Investing in the Middle Market for Compute

This note is about a specific segment: compute deals below the hyperscale tier, where the borrower is a real operating business and the offtaker is not investment grade. It is written for credit investors, and what it offers is the domain detail that is hard to get from outside the flow, along with an account of how transactions in this segment are actually being structured.

We are a brokerage for physical compute capacity. We also publish a price index and arrange financing. That means we see the offtake contracts, the quotes behind them, and the financing conversations that follow, usually before any of it reaches a lender. Most of what is below comes from that position rather than from research.

We wrote it because the same four questions come up in nearly every conversation we have with credit investors: what does the market actually price, is there a secondary market for the hardware, how do you tell good offtake from bad, and what does the depreciation curve look like. Those are data and documentation questions more than credit questions, which is why an outsider’s frameworks transfer here more easily than most people expect.

The shape of the opportunity

What the segment looks like from where we sit:

The structural mismatch

Loan tenor gets sized to contract tenor. The assets take three to four years to pay back. Where the offtake runs three years or longer the match is workable; where it runs twelve to twenty-four months, few lenders will bridge the difference, and that gap is where the return is.

The arithmetic, since the inputs are not widely published. A 1,024-GPU cluster at roughly $98,000 per GPU all-in is a $100m asset. Contracted at $4.40 per GPU-hour it produces around $39.5m of gross revenue a year. At 30 to 40 percent operating cost, payback lands between 3.6 and 4.2 years. An operator put it to us in one sentence: the down payment and amortization have to work over four years for the business to cash flow.

Offtake tenor varies more than is generally assumed. Twelve to thirty-six months is the common range, three years is a frequent minimum on new-generation capacity, and five-year contracts are available where the offtaker is strong enough to sign one. Shorter blocks of one to six months trade actively alongside all of it.

The financing question is therefore less about the market as a whole than about the specific deal in front of you. A three-year contract against a four-year payback is a manageable structuring problem. A twelve-month contract against the same asset is the one that goes unfunded, and it is also the one where pricing is most attractive for a lender willing to look at the re-let rather than only the contract.

The formats we see used to bridge it are familiar ones: revolving borrowing bases, utilization-based sizing, portfolios of staggered contracts. Nothing exotic. The point is that the mismatch is mechanical rather than fundamental, and lenders willing to work with it are being paid for a structuring problem rather than a credit problem.

Terms of art

Reference only. Skip if you are already transacting in the sector.

The institutional market, and what carries down

Compute financing is settled at the top of the market. Several very large transactions have priced in the past eighteen months, and a coalition of major alternative managers has announced platforms to mobilize third-party capital into compute infrastructure at scale.

Four things from those deals travel down-market: contracted revenue rather than hardware value as the underwriting basis; coverage and liquidity tests as the accepted covenant format; asset-level SPV structures; and credit support as the principal lever on pricing.

That last one is worth a sentence of detail. Applied Digital, whose CoreWeave leases run roughly fifteen years across approximately 400MW, has stated that it obtained credit enhancements tied to the tenant subsidiary securing an investment-grade refinancing, and that reducing its own project-level debt cost depends on tenant credit improving. In March 2026 those leases were amended and assigned to a CoreWeave subsidiary, with guarantees and a letter of credit added. The financing cost of a compute asset tracks the offtaker’s credit far more closely than the hardware’s.

What is different below that tier

The same principles have to be delivered by different means, because there is no tenant rating to lean on. In practice the documentation carries the weight the rating would otherwise carry, and the spread reflects that.

What we see in the segment

The defining characteristic, and it is more specific than most descriptions of a middle market: good businesses cannot get financed because their customers are not credit quality. These are often profitable, high-utilization operators with demonstrated pricing power. What they lack is an offtaker a traditional lender recognizes.

Three situations from the past year, anonymized:

Prepayments, and what they actually are

Upfront payments of 20 to 35 percent are standard. Two economically different things go by that name. Where the payment is credited against rent, it reduces the buyer’s balance and the contract is self-liquidating in the ordinary way. Where it funds the operator’s contribution to the capital stack, the buyer is effectively a subordinated capital provider sitting behind the operator’s senior lender.

