Three layers now surround the chip: capital that funds AI factories, offtakes that turn them into cash flow, and futures that price them. The largest exposures sit with lenders, lessors and guarantors who must mark, hedge and transfer them. That is demand for every line of the Firm.
Six asset managers commit to dedicated pools of capital targeting over $500bn, with final agreements still to be executed. Capital funds AI factories inside special purpose vehicles and debt funds, off the operator’s and the vendor’s balance sheets, and contracted revenue from leases and capacity contracts services the debt.
The vendor positions its systems as offering the lowest token cost, highest revenue and longest life. A common software stack makes capacity fungible across customers and operators, so a defaulted contract can be re-leased, and chips cascade from frontier training to inference and batch work.
Take-or-pay contracts with investment-grade offtakers amortize principal inside the contract term, so residual value is tested only when the offtaker weakens. That single variable is what every compute financing depends on, and the one the Firm prices daily.
| Reference Transaction | Size | Rating | Pricing | Tenor | Credit Support |
|---|---|---|---|---|---|
| CoreWeave facility | $8.5bn | A3 | SOFR + 225bp | 2032 | GPUs and customer contracts |
| Meta Hyperion SPV | $27bn | A+ | 6.58% at issue | 2049, fully amortizing | Lease with a 16-year residual value guaranty |
| SB Energy Ohio campus | Up to $105bn guaranty | n/a | n/a | 20-year lease | Vendor residual value guaranty, with offtaker indemnity |
| Term | Energy PPA | Compute Capacity Offtake |
|---|---|---|
| Unit and price | $/MWh, fixed or hub index plus a differential | Reserved GPUs or MW at $/GPU-hour or $/MW-month |
| Volume terms | As-generated or baseload shape | Take-or-pay minimum utilization on reserved capacity |
| Tenor against asset life | 10 to 25 years on a 30-year plant | 1 to 5 years, often 12 to 18 months, on 5-year debt |
| Residual value | Stable and well understood | Contested: 2 to 3 year competitive life against 5 to 6 year books |
| Hedge market | Deep: Henry Hub, hub power futures, options | New: NYMEX compute futures from 05 October 2026 |
| Merchant tail tools | Revenue puts, collars, tolls, hedge covenants in lending | Compute collars, deferred strip sales, residual value puts |
| Settlement benchmark | ISO locational marginal price or hub index | Silicon Data H100 and B200 rental indices |
Short tenor against long debt, a contested residual value, and a price market that opens in October. Each gap is closed in energy with hedging, independent marks and structured collars, and each is a revenue line for the Firm.
The index level is a joint statement about useful life and operator margin. When the sector extended disclosed accounting lives from four years toward six, it asserted a point on this curve. The divergence between the life a lender underwrites and the life the rental index is actually pricing is tradeable in the deferred strip and billable as an independent mark.
| Funded Useful Life | Implied Index Floor At A 25% Operator Margin | Against A $2.64 Spot Index |
|---|---|---|
| 3 years | $2.33 | -12% |
| 4 years | $1.92 | -27% |
| 5 years | $1.67 | -37% |
| 6 years | $1.50 | -43% |
| 7 years | $1.39 | -47% |
| Virtuous Loop | Circular Unwind | |
|---|---|---|
| What happens | Low-cost, long-duration capital funds capacity, offtakers lock cost, volume grows | Offtaker revenue lags commitments, defaults trigger re-leasing, residual values fall faster than underwritten |
| What the market needs | Hedge programs for lenders and operators, covenant design, offtake collars, liquidity | Independent marks, downside protection on the deferred strip, pricing for re-leased capacity |
| Regime | Power-constrained, extending | Demand-constrained, reflexive |
| Firm revenue line | Structuring mandates, account growth, the pooled vehicle on its trigger | Advisory marks, residual value structuring, optionality in the book |
| Book posture | Relative value on the curve and the cross-generation ratio | Own optionality over linear shorts, because the belly gaps |
A direction-neutral desk with no chip inventory, no index affiliation and a residual value model earns from the need to price the outcome, whichever path the cycle takes. That is the structural advantage of standing beside the financing rather than inside it.
| Line | Per MWh Of Facility Energy |
|---|---|
| Compute revenue | $1,768 |
| Power at $45 per MWh | ($45) |
| Network, staff and overhead | ($202) |
| GPU amortization | ($618) |
| Facility amortization | ($104) |
| Compute spark spread | $799 |
A gas plant’s spark spread runs about $15 to $45 per MWh, fuel margin only, on roughly $1mm of capital per MW. Compute earns $1,768 per MWh of facility energy but carries about $33mm of capital per MW, so the true comparison is after capital recovery. Leaving out capital recovery overstates the spread by about $720 per MWh. The spread is set by utilization, rental rate and useful life, which is why it must be hedged and marked like a commodity margin.
