GPU lease structures: operating, finance, FMV, leaseback

Lease taxonomy is easier to follow once the organising question is clear, and the organising question is not tax treatment or balance-sheet presentation. It is this: when the term ends and the equipment is worth a fraction of what it cost, who absorbed that?

Every structure below answers it differently, and each answer implies a different payment profile, a different set of end-of-term rights, and a different party bearing the technology risk.

This matters more for accelerators than for most equipment because the decline is steep and driven by product cycles rather than wear. A structure that quietly leaves the residual with a party that cannot price it is not a cheaper structure. It is an unpriced position that will be discovered later.

The user takes possession and use for a defined term, and the equipment returns at the end. The lessor retains ownership and retains the residual position.

What it does. Converts a large purchase into a periodic payment for a period shorter than the asset's full life, and hands the obsolescence question to a party whose business is holding it.

When it fits. Where the compute contract is finite, where the user does not want to hold technology risk, and where the equipment is current enough that a lessor will underwrite a meaningful residual. Users whose demand is genuinely uncertain past the contract term are the natural case.

What to read closely. Return conditions are the substance of the structure, not boilerplate. What state the equipment must be returned in, who bears de-installation and freight, what counts as excess wear, and whether the return window is workable. Onerous return terms are how an apparently attractive rate is recovered at the end.

The user takes possession for substantially the asset's useful life and, in economic terms, most of the risks and rewards of ownership along with it. Frequently there is a purchase right at the end at a nominal amount.

What it does. Functions as financed ownership. The user is exposed to the residual, because the user will in practice end up holding the asset.

When it fits. Where the user wants the equipment for its full productive life, where there is confidence in sustained utilisation, and where holding the technology risk is acceptable or unavoidable. It is also common where no third party is willing to underwrite a residual, which is a signal in itself.

What to read closely. Whether the arrangement is genuinely a lease or ownership described as one. That distinction drives accounting and tax treatment, and it determines what the user actually holds if things go wrong.

The end-of-term option is where the residual allocation becomes explicit.

Fair-market-value option. The user may buy at the value at the time. Because the lessor keeps the upside and the downside of that value, payments during the term are lower and the user keeps genuine flexibility to return, extend or buy. This is the structure that most cleanly separates the compute layer from the user's balance sheet.

Nominal purchase option. The user may buy for a token amount, which means the asset was always going to end up with the user and the payments amortised effectively all of its value. It is a financing in lease form, and it should be evaluated as one.

Fixed-price option. A middle case, and the one that needs the closest reading. A fixed price set years in advance is a view on where values land. Whether it is favourable depends entirely on how the curve actually behaves, and on hardware of this kind the range of outcomes is wide.

The useful discipline is to ignore the label and ask what the option implies about who expects to hold the asset.

Operating lease and finance lease are useful shorthand, but accounting classification is the result of applying rules to a contract. It is not a substitute for reading who bears loss in the real transaction. Two arrangements reported differently can leave the user with similar cash and residual exposure; two arrangements carrying the same commercial label can produce different remedies.

Begin with the payment obligation. Is the rent cancellable, or is the full stream effectively owed once equipment is accepted? Does a customer-contract termination excuse anything? Are power, maintenance and insurance separate pass-throughs or embedded in rent? A lease whose payments continue in every downside behaves like debt even if its presentation receives attention for another reason.

Then read ownership economics. Who receives sale proceeds, bears casualty shortfall, decides whether equipment is upgraded, and benefits if value is higher than expected? A nominal purchase option normally leaves those economics with the user. A genuine return right and market-value option normally leave them with the lessor. Fixed-price options, guaranteed residuals and mandatory extensions divide them.

Next read control. The legal owner may restrict location, configuration, subleasing, workload, software, maintenance provider and modifications. Those controls protect collateral, but they can also prevent an operator from reallocating equipment across customers or facilities. Commercial ownership belongs partly to the party that controls the asset's next use, not only to the party named on title.

Finally read remedies. On default, can the lessor accelerate all rent, take the equipment, require a make-whole payment, or do several of those subject to a proceeds credit? Can it enter the data centre? Does the hosting provider recognise its ownership? An asset-return remedy without access is theoretical; an acceleration remedy plus a residual claim can give the lessor substantially more than a simple equipment recovery if not limited by the documents and applicable law.

