In this note
- Lease, reservation or offtake
- What the commitment is worth
- The checklist
- Term and what the rate covers
- Delivery, acceptance and when billing starts
- SLA: what counts as downtime
- Prepayment and how it is protected
- Exit rights and termination fees
- Hardware refresh
- Placement, region and moves within a region
- Data handling
- Capacity expansion
- If the provider fails
- Reserving with Lightyard
A multi-year GPU lease for a dedicated cluster is a large contract. Every US$1.00 of rate per GPU-hour on a 32-node, 256-GPU cluster is US$11.2M over 60 months, and on 128 nodes it's US$44.9M. Multiply by a Blackwell rate and a single signature commits tens of millions of dollars.
The rate gets most of the negotiation. The clauses that decide what you pay per useful GPU-hour get less attention, and the short version of what to ask for is: billing that starts wave by wave at acceptance, not at signing; an SLA measured per node, with a time to restore a failed node rather than a response time; prepayment held by an independent escrow agent and released against milestones; one defined exit with a known cost; and a plan for the provider failing. This post goes through those clauses in the order a contract runs, with a checklist table you can take into the room and worked numbers you can redo with your own quote.
Where Lightyard's contract answers a point, we say how. Everywhere else it's general advice from the buyer's side. It isn't legal advice, and your counsel should read the actual paper.
Lease, reservation or offtake
Buyers search for "GPU lease". Providers write "reservation", "GPU-as-a-service" or "capacity agreement". Lenders call the same document an offtake. Underneath those words there are two different structures, and they put the risk of a dead GPU in different places.
| Equipment lease | Capacity reservation (GPU-as-a-service) | |
|---|---|---|
| Who owns the hardware | The lessor; you may buy it at the end | The provider |
| Where it runs | Your site or your colo cage | The provider's facility, in a region named in the contract |
| Who runs and repairs it | You | The provider, to an SLA |
| If a GPU fails | You keep paying and arrange the repair | The provider replaces it; credits if it misses the SLA |
| What payment depends on | Acceptance of the goods | Capacity delivered, accepted and kept within the SLA |
| Power, cooling, network, staff | Yours, on top | Usually in the rate |
The difference shows up in the payment clause. Under US commercial law, in a finance lease that isn't a consumer lease, "the lessee's promises under the lease contract become irrevocable and independent upon the lessee's acceptance of the goods" (UCC 2A-407 (opens in a new tab)). That's the "hell or high water" rule. It's normal for leased equipment you run yourself. If a document titled GPU-as-a-service carries the same language ("payment obligations are absolute and unconditional, without set-off or abatement") and also leaves you with a thin SLA, you're carrying hardware risk at a service price. Paying for every reserved hour whether or not you use it is fine, because that's what a reservation is. Paying for hours the provider failed to deliver isn't.
"Offtake" is the lender's word. Clifford Chance's 2026 outlook says GPU financings are increasingly "backed by minimum commitment levels or take-or-pay style compute contracts" and sized "to the expected contracted revenue rather than the neo-cloud's balance sheet" (Clifford Chance (opens in a new tab)). So the provider's lenders will read your payment, assignment and termination clauses too. That's normal. What you want in exchange is obligations that run both ways.
Will the contract count as a lease in your accounts?
Ask your auditors before signing. Under ASC 842 a contract "is or contains a lease if the contract conveys the right to control the use of identified property", and a supplier's right to replace equipment that is "not working properly" doesn't stop an asset from being identified (PwC (opens in a new tab)). Single-tenant nodes can meet that test; whether you direct their use depends on the contract. The answer decides whether the commitment goes on your balance sheet as a right-of-use asset and a lease liability.
What the commitment is worth
Contracts bill GPU-hours, so start there. A 730-hour month is the usual basis (8,760 hours a year divided by 12). The table gives the value of every US$1.00 of rate, so you can multiply by the rate in your quote.
| Cluster | GPUs | GPU-hours per month | GPU-hours over 60 months | Per US$1.00 of rate, monthly | Per US$1.00 of rate, 60 months |
|---|---|---|---|---|---|
| 8 HGX nodes | 64 | 46,720 | 2.80M | US$46,720 | US$2.80M |
| 32 HGX nodes | 256 | 186,880 | 11.21M | US$186,880 | US$11.21M |
| 64 HGX nodes | 512 | 373,760 | 22.43M | US$373,760 | US$22.43M |
| 128 HGX nodes | 1,024 | 747,520 | 44.85M | US$747,520 | US$44.85M |
| 4 NVL72 racks | 288 | 210,240 | 12.61M | US$210,240 | US$12.61M |
| 8 NVL72 racks | 576 | 420,480 | 25.23M | US$420,480 | US$25.23M |
Rates by platform are on the pricing page. At this size, a clause that costs you 1% of GPU-hours is worth more than most rate concessions you'll win, which is why the rest of this post is about clauses.