The second version is more common than it looks, and it has practical consequences: the money cannot be escrowed to delivery without breaking the operator’s funding condition, and any facility financing that prepayment is not lending against a self-liquidating contract. It is worth resolving early, and it is the first thing we test on a deal of this shape.

The intermediation chain

Deals often run several parties deep: facility operator, GPU cloud, reseller, end user. Credit quality and capital thin at each hop. The useful question on an offtake contract is not only who signed it but whether that signer has contracted demand behind them or is warehousing capacity against demand they expect.

Depreciation, residual value, and the secondary market

Residual is unpriced rather than unknowable

No meaningful fleet has come back and been re-let at scale, so the first one that does will set the recovery assumption everyone marks against. Every institutional lender we have spoken with has said some version of the same thing: future value is a blind spot and the data does not exist.

We would read that as an opportunity rather than a reason to stay out. Term prices are observable, decline rates are measurable, utilization data exists. The premium currently attached to this uncertainty is compensation for information work, and it compresses once the first liquidations are public.

Obsolescence is a jump, not a slide

The one modeling convention we would push back on, because it is where structures in this sector have failed: decline is usually modeled as a smooth annual percentage. Value in fact steps down discontinuously, when a model release crosses a memory or interconnect threshold and a hardware generation stops being able to run the workloads buyers want. Term pricing for a previous-generation chip deteriorated sharply following a recent frontier release on exactly this mechanic. A structure that survives a smooth 30 percent annual decline may not survive a 60 percent quarter.

Prior-generation flagship pricing has fallen substantially over the past eighteen months, and the useful observation is not the magnitude but the distribution: the declines cluster around specific model releases and supply events rather than spreading evenly across quarters. Averaging them into an annual rate produces a curve that no individual quarter ever looked like.

The contract-level version is what we call spec expiry. A compute contract is a claim on a specific instance type. If the frontier moves up in resource requirements, a multi-year contract can lose economic value well before legal maturity. A three-year contract with six months of specification headroom behaves more like a shorter one, and the specification headroom is knowable at close.

Observable price behavior

Two readings. Near-term availability is scarcer than cheap capacity, so immediacy carries a measurable premium and the curve is not reliably upward-sloping in the way commodity intuition suggests. And the upfront requirement, which sits around thirty percent and is negotiable on most deals, is a financing curve rather than a pricing one: prepayment moves inversely with tenor because it substitutes for a capital layer operators cannot otherwise raise. When a buyer cannot fund the deposit, the lever they pull is tenor, not price.

Remarketing rather than repossession

A provisioned GPU is difficult to move, so recovery in practice runs through servicing the existing customer or selling in place. That is less of a weakness than it first appears, since a cluster that keeps running keeps producing revenue through a workout. What it requires is arranging in advance: step-in rights, colocation access, and a named party capable of re-letting. All three are cheap at signing and expensive to assemble at default.

The constraint is that re-letting capacity is a different activity from liquidating equipment. It needs live buyer demand, a current view of where comparable capacity is clearing, and paper the replacement customer will actually sign, none of which desk assembles on short notice. Recovery in this asset class is a distribution problem more than a disposal problem, and the difference between the two shows up directly in the advance rate a lender can justify.

Two opportunities in the capital stack

These two structures account for most of the deals we see stall, and in both cases what is missing is a provider rather than a solution. Neither is a credit problem in the conventional sense, and both carry returns well above what the underlying risk supports once the structure is right.

The down payment problem

Most lenders arriving here picture the same transaction: a secured loan against a GPU cluster with contracted offtake behind it. That is the right picture, and it is the deal we see most often. There is one wrinkle worth knowing before the first one, because it is where these deals stall.

Hardware is not bought on delivery. The vendor requires roughly twenty to thirty-five percent up front before it will build and ship. A lender advancing, say, seventy percent against equipment value cannot fund that deposit in the ordinary way, because at the moment it is paid there is no equipment to secure. The operator is usually expected to cover it, and in the middle market the operator frequently cannot. That is the single most common reason a well-documented cluster with signed offtake does not close.

The deals that do close solve it one of three ways:

None of this is exotic. It is the ordinary shape of financing capital equipment that has to be paid for before it exists, and the return available for solving it is well above what that risk supports.

Prepayment financing for compute buyers

The demand-side mirror. A buyer with real workloads and a signed contract cannot fund the upfront payment, so the deal shrinks, shortens, or dies.