The right conclusion is therefore transaction-specific: describe the payment profile, residual, control and remedies first, then state the accounting treatment. Reversing that order lets a reporting label do work it was never designed to do.

DigitalOcean's September 2026 equipment facility is useful because the public filing exposes the mechanics rather than only the facility size. It provides committed financing through a lessor, funds up to 90% of eligible equipment cost while the lessee contributes the balance as prepaid rent, amortises the advances fully to a common 2030 maturity, and transfers title for nominal consideration after payment. (DigitalOcean Form 8-K, as of September 20, 2026)

That combination answers the residual question. The lessor holds legal title during the term, but the lessee funds the first-loss contribution, repays the advanced cost and receives the equipment for a token amount. The structure is financed ownership rather than a bet by the lessor on end-of-term resale. Its collateral and guarantee package protects repayment; a meaningful unamortised residual is not needed to make the economics work.

The example also shows why advance mechanics matter. A committed facility is not necessarily funded on day one. Equipment is acquired in draws, each draw needs eligibility and documentation, and unused commitments can carry fees. The operator must align purchase orders, delivery evidence, invoices, acceptance and draw availability. A supplier deposit due before the lessor is prepared to fund remains an equity or bridge requirement even when the eventual equipment is eligible.

The fixed final maturity creates another discipline. A later draw amortises over a shorter remaining period unless the documents provide a separate schedule, so equipment ordered near the end of the availability window can carry a heavier payment profile. A sponsor comparing this structure with a vendor programme should model each anticipated draw rather than divide the headline commitment by the facility term.

Most importantly, the public terms demonstrate that lease form does not imply residual transfer. A lessor can be the legal purchaser and the lessee can still retain essentially all technology risk. Anyone choosing a lease to avoid obsolescence has to find the return right and the lessor's unamortised value in the economics. Without them, the lease changes the funding path, not who owns the downside.

GPU transactions have a period conventional equipment summaries often compress: cash is committed before a productive asset exists, and the lease may not begin until after delivery and acceptance. The party carrying that interval is financing procurement risk rather than operating equipment.

The timeline begins with a reservation or purchase order and deposit. It continues through manufacture, shipment, customs, delivery to the site, installation, network and cooling integration, testing and customer acceptance. During that period the configuration can change, the site can slip, a successor architecture can be announced and the end customer can delay its own commencement. Yet the supplier still expects payment under the equipment contract.

A lease should specify when the lessor is required to purchase, when the lessee's obligation becomes irrevocable, and what document proves acceptance. If the lessor funds a supplier before acceptance, it will usually require the lessee to bear non-delivery and specification risk or provide progress-payment indemnities. If it funds only on acceptance, the sponsor needs another source for deposits and interim payments. Neither answer is wrong; leaving the interval unfunded is.

Partial delivery complicates the result. A rack-scale system may arrive in components that have little productive value separately. Rent commencing on each invoice can force payment before a complete system operates. Rent commencing only on full system acceptance leaves the lessor exposed to paid-for components at the site. Milestone funding, title vesting, storage and insurance provisions should follow the unit that can actually be completed and deployed.

The customer contract needs the same map. Ideally, equipment acceptance, lease commencement and compute-service payment begin close together. Where they do not, the model should show lease carry before revenue and identify who covers it. A nominally matched four-year lease and four-year customer contract can contain a material mismatch if the lease begins six months earlier. Matching tenor means matching dates and conditions, not matching the number printed beside years.

A static lease assumes the original equipment stays in place until maturity. An AI platform may need to refresh sooner to preserve customer pricing or performance. The documents should decide how that change reaches the lessor before commercial pressure makes the answer urgent.

One approach is a scheduled refresh option. The lessee may replace equipment at specified dates if it pays an agreed termination amount, sells the old units through an approved channel or rolls remaining value into a new schedule. This preserves flexibility but can hide negative equity: the old lease balance does not disappear because the new equipment is better.

