Power, priced by tenor

On 2026-07-06, TeraWulf announced a 20-year lease with Anthropic at its Justified Data campus in Hawesville, Kentucky: about 401 MW of critical IT load, roughly $19B of contracted revenue over the initial term, first capacity in the second half of 2027 and the full 401 MW by early 2028, with investment-grade credit support expected behind it. Divide the one number by the others and the lease clears at ≈$270 per IT-megawatt-hour, or ≈$197 per kilowatt-month. Two weeks earlier, on 2026-06-22, Bloomberg, CNBC and Axios had reported a SpaceX–Reflection AI agreement: $150M a month from 2026-07-01 through 2029, up to $6.3B, for NVIDIA GB300 capacity at Colossus 2 outside Memphis, with either party free to walk on 90 days’ notice after an initial three-month commitment.

These are not the same product, and the difference is the point. TeraWulf sells a powered campus; on Somani’s reading the tenant brings the GPUs, though the announcement itself is silent on who supplies them. SpaceX, as reported, sells the compute itself; the GPUs are in the price. Neel Somani, a former hedge-fund power quant whose free primer Power 2026 is the occasion for this note, puts the SpaceX rate on the order of $5,000 per megawatt-hour on his own napkin arithmetic. The megawatt allocation behind that figure is not public, so treat the level as an estimate; the ordering is not in doubt. Call it eighteen-fold. Most of that gap is silicon rent, not electricity, and with the allocation unpublished the two cannot be cleanly separated. So the two prints do not draw one curve for one product. What they show is the two ends of the market, and the variable that divides them on the other axis. The cheap contract runs twenty years. The dear one can end in ninety days.

Consider what has to be built to deliver the cheap end. A behind-the-meter gas plant for a compute campus runs on the order of $300M; construction debt prices around SOFR plus 225 basis points for an established developer with a prime tenant, and the lender wants years of contracted cash flow before it funds. Homer City in Pennsylvania, the 2 GW coal site shut in 2023, is being rebuilt as 4.4 GW of gas at roughly $10B on the strength of exactly this kind of offtake. Twenty years underwrites that. Ninety days underwrites nothing; it is an option on capacity, and options on scarce capacity are expensive. That is the sense in which tenor prices power: not as a coefficient on one curve, but as the condition under which a megawatt gets financed and built at cost-plus rather than rented at scarcity. Sign long and the plant gets built for you. Sign short and you pay for someone else’s.

The gap is wide because supply cannot answer on the buyer’s clock. The three makers of heavy gas turbines — GE Vernova, Siemens Energy, Mitsubishi Power — are sold out, with backlogs running from months to years; the relief valves in use are repurposed jet engines and Chinese imports, which is what a binding constraint looks like when it is being routed around at a premium. Interconnection queues run years. Homer City’s air permit arrived in November 2025 for a site that has been wired to the grid for half a century. Against that, data centers already draw about 5% of US power and the load is doubling every two years or so, on Somani’s figures. While the constraint binds, the premium for short tenor stays wide, and every buyer who cannot sign for twenty years pays for the ones who can.

What would narrow it is worth writing down in advance. First, the turbine backlogs: not the level but the second derivative, the quarter in which the queue stops lengthening. Second, the power-only rate and tenor of each new compute lease; if the next twenty-year print comes in well under $270, or the next short-tenor deal comes in far under Somani’s figure, the scarcity rent is compressing. Third, the policy layer. A non-binding Ratepayer Protection Pledge, a NERC warning that data centers can lift capacity prices, a city (Monterey Park) that has already voted its data centers out; a rate-class ruling reprices this complex faster than any turbine delivery. And one read-through we watch for its own sake: a bitcoin miner’s legacy power contract at $30–50 per MWh, marked against a $270 twenty-year AI lease. TeraWulf was a miner. That spread is the conversion thesis, and it is now a print rather than a projection.

We filed the power argument in May and the term-sheet reading in August. This is the same thread pulled one step further. The open leg in the Volta structure was duration: revenue for six years against a lease for sixteen. Here duration is what decides which end of the market a buyer is standing at.

Twenty years buys a megawatt near cost-plus; ninety days rents compute at scarcity, silicon included. Two products, one dividing variable. Not a position; two prints worth filing.

Sources.The TeraWulf–Anthropic terms are from TeraWulf’s announcement of 2026-07-06, which does not state who supplies the GPUs; the SpaceX–Reflection terms are as reported by Bloomberg, CNBC and Axios on 2026-06-22; the Volta–Bitdeer rate is from Bitdeer’s announcement of 2026-08-04. Wholesale ranges, plant costs and financing spreads, the turbine backlog, the Homer City figures, the data-center share of US load and the ≈$5,000 estimate are from Neel Somani, Power 2026: Electricity Pricing in the Age of AI (power2026.ai), whose primary sources are EIA, FERC and ISO publications. Per-MWh and per-kW-month figures are our own arithmetic on those disclosures.