Each line opens only after the information boundary between the client business and the proprietary book is documented. That boundary is described in full under Risk And Governance.
| Mechanism | How It Works In Energy | The Compute Parallel |
|---|---|---|
| Asset-backed trading | Owning storage, shipping or spare capacity creates optionality and information the trading arm monetizes | Hyperscalers and neoclouds choose to run a cluster, rent it out or idle it |
| Administrative pricing | A national oil company sets a monthly official selling price against a benchmark and the market reprices around it | Allocation, list prices and backstop terms do the same job for the chip |
| Tolling and risk transfer | A toll converts fuel into power for a fee with no commodity risk, and the offtaker holds the spread | The key structural choice for a data center operator |
| Mandatory hedging | Reserve-based lending requires hedging a fixed share of production: the largest source of hedge supply in energy | Its compute equivalent is when listed liquidity becomes real |
Separation: a separately incorporated trading entity with its own profit and loss, its own compliance officer, and information barriers between operations and the book. Compute Capital Partners is built on that model from inception.
A merchant plant buys gas, sells power and owns the spark spread, so earnings are volatile and the equity trades on commodity prices. Under a toll the offtaker supplies fuel and pays fixed capacity and conversion fees. Same steel, two businesses, two costs of capital.
A neocloud buys GPUs, sells GPU-hours and owns the compute spread: the merchant plant. A colocation or powered-shell operator sells space, power and operations for a fee while the customer owns the metal: a toll with no GPU-hour price risk. Most operators sit between the two.
| Structure | GPU Owner | GPU-Hour Price Risk | Natural Hedge | Capital Cost |
|---|---|---|---|---|
| Powered shell / colocation | Customer | Customer | None needed | Lowest |
| Toll with conversion fee | Customer or a special vehicle | Customer | None needed | Low |
| Contracted GPU-as-a-service | Operator | Operator, after contract expiry | Sell the deferred strip | Medium |
| Merchant neocloud | Operator | Operator, in full | Sell the strip | Highest |
Selling megawatts under a toll or GPU-hours as a merchant is the decision to shed or hold compute price risk. From the listing, that choice is quantifiable for the first time, and pricing it for operators is a core structuring product.
Reserve-based lending agreements filed this year require borrowers to enter swaps, puts or collars on at least half of production for the first twelve months, with a further tranche beyond. Others tie the requirement to utilization thresholds. The flow is non-discretionary: lenders demand it.
A $2.6bn neocloud facility struck in August 2026 carries roughly a five-year maturity against customer contracts averaging about three years, and most offtake agreements with AI developers run twelve to eighteen months. Rating agencies haircut revenue beyond the contracted term.
Once a GPU-backed credit agreement requires the borrower to hedge a share of uncontracted rental revenue against the listed index, the sector becomes a scheduled, price-insensitive seller of the strip: the mechanism that made oil and gas deep. The hedge is worth advance rate and cost of capital, because a hedged post-contract revenue line takes a smaller haircut and supports more debt against the same asset.
The scale trigger. The first GPU-backed credit agreement carrying a compute hedging covenant, visible in a quarterly filing footnote or filed exhibit before any exchange volume statistic. The Firm monitors for it continuously.
Training and inference budgets are set annually against a price that fell 75 percent and partly recovered inside thirty months. A lab buying at spot carries an unhedged input cost on the largest line of its income statement.
Hedge: buy the nearby strip, one to twelve monthsMulti-year capacity commitments are forward purchases of a commodity that cannot be stored. If demand disappoints or an efficiency gain resets the compute-to-revenue ratio, the lab holds long compute it cannot use.
Hedge: sell the deferred strip against the uncommitted portionThe listed strip is the first instrument that expresses a view on the AI capital cycle without a view on any single company. Selling deferred compute isolates whether the marginal cost of AI compute falls faster than demand rises.
A long-dated book is valued against a price the holder does not set. An independent, governed settlement price turns a model valuation into a mark. Three features of this market make that harder, with no implication of wrongdoing: settlement index providers are backed by proprietary trading firms that will also trade the contracts; nine and twelve-figure residual guaranties and multi-year collars are carried against a curve that has yet to trade; and a vendor that sells the chip, finances the buyer, shares rental revenue and guarantees residual value holds four exposures to one price.
A hedge is margined and the underlying asset is not, so cash timing runs opposite to economics. When price spikes, the hedge pays out immediately in cash while the offsetting gain arrives slowly as revenue. European utilities needed state liquidity support in 2022 for this reason. For a neocloud, collateral is a depreciating GPU and the margin call is due tomorrow, so the facility is sized first. That is the first question the Firm asks a prospective client.
An independent desk with no index affiliation and no chip inventory is the natural party to mark a book and structure a hedge for a client that needs both. Meeting the two conditions is what the structuring business sells.
Not with the hedge ratio, but with liquidity facility sizing. It is the question that separates an adviser from a salesperson, and it is where we start.
Start A Conversation