Another is collateral substitution. New equipment replaces old collateral after value, title, location and lien conditions are met. That can protect the lender while permitting operational upgrades, but only if proceeds from the released units and new equity keep the borrowing position covered. Substitution based solely on original invoice cost invites weaker collateral to replace stronger coverage.

A third is portfolio treatment. Several vintages sit inside one borrowing base and newer assets compensate for older ones subject to concentration and advance-rate rules. This is efficient at scale and introduces cross-collateralisation: a problem in one deployment can trap equity across the fleet, and selling a single cluster may require a facility-wide test.

The customer side must move with the equipment. A refresh can require downtime, new acceptance, software migration and a revised service specification. If the end customer can reject the new configuration while the lease has already financed it, the operator owns the transition risk. Conversely, a customer contract that requires continuous current-generation equipment can force capital spending faster than the lease permits.

A financeable refresh provision therefore connects four consents: supplier, lessor, hosting provider and compute customer. It allocates removal cost, data handling, sale proceeds, remaining rent and new acceptance. Flexibility is not a sentence saying upgrades are permitted. It is a waterfall showing what happens to every dollar and obligation when the old equipment leaves.

In a true operating lease the return obligation is the price of the flexibility. It deserves the same modelling attention as scheduled rent because it can produce a large final cash requirement at exactly the moment the equipment stops earning.

The return standard should identify configuration, component completeness, operating hours or condition metrics, maintenance records, firmware, cosmetic damage, packaging, location and testing protocol. Missing network components, proprietary cooling parts or matched interconnect can reduce the value of a complete system even when every accelerator functions. The definition of ordinary wear should fit equipment run continuously in a data centre rather than a generic office-equipment form.

Timing matters as much as condition. A short return window may require de-installation before the replacement system is ready. A missed window can trigger holdover rent or an automatic extension, converting a planned exit into another payment period. If the lessor controls the inspection date or refuses delivery while disputing condition, the documents should prevent an unresolved inspection from extending rent indefinitely.

Location and logistics determine the real cost. The lessee may have to disconnect liquid cooling, sanitise data, obtain export or transfer approvals, pack to manufacturer specifications, insure transit and deliver to a designated destination rather than make the units available at the site. Taxes and duties can change depending on that destination. A reserve for return cost is prudent where the customer contract ends at the same time and no operating cash remains.

There should also be a procedure for disagreement. Independent inspection, cure periods, a schedule of stipulated damage amounts and a cap where appropriate are more financeable than an open obligation to restore equipment to the lessor's satisfaction. The lessee needs a measurable exit; the lessor needs equipment it can actually remarket.

An operating lease is valuable precisely because the user can return the residual. If the return conditions make return impractical or unknowable, the economic choice collapses toward purchase regardless of the option language. The final diligence question is therefore simple: could the operator comply on the worst plausible end date, and has the cash and downtime required to do so been funded?

Equipment already owned is sold to a third party and leased back, so the user keeps using it while the ownership position — and the residual with it — moves elsewhere.

What it does. Releases capital committed to equipment that is already deployed and productive, and transfers the technology risk to the buyer.

When it fits. Where a sponsor has funded equipment directly, wants the capital back for the next stage, and is willing to give up the residual to get it. It is also a way to bring the compute layer into a structure matched to its life after the fact, where it was originally bought outright and bundled with everything else.

What to read closely. Whether the equipment can actually be sold — existing security interests, hosting agreements that restrict transfer, and whether the site arrangements survive a change in the equipment's owner. A sale-leaseback over equipment that a facility agreement will not permit to change hands is not available, however attractive it looks.

The same logic applies across the stack; on the generation and facility layers it is a long-established structure, and only the residual behaviour differs.

The structure should follow from the compute contract, not the other way round. In practice three questions settle it:

1. How long is the demand actually contracted for? A short, firm contract points toward a structure the user can exit. A long one, or genuine confidence in reuse afterwards, supports taking the residual. 2. Can the user price obsolescence? Operators with a view on the product cycle and a route to redeploy or remarket equipment can hold residual risk economically. Most users cannot, and pay less in the end for handing it to someone who can. 3. What does the rest of the stack look like? Where the facility and power are held on long structures, keeping compute short and separable preserves the separation the stack needs. Bundling it back in reintroduces the mismatch the structure was meant to solve.

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