The scarce resource in AI is not models, it is verified compute. Financing that hardware, secured by the hardware itself, is an underwritable yield story.
Everyone is chasing the AI application layer. The quieter, more underwritable opportunity is one layer down: financing the verified compute that the whole stack runs on, secured by the machines themselves.
Underneath every AI model sits a scarcer resource than the model: verified compute, the actual GPU hardware that produces the output. Most of the capital and attention is crowding the application layer at the top of the stack. The more underwritable question lives one layer down, in how that hardware gets financed. Lend against the machine, with the machine itself as collateral, and you are asking a credit question rather than placing a bet on which model wins. Credit questions are underwritable in a way a blind-pool venture bet never is. This piece is a walk through that thesis as an idea, including the places it breaks, not a description of any vehicle that exists.
A blind-pool equity fund asks an LP to back a manager's judgment about uncertain future outcomes, with nothing tangible behind the dollar and a long wait for any return. An asset-backed credit fund asks something narrower and easier to say yes to: underwrite a loan against a physical thing that can be resold, at a stated rate, over a stated term, with current yield from early in the loan's life. We draw out that distinction in full in asset-backed credit versus blind-pool VC.
Applied to compute, the asset is a high-end GPU rig. The financing is repaid from the operator's covenant to pay and the hardware's earnings, secured by a first-priority interest in the hardware plus an assignment of the operator's compute earnings. That is a dual-security profile: backed by an asset and by a cash flow. It is the shape specialty-finance investors recognize, and it is far more legible to a yield allocator than "trust our picks."
The AI scarcity is verified compute. Finance the machine, secured by the machine, repaid from what the machine earns. That is collateral, not conviction.
A single detail separates a real credit thesis from a sub-prime hardware-leasing scheme. The collateral is only as good as it is verified and earning. In decentralized compute, a large share of registered hardware sits unverified or simply unusable, with public reporting on some networks putting the genuinely verified, usable share in the low single-digit percentages of what is registered. Hardware that is idle or cannot be verified does not function as collateral at all; it is just inventory that fails to pay.
So the underwriting rule writes itself: finance only verified, earning hardware. The verification stack that confirms a machine is real, performant, and actually producing income is not a product nicety here. It is the collateral-integrity mechanism the entire credit case depends on. Closing that gap is exactly what purpose-built, verification-first compute networks are designed to do, with proof-of-work-style attestation, anti-abuse controls, and earnings telemetry standing in for a lender's collateral monitoring. The credit safety and the verification technology are the same problem viewed from two sides. (Griddly, a verification-first compute network in which the studio has an interest, is one example of this architecture; see griddly.ai. The reference is illustrative, not an endorsement of any return.)
This is where the thesis earns its keep or quietly rots. The most dangerous failure mode is circularity: if the only place the financed hardware can earn is one immature marketplace, then the collateral's income is captive to that marketplace's unproven demand. A strategy whose collateral only pays inside your own young network has stopped being credit and become vendor finance wearing a costume, and that is the version that defaults.
The discipline that prevents this is the same three-rule asset-backed-credit discipline we set out in full in asset-backed credit versus blind-pool VC, here applied to compute.
Ability-to-pay independent of any single network. The borrower should be able to service the loan from other revenue or reserves even if network earnings come in well below plan.
Conservative loan-to-value and fast amortization. A meaningful borrower down-payment and principal that amortizes faster than the hardware depreciates. This rule does the most work in compute, because GPUs lose value quickly as new generations ship, so the amortization has to outrun the depreciation or the loan goes underwater on a wasting asset.
Recovery to off-network resale. The repossessed hardware has to be worth something in the open market, not only redeployable on your own network. That is real for compute specifically because high-end server GPUs have a broad secondary market and alternative uses, so prefer generic, liquid hardware over bespoke configurations locked to one platform.
Put those three together and the captive risk is contained. Skip any one of them and the thesis collapses into the exact trap it claims to avoid.
The shape this produces is bond-like: a current-yielding, collateralized return rather than a venture multiple. The strategy earns the spread between the borrower's rate and its own cost of capital, servicing, and credit losses. The credit-loss provision is the swing variable, and on an unproven borrower class with collateral partly exposed to a young marketplace, it should be modeled pessimistically and reserved heavily from day one.
So the honest description is qualitative, not a printed number. The return an allocator could underwrite here is collateralized and current-yielding, secured by a thing that can be resold. Think of it as a bond-like return rather than a venture multiple. The upside is capped, the way credit upside always is. The downside is real, because a hardware-price collapse or a demand failure can drive losses well past a comfortable assumption. Any early book should be sized to prove the loss assumptions, not to chase a headline yield, and the credibility move, as with any first raise, is to under-promise on return and over-deliver on discipline. We work the underwriting in depth in asset-backed credit versus blind-pool VC.
As a matter of strategy, the form reaches capital that blind-pool venture simply cannot. Credit and specialty-finance allocators buy paper rather than picks, and their diligence runs on the collateral and the loan book instead of a manager's exit history. That makes the form more raisable for a manager still building a track record, because the quality lives in the structure and the documents, which are auditable. It also widens the audience: a yield allocator who would never touch venture can still be underwritten into AI infrastructure through credit.
None of which makes it a shortcut. The form does not raise day-one money any more easily than equity, because credit LPs and warehouse lenders both want a seasoned book first. The sequence is the one that runs through this entire segment: originate a few real loans, prove they perform, then raise the pool. The advantage of the credit form shows up at the scale stage, well after the survival stage.
Pull it together and the most underwritable opportunity in the AI build-out is one layer below the model layer everyone is crowding: financing the verified compute the whole thing runs on, secured by the machines, repaid from what they earn, with enough underwriting discipline to keep the collateral from being captive to any single immature market. Done with the verification and the loan-to-value and the off-network recovery in place, it reads as a real, collateralized yield story. Done without them, it degenerates into vendor finance with extra steps. The picks and shovels are fundable. Whether they are well-funded comes down to the discipline.
Read next: Exempt Reporting Adviser path: 203(l) vs 203(m)
For managers and operators reading this: none of the above is legal or financial advice, and nothing here describes a fund that is open, raising, or planned. It is a structural overview of how an asset-backed compute-credit strategy could be built. Asset-backed credit carries real risk of loss, and any actual structure depends on underwriting, market pricing, and the applicable law in each jurisdiction.
Nothing here is an offer to sell a security or investment advice; offers are made only to verified accredited investors via definitive documents.
Field notes on venture building, AI, and capital. No spam, unsubscribe anytime.
By subscribing you agree to receive AOS Insights e-mails. We use your address only for this newsletter - see our Privacy Policy.