The important point about these buyers is that they are generally not short of money. They are venture-backed companies with substantial cash, and they are funding compute out of equity because no one has offered them anything else. That is the most expensive capital available, priced for company-building risk, being used to pay a deposit on a contract that consumes it in the ordinary course. Founders dislike it, boards dislike it, and it persists only because the alternative has not been built.

The efficiency argument is straightforward. A buyer paying a deposit out of a venture round is financing a known, contracted, consumable input with capital that costs several times what debt against that same contract should cost. Substituting one for the other lets the company hold its equity for the things equity is actually for, and lets it commit to longer tenor at better pricing, which improves the offtake quality for everyone above it in the structure.

Where the prepayment is credited against rent, a facility funding it is lending against a contract the borrower consumes and pays down over the term, and the collateral is the contract rather than the borrower. On default the lender holds re-lettable capacity rather than a claim on a struggling company. Three conditions make that work, and all three are drafting decisions: the contract must be assignable, funds escrowed and released against delivery milestones, and a remarketing agent named at close.

Two issues are unresolved and we would rather name them. Tickets are small against most facility minimums, which argues for pooling. And a pool of venture-funded buyers carries correlation that aggregation does not fix. The mitigants we think are real are contract-level rather than borrower-level collateral, concentration limits by operator and chip generation, and originator first-loss. We are working on both and would rather shape them with a lender than present a finished term sheet.

What operators will and will not give

The security package below is not novel as credit structuring. What is specific to compute is which items are actually obtainable, which get resisted, and which turn out to matter more than most people think. This is our read from deals in the market.

The four that carry the structure

Worth asking for, negotiable in practice

• UCC-1 against specific equipment by serial number, with a pre-close lien search on the operating entity and parent.
• Confirmation of ownership versus lease. Many operators lease and cannot pledge what they do not own, which is not always volunteered.
• Mortgagee or ground-lessor consents where the facility is itself financed.
• Escrowed prepayments released against delivery and acceptance milestones, where the operator’s own funding structure permits it.
• Acceptance testing as a funding condition: node count, throughput, fabric configuration, burn-in.
• Reserve accounts for colocation and power.
• Step-in and novation rights on operator default or insolvency.

Sizing conventions we would flag

• Advance against NOLV rather than cost, with the residual curve stated explicitly rather than left implicit in the advance rate.
• Amortization ahead of both the depreciation curve and the contract decay curve, tested against a step-down rather than a straight line.
• Concentration limits by end customer, chip generation and facility. A single offtake agreement is a concentration position; several across distinct end customers is a portfolio. In our view those should be priced differently than one declined outright.
• Utilization floors and re-let reporting, so deterioration is visible before it reaches the coverage test.

The compute-specific questions

Not a diligence checklist. These are the ones that are particular to this asset and easy to miss from outside it.

1. Does the borrower own the hardware outright, or lease it? The pool of true owners is smaller than the seller count suggests.
2. How many distinct offtake agreements, across how many distinct end customers, and what is the revenue source behind each?
3. Weighted average remaining offtake tenor against the asset’s payback period.
4. Is each contract assignable without end-customer consent, and can waivers be obtained at signing?
5. Take-or-pay or usage-based, and what are the termination rights?
6. Is any prepayment credited against rent, or funding the operator’s own contribution?
7. Who holds the colocation contract, and is there a direct agreement, bailee waiver or access right?
8. What residual curve is the advance rate implying, and does the structure survive a step-down rather than a slide?
9. Who re-lets the fleet on default, on what economics, and are they named in the documents?
10. What is the specification headroom on the contracted hardware against the current frontier?

Working with Liquid Compute

Liquid Compute operates a brokerage for physical compute capacity, publishes a daily reference index governed by an independent third-party benchmark administrator, and arranges financing against fleets and contracts. We came to this market from institutional credit rather than infrastructure, which is why this note reads the way it does. Three ways we work with lenders, independent of one another.

We have live situations looking for capital, including several of the shapes described in “Two opportunities in the capital stack.” Happy to walk through specific transactions, share pricing and index methodology, or be named as remarketing agent on deals you are already underwriting. If anything here is wrong, we would like to know.