Collateralizing compute: taking security over GPUs

Security over equipment is straightforward in principle. A creditor takes an interest in identified assets, perfects it according to the rules of the relevant jurisdiction, and can realise it if the obligation is not met.

Compute complicates every step of that, and in a specific way: the legal position is usually fine and the practical position is much weaker than it reads.

What determines actual recovery is a short list of physical and contractual facts — whether the equipment can be identified, whether it can be reached, whether it can be sold, and whether anyone else has a competing claim. A perfected interest that fails on any of those is a document rather than a remedy.

The mechanics of perfection are jurisdiction-specific and belong to counsel. What follows is what capital tests commercially, which is consistent across jurisdictions and is where transactions actually run into trouble.

A cluster is a large number of near-identical units. A security description that does not distinguish them from every other similar unit in the same building is a description that will be argued about.

What makes a position durable:

• Serial-level schedules. Identification down to individual units, maintained as a live record rather than captured once at closing. Equipment is replaced under warranty, moved between racks, and swapped during servicing, and a schedule that reflects the position at signing describes a set of assets that no longer exists.
• Physical marking, where the facility permits it. Unglamorous, and repeatedly decisive when ownership is contested.
• An update obligation. A covenant requiring the schedule to be maintained and shared, with consequences for failing to. Without one, the record decays silently.
• Reconciliation against the facility's own records. Where the operator's inventory and the security schedule disagree, that discrepancy surfaces at the worst possible moment.

Commingling is the underlying problem. Equipment belonging to several owners, in one facility, in identical racks, is the normal case rather than the exception.

Compute sits in a building somebody else controls. That single fact governs more of the recovery analysis than the security documents do.

A creditor that cannot enter the facility cannot inspect, cannot verify, and cannot remove. So the arrangements with the facility are part of the security position whether or not they are drafted that way:

• Consent and access rights from the facility owner or operator, agreed directly rather than assumed from the customer's own agreement. This is the single most valuable document in a compute security package and the one most often missing.
• Acknowledgement that the equipment belongs to someone other than the occupier, which is what prevents a claim arising against it in the operator's own insolvency or by way of a lien.
• What happens on termination of the hosting agreement. Whether equipment can be removed, on what notice, and whether the operator can withhold it against unpaid fees. A right to detain equipment for unpaid hosting charges sits ahead of a great deal of carefully drafted security.
• Continuity on the operator's own failure. If the facility operator fails, the question is whether the equipment can be reached and whether the site keeps running long enough to matter.

The test is simple to state and frequently fails: if this needed to be enforced next month, who would have to agree, and has anyone asked them?

Compute equipment frequently attracts more than one claim, and they are not always visible from the outside.

The parties who may already have a position include the supplier or integrator, where equipment was sold on terms retaining title or granting security until payment; an existing lender with a general security interest capturing equipment as a class; a lessor, where part of the estate is leased rather than owned, in which case it is not the borrower's to pledge at all; and the facility operator, whose rights over equipment on its site may arise by contract or by operation of law.

Two practical consequences follow.

Search before advancing, not after. Registries in the relevant jurisdictions will reveal much of this. What they will not reveal is a title-retention arrangement in a supply contract or a hosting agreement's detention rights, which have to be read out of the documents themselves.

Settle priority explicitly. Where several parties have positions, an intercreditor arrangement recording who ranks where and who may do what is worth more than a stronger position on paper that is contested. Mixed estates — some owned, some leased, some financed by different providers — are common and manageable, provided the boundaries are recorded before anyone needs them.

Equipment moves, and security follows the rules of the place where the equipment is rather than the place the contract was signed.

A structure whose equipment sits in one jurisdiction, whose borrower is incorporated in another, and whose documentation is governed by a third is entirely normal and requires perfection to be considered in each. Relocation is the live risk: equipment moved to another jurisdiction may require steps there for the position to remain effective, and a covenant restricting movement without consent is the usual response.

This is legal territory and counsel owns it. What matters commercially is that the question is asked early, because the answer occasionally changes the structure — and because the cost of discovering a defective position is not the cost of fixing it, but the loss of the remedy at the moment it was needed.

Assembled deliberately, a security package over compute contains:

• A specific, maintained equipment schedule, with an obligation to keep it current.
• Perfection in every jurisdiction where the equipment sits, confirmed by counsel rather than assumed.
• Direct agreement with the facility covering access, acknowledgement of ownership, waiver or subordination of the operator's own rights, and treatment on termination.
• Searches and a settled priority position, including any supplier title retention and any existing general security.
• Assignment of the compute contract, so the cash flow and the equipment producing it do not separate at the point of enforcement — this is frequently the more valuable half of the package.
• Insurance, properly noted, covering the equipment and naming the interested parties.
• Restrictions on movement, with consent required before relocation.

None of this is exotic. It is ordinary equipment-finance discipline applied to an asset class where the margin for error is narrower, because the collateral is depreciating while the question is being resolved.