The checklist
One row per clause. The middle column is what to ask for. The last is what to walk away from, or at least price in.
| Clause | Ask for | Walk away from, or price in |
|---|---|---|
| Term | One term, its start and its end in the order form; no automatic renewal | Auto-renewal with a short notice window |
| What the rate covers | Hardware, facility, power, network and operations in one rate; anything extra listed with its price | Power, support or fabric billed separately at the provider's discretion |
| Billing basis | Hours per month stated; escalators, if any, tied to a named index | Unindexed annual increases |
| Delivery | Target dates per wave, counted from signing | "Commercially reasonable efforts" and no date |
| Acceptance | A written test plan with pass criteria, run before you accept | Acceptance deemed at power-on or after a few days of silence |
| Term start | Billing for each wave starts at that wave's acceptance | The term starting at signing or at first delivery |
| Late delivery | A right to cancel late capacity and get its prepayment back | Delay credits only, capped low |
| SLA scope | Availability per node, with degraded GPUs, NICs and links counted as down; fabric and storage covered | One cluster-wide uptime figure |
| Node replacement | A time to restore a failed node, and spares held on site | "Next business day response" with no restore time |
| Credits | A sliding scale on the affected capacity, and termination for repeated failures | Credits as the sole remedy, capped near zero |
| Prepayment | Independent escrow, staged release against milestones, credit against invoices | Prepayment paid straight into the provider's operating account |
| Exit | A defined exit point, notice date and known cost | A termination fee equal to all remaining payments |
| Hardware refresh | A plan for the next generation: an exit point, rollover, split commitment or priced upgrade | Nothing on the subject, with a rate fixed for five years |
| Placement | Region and facility tier in the order form; moves only with consent, notice and at the provider's cost | A right to relocate at will |
| Data | A DPA, access rules, failed-drive handling, deletion with an attestation | Failed drives leaving the site; no deletion terms |
| Expansion | Fabric sized for your target size; how added nodes are priced and when their terms end | No path to grow on the same fabric |
| Provider failure | Information rights, escrow for undelivered capacity, a data export window | A large unprotected prepayment and no visibility |
Term and what the rate covers
Check that one number buys the whole service. A per-GPU-hour rate should cover the servers, the facility, power and cooling, the compute fabric and the people who run it. Then list what it doesn't: storage beyond local NVMe, egress, software licences, smart hands for your own gear, and managed services above the operating system. Each should have a price in the order form, not "at then-current rates".
Look at the end of the term as hard as the start. Automatic renewal on a five-year GPU contract with a 90-day notice window is an easy way to pay year-one rates for hardware that's two generations old.
Delivery, acceptance and when billing starts
Ask for delivery dates as months after signing, per wave, rather than calendar dates. A calendar date in a contract signed three months late is a date the provider has already missed. How long the dates should be depends on whether the capacity is already running; our deployment timeline post covers what fills those months.
Acceptance is where you have the most leverage, and it's gone once you sign it off. Write the test plan into the contract: node diagnostics, all-reduce bandwidth across the whole fabric, and a multi-day full-load run, each with a pass threshold. For a cluster of a few hundred GPUs, two to six weeks of burn-in is a reasonable plan, mostly spent replacing parts that fail early.
Then tie billing to acceptance, wave by wave. A worked example: 128 nodes handed over in four monthly waves of 32. If billing for all 128 starts at the first wave, you pay for 96, then 64, then 32 nodes you don't have yet: 192 node-months, or 1.12M GPU-hours. That's US$1.12M for every US$1.00 of rate, paid for nothing.
SLA: what counts as downtime
An availability percentage means little until you know what it's measured on. On a 256-GPU cluster, 99.9% averaged across the cluster is 187 GPU-hours a month, so one node can sit dead for 23 hours every month without a breach. Measured as "the cluster is reachable", a dead node may not count at all. A synchronous training job stops when one node does, so the per-node definition is the one that matches your losses. Our post on GPU cluster SLAs and goodput goes through the difference in detail.
Define "down" to include partial failures: a GPU that has fallen off the bus or keeps throwing memory errors, a dead NIC on a rail-optimized fabric, links running degraded, a storage mount that hangs. Get scheduled maintenance windows capped, announced in advance and counted against the SLA beyond the cap.
