. After the $10m outlay, finishing beats selling by about $455,000. These are illustrative underwriting inputs, separate from the Hawesville lease terms.

Now seat two lenders at the table. The incumbent has a $100m claim; selling today puts $63m in its pocket. The newcomer wants at least $11m in expected year-end cash for advancing $10m. Under an offer that caps its total claim at $14m, the two negotiate how much gets paid before the incumbent, with the rest paid after it. In this example, a first-priority slice of $12.750m–$13.375m works for both. At the low end the newcomer just earns its hurdle. At the high end the incumbent just matches the sale.

Change one term New lender’s position What can close the gap?
Cap the new claim at $12m; promise first priority on every dollar Gets $12m in success and $4m in failure. Expected recovery: $10.4m, short of $11m. Giving it yet more seniority is impossible.
Keep the $12m cap; cut the outside advance to $9.455m (rounded) The same $10.4m now earns the 10% hurdle on the exact advance of $10.4m / 1.10. The incumbent itself supplies the remaining roughly $545,455 upfront.

Why would the existing lender pay? When the newcomer takes the first $12m, the incumbent receives $88m if the build succeeds and zero if it fails. Its expected $70.4m is worth $64m today. Put in the missing $545,455 and the incumbent keeps roughly $63.455m of net present value, against $63m in a sale. A developer asking an unrelated equity investor for the same check would face a harder pitch: ordinary equity gets nothing in either outcome under this waterfall. The natural source of the last check is the creditor whose existing claim it protects.

Try it: download the editable Excel model. Change C15 from 14 to 12: the consent window disappears. Then set C14 to =10.4/1.1 for the smaller outside advance. C8 now treats the smaller advance as if it were the whole outlay. Deduct the incumbent’s $0.545455m contribution: project NPV stays at $0.454545m. Compare the incumbent’s net recovery against the sale after that same contribution, as above. The calculation and exact inputs also separate cash invested by each party.

When the facts change, the thesis should too.Get the research ↗