Lease vs own compute

The two routes deliver the same compute to the same customer. What differs is where the asset sits, who is exposed to what it is worth in three years, and how easily the position can be changed.

Every row is downstream of one fact: accelerator values fall in steps driven by other people's product roadmaps, not smoothly with age or use. That behaviour is set out in detail on GPU residual value and depreciation, and it is the reason this decision is not the same as leasing a truck.

An owner is long that curve. If the equipment stays useful and utilised past the first contract, ownership was the right call and the operator keeps the whole benefit. If a new generation lands and the workload migrates, the owner discovers it has been holding a position it never priced.

A lessor under an operating structure is short the same curve, and is compensated for it inside the payment. That is why leasing looks more expensive in nominal terms and frequently is not more expensive in substance: part of what is being bought is the transfer of a risk the user could not otherwise place.

The balance-sheet row is genuinely conditional, which is why the table hedges it. Whether a lease keeps the asset off the sponsor's books depends entirely on the structure, and the line between the two treatments is examined in operating vs finance lease. A sponsor that leases in order to keep the asset off its balance sheet and then signs a full-payout structure has bought the cost of leasing without the effect it wanted.

The last row is the one that surprises people. Return conditions, de-installation, freight and excess-wear standards are the substance of a lease, not boilerplate. An attractive payment recovered through onerous return terms is a common shape, and it is where a lease that looked cheap stops being cheap.

Ownership wins under four conditions, and they compound — a sponsor that meets one of them and not the others is usually reaching for the wrong answer.

Utilisation is genuinely durable past the current contract. Not forecast, not hoped for. An operator with a pipeline, a track record of recontracting, and workloads that do not migrate to the newest silicon on announcement is describing an asset with real life left in it. That operator should keep the upside rather than sell it.

There is a route to redeploy or remarket. Ownership is only economic for a party that can do something with the equipment after the first use. An operator running several sites can cascade older parts to less demanding workloads; a single-purpose owner cannot, and is holding a position it has no way to work out.

The balance sheet can carry it. The capital committed up front is not available for anything else. Where that capital is the constraint on building the next site, ownership has an opportunity cost that never appears in the purchase comparison.

No third party will underwrite a residual anyway. This is the honest case and it is common on older or less liquid configurations. Where the market will only offer full-payout terms, the sponsor is being told that nobody wants the residual — and owning at least removes the intermediary. But read the signal: a lessor's unwillingness to write a residual on a given generation is a view on values, and it is usually worth more than the sponsor's own view.

Ownership is the wrong answer where the demand behind the equipment ends on a known date and nothing is contracted after it. That is a finite position, and holding an infinite-life asset against it is the mismatch in miniature.

Leasing wins where the sponsor's exposure needs to end when the demand does, or where the residual belongs with someone who can price it.

Demand is contracted for a finite term. The cleanest case. A structure whose term matches the compute contract leaves no tail — no equipment to house, insure and remarket after the customer has gone. Where the two cannot be matched exactly, the gap should be allocated deliberately rather than discovered at expiry.

The sponsor cannot price obsolescence. Most sponsors cannot, and there is no shame in it. Pricing a residual on accelerators requires a view on product cycles, a remarketing channel and enough volume to be wrong occasionally without it mattering. A party with none of those is not taking a risk when it buys; it is taking an unpriced position.

The rest of the stack is long-dated. Where the site, the shell and the generation equipment sit inside structures measured in decades, keeping the compute layer short and separable is what preserves the separation the stack needs. Buying the compute into the same entity reintroduces exactly the tenor mismatch the structure was built to solve.

Capital is better used elsewhere. A sponsor whose binding constraint is equity for the next site, not cost per unit of compute, is usually better off converting a purchase into a periodic obligation and deploying the difference.

Leasing is the wrong answer where the operator genuinely intends to run the equipment for its whole productive life and has the means to do so. At that point the structure is financed ownership wearing a lease label, and it should be evaluated as ownership — which is precisely the distinction drawn in operating vs finance lease and made explicit by the end-of-term option in FMV vs dollar-out.

One route sits between the two: equipment already bought can be moved off the balance sheet after the fact, which is the subject of sale-leaseback vs refinancing.

In practice the decision resolves against four answers, in this order.

1. How long is the compute actually contracted for, and by whom? This sets the term the position can support. Everything else is arranged around it, and it is the first thing to establish rather than the last. The tests that separate a bankable contract from a signed one are on compute offtake as credit. 2. What happens to the equipment after that term? If the answer is a specific plan — a named follow-on workload, a cascade to a second site, a remarketing relationship — ownership is defensible. If the answer is that something will presumably turn up, it is not. 3. Who is best placed to price the residual? Compare honestly: the sponsor's view of accelerator values against that of a party whose business is holding them. Where the sponsor's view is weaker, paying for the transfer is rational. 4. What does the rest of the structure need? A compute position inside a ring-fenced infrastructure entity constrains that entity's tenor and pricing. Where the project structure is the point, the compute usually has to be separable — see corporate credit vs project finance.

A useful discipline: write down the assumed value of the equipment at the end of the term before choosing. Ownership is a bet that the number is higher than the lessor thinks; leasing is a decision not to take that bet. Sponsors who cannot state the number are usually taking the bet without knowing it.

Continuum structures and arranges across both routes; it does not lend, take equipment positions, or hold client funds.