Node replacement time matters more than the percentage
Failures are routine at this scale. During Llama 3 405B pre-training on up to 16,384 H100s, Meta (opens in a new tab) logged 419 unexpected interruptions in 54 days, about 78% of them hardware, with faulty GPUs (148) and HBM3 memory (72) the largest causes. That works out to about 0.135 hardware interruptions per GPU per year, or about 35 a year on 256 GPUs, the same rate our cost model uses. Nobody has published comparable data for Blackwell yet.
At that rate, how fast a failed node is restored decides how many GPU-hours you lose. Our estimate, for 256 GPUs and about three node failures a month:
| Time to restore a failed node | GPU-hours lost per month | Share of 186,880 |
|---|---|---|
| 4 hours (hot spare in the rack) | about 90 | 0.05% |
| 24 hours | about 550 | 0.3% |
| 72 hours | about 1,660 | 0.9% |
| 7 days (part shipped under RMA) | about 3,870 | 2.1% |
The difference between the first and last rows is 2% of the cluster, every month, for five years. Ask for a restore time, not a response time, and for spares on site.
Credits and remedies
Credits are usually a reduced fee on a sliding scale, often with an overall cap. UNCITRAL's notes on cloud contracts point out that making credits "the sole and exclusive remedy" may "limit the customer's rights to other remedies, including suing for damages or terminating the contract" (UNCITRAL (opens in a new tab)). You won't get uncapped damages for GPU downtime, and you shouldn't expect to. Do get credits calculated on the affected capacity, and a right to terminate the affected nodes, or the contract, after repeated breaches in a rolling window.
Prepayment and how it is protected
Long-term reservations are commonly priced lower when part of the contract is paid up front. That's a fair trade, but a prepayment is also the largest unsecured exposure in the contract if it isn't protected. Four things to check:
- Where it goes. An independent escrow agent, not the provider's operating account.
- When it's released. In stages, against milestones you can verify, with the evidence for each release request sent to you.
- How it comes back. As a credit against invoices once the capacity is running, so it's used up over the term rather than sitting as a balance you have to chase.
- What happens in each exit route. Late delivery, early exit, and provider default each need an answer.
| Scenario | What should happen | Lightyard |
|---|---|---|
| Signed, nothing delivered | The money sits in escrow | Paid into escrow with an independent escrow agent |
| Milestones met | Released in stages, against evidence | Released in stages as supply, installation and acceptance milestones are met |
| Capacity running | Credited against invoices | Credited in equal parts against each monthly invoice over the term, from the first full month |
| Capacity late | Refund for the late part | Agreed in the order form |
| You exit early | Known in advance | An exit option, on terms set in the proposal |
| Security deposit on top | None | None |
Exit rights and termination fees
Most reserved contracts run to term. Where early termination exists, the fee is often close to the remaining payments, which makes it a payment schedule, not an exit. UNCITRAL notes that providers' standard terms for multi-tenant cloud services "usually reserve the right of the provider to terminate the contract at any time without customer default" (UNCITRAL (opens in a new tab)); check that yours doesn't, outside clear causes such as non-payment.
Termination for cause should run both ways. Your list: missed delivery beyond a long-stop date, repeated SLA failure, a security breach involving your data, and a change of control you can't accept.
An exit option changes the size of the bet: the firm commitment is the early part of the term, and the rest of the 60 months becomes a decision you make later.
Hardware refresh
NVIDIA has said it runs on a "one-year rhythm" (NVIDIA (opens in a new tab)), so a 60-month term will see three or four newer generations. Clifford Chance expects agreements to "embed clear mechanisms for chip refresh, upgrades and technology substitution over multi-year terms". Read any refresh clause carefully: a swap to new hardware mid-term usually resets the economics, so it tends to come with a higher rate, a longer term, or both.
The practical defences are simpler. Split the commitment across generations, keep an exit point where the next generation is likely to make sense, and plan to move older capacity to work that still pays for it, such as smaller models, batch inference and evals. Our post on GB300 or waiting for Vera Rubin runs the numbers. On Lightyard, part of the commitment can roll into a newer GPU generation when it ships, and the exit option is a second route.
Placement, region and moves within a region
The order form should name the region and the facility tier. Data residency, latency to your other systems and your auditors' site visits all depend on it. Then cover moves: a provider may need to consolidate halls during a five-year term. Ask that any move needs your consent, comes with long notice, stays inside the region and any residency boundary you've set, is paid for by the provider, and has a cap on downtime that counts against the SLA.