| Lever | How It Works | Why It Pays |
|---|---|---|
| Power-first siting | Stranded industrial sites, co-location with generation, behind-the-meter supply | Removes the transmission queue from the critical path |
| Flexible load | Curtailable, latency-insensitive workloads during grid stress | Faster or larger interconnection where grids accept flexible load |
| Long-lead equipment | Transformers, switchgear and turbines reserved against a dated plan | Protects the date the spread depends on |
Five to eight points of illustrative equity return are given up for each twelve months of interconnection delay. A year of slippage gives back most of what the hedging overlay adds, which is why time-to-power diligence leads every mandate.
| Structure | How It Works | What It Captures | What It Needs |
|---|---|---|---|
| Fixed power, floating compute | 15 to 20 year fixed or virtual PPA from nuclear, geothermal or solar plus storage, against 3 to 5 year offtakes | Power capped for 15 to 20 years while compute resets upward in scarcity | Short offtakes priced and collared |
| Grid-tied dual market | An ERCOT, PJM or CAISO facility curtails batch work in price spikes and sells the power position | Pays above about $1,400 per MWh; the larger prize is flexible-load interconnection | Dispatch rules and a financial power hedge |
| Virtual compute spark hedge | A fixed power hedge plus a forward compute sale for the first operating years, locked before operation | A financeable spread that supports non-recourse debt | OTC cover for 2028 to 2030, since the listed strip spans months |
The first needs the offtake tail priced, the second needs dispatch economics and a power hedge, and the third needs a counterparty for multi-year compute cover before the listed curve reaches those dates. The Firm structures the OTC cover, runs the hedge and lays risk off into the listed market as the strip lengthens.
Land, power, interconnection, shell and cooling. Fifteen-year amortizing debt at about 60 percent of cost, from infrastructure-grade lenders at lower spreads, against an asset that outlives several chip generations.
Take-or-pay offtake for years one to three, collars on the cascade for years four to seven, a fixed power hedge on the input, and independent marks each quarter for both lenders.
GPUs and networking at commercial operation. Three-year amortizing debt sized to hedged cash flow, with principal repaid inside the contracted term.
| Illustrative, Per MW | Merchant, One Vehicle | Overlay, Two Vehicles |
|---|---|---|
| GPU debt capacity at a 1.20x debt service coverage ratio under a $1.00 stress | 26% of GPU cost | 70% of GPU cost |
| Minimum debt service coverage ratio, base case | 2.75x | 1.54x |
| Equity internal rate of return in the $1.00 stress case | -8.2% | +22.9% |
Hedged cash flow supports debt sizing the merchant case cannot. That is the whole of driver three, and it is why every project can carry the overlay.
| Tier | Workload | Counterparty | Illustrative Rate |
|---|---|---|---|
| Years 1 to 3 | Frontier training | Tier-one lab or hyperscaler, take-or-pay terms | $2.20 per GPU-hour |
| Years 4 to 5 | Enterprise fine-tuning | Pre-arranged secondary offtake at a 30% step-down | $1.54 per GPU-hour |
| Years 6 to 7 | Baseline inference | Distributed GPU networks and edge inference | $1.10 per GPU-hour |
| Test | The Stress | Overlay Response |
|---|---|---|
| Rate compression | Next-generation supply outruns demand and rates fall from $2.50 toward $1.00, with lenders testing $0.70 | Take-or-pay front years and a collared tail, with debt amortized inside the contracted term |
| Interconnection delay | A 12 to 24 month slip carries land, deposits and power commitments before revenue | Power-first siting, flexible-load interconnection and dated equipment reservations |
| Cooling efficiency | Power usage effectiveness drifting from a liquid-cooled 1.15 design to 1.40 raises power per compute unit by about 22 percent | Power is under three percent of revenue, so a 22 percent rise costs roughly half a point of margin |
| Rack density | Next-generation racks exceed 100 kW and need direct-to-chip liquid cooling | Shell specification for 120 to 140 kW racks from day one |
| Offtaker credit | Front-year take-or-pay depends on the lab or hyperscaler paying through the cycle | Investment-grade or guaranteed offtakers, with independent marks on replacement value |
| Chip transition | An annual architecture cadence discounts deployed chips by the fourth year | Cascade tiers priced from the curve and collared before commercial operation |
Site, queue and equipment path scored for energization date and megawatts available.
A compute spark spread model with implied-life residuals and lender stress cases.
A fixed power hedge and OTC compute cover, laid off into the listed market as the strip extends.
Tiered offtakes and collars priced from the curve before commercial operation.
Infrastructure and GPU special purpose vehicles, with lender packages built on hedged cash flow.
Quarterly independent marks and covenant reporting for sponsors and lenders.
The same research platform, hedge book and marking engine serve each mandate, so the program scales with the development pipeline rather than with headcount, and every project it touches adds hedging flow to the listed market.
The overlay is priced per project on its own inputs. The first conversation is time-to-power diligence, and it costs nothing.
Start A ConversationThe Hedging And Structuring