Data handling
On bare metal your data lives on the provider's drives and passes over their fabric, so the paper needs to cover:
- A data processing agreement, and a BAA if you'll hold health data.
- Who can reach the nodes, how that access is logged, and when the provider's out-of-band access is used.
- What happens to failed drives. Ask that they're wiped or destroyed on site to a named standard, such as NIST SP 800-88 Rev. 2 (opens in a new tab), rather than shipped back to a vendor intact.
- End of term: an export window, then deletion with an attestation. UNCITRAL suggests the customer "will be served with an attestation, report or statement of data deletion" (UNCITRAL (opens in a new tab)).
Capacity expansion
If you expect to grow from 32 to 128 nodes, the time to say so is before the fabric is designed. Adding nodes to a fabric built for 32 can mean new spine switches or a second fabric your jobs can't span. Ask for the fabric to be sized for your target size, a stated way to price added nodes, and whether added nodes co-term with the original contract or start their own term.
If the provider fails
Insolvency clauses don't do what they appear to. Under the US Bankruptcy Code, a trustee "may assume or reject any executory contract", and a contract generally "may not be terminated or modified, at any time after the commencement of the case solely because of" a provision tied to insolvency (11 U.S.C. 365 (opens in a new tab)). UNCITRAL (opens in a new tab) makes the same point more generally: "Mandatory provisions of insolvency law may override those clauses." SLA credits owed by an insolvent provider are worth little.
So protect what you can before anything goes wrong:
- Keep undelivered money in escrow, released only against milestones. That's the largest exposure in most contracts.
- Get information rights: notice of a change of control, and of the loss of a facility or power contract.
- Agree a data export window that holds even if the contract ends, and keep your own copies of checkpoints and weights outside the cluster.
- Ask who would step in. Clifford Chance notes that lenders want "step-in or controlled termination in distress"; ask whether that would keep your capacity running.
Reserving with Lightyard
Lightyard's contract is a master services agreement with an order form under it, for a single-tenant cluster on bare metal or managed, reserved for 60 months at a per-GPU-hour rate fixed for the term and run to an SLA. The terms are set out under contracts. If you're comparing agreements now, send us yours with the cluster you need, and we'll reply with a rate, a region and a delivery date counted from signing.
Questions this note answers
Usually a multi-year contract for dedicated GPU capacity: a set number of nodes or racks, reserved for one customer at a fixed rate per GPU-hour and run by the provider to an SLA. A true equipment lease is different: you take the hardware, and under US commercial law your payments in a finance lease become irrevocable once you accept the goods, even if they later fail.
It is the same kind of long-term capacity contract seen from the provider's lenders' side. GPU financings are increasingly sized against contracted revenue, so the lenders read the payment, termination and SLA terms closely. For the buyer, the checklist is the same as for any reservation.
Availability measured per node, not as one cluster-wide figure, with a GPU that has dropped off the bus, a failed NIC or degraded fabric links counted as down. Add a time to restore a failed node, credits on a sliding scale, and a right to terminate for repeated failures, since credits alone are often capped and the sole remedy.
Only if the contract says so. Most reserved contracts run to term or charge a termination fee close to the remaining payments. At Lightyard, the contract includes an exit option, on terms set out in your proposal.
Held by an independent escrow agent, released in stages only as supply, installation and acceptance milestones are met, credited against invoices once the capacity is running, and refundable for capacity that is not delivered. Also check what happens to prepayment not yet credited if you exit early.
Sources
- Clifford Chance: Data centres and AI compute infrastructure, insights 2026 (opens in a new tab)
- Cornell LII: UCC 2A-407, irrevocable promises in finance leases (opens in a new tab)
- PwC Viewpoint: Leases guide, definition of a lease (ASC 842) (opens in a new tab)
- UNCITRAL: Notes on cloud computing contracts, remedies (opens in a new tab)
- UNCITRAL: Notes on cloud computing contracts, term and termination (opens in a new tab)
- UNCITRAL: Notes on cloud computing contracts, end-of-service commitments (opens in a new tab)
- Cornell LII: 11 U.S. Code 365, executory contracts and unexpired leases (opens in a new tab)
- NIST SP 800-88 Rev. 2: Guidelines for Media Sanitization (September 2025) (opens in a new tab)
- Meta: The Llama 3 Herd of Models (reliability section) (opens in a new tab)
- NVIDIA: Computex 2024 keynote, one-year rhythm (opens in a new tab)