GPU capacity
Five clauses that quietly cost you in GPU-capacity agreements
Five clauses that quietly move money in GPU-capacity agreements, and what to negotiate on both sides.
LegalBooks Legal Team
Attorneys licensed in the US (California & Nevada) and Canada (Ontario)
Aug 5, 2026 · 9 min
Five clauses quietly move money in GPU-capacity agreements: service credits framed as your sole remedy with no exit from chronic downtime; GPUs that the provider doesn't own free and clear; open-ended power-price pass-throughs; a soft default exit when the buyer stops paying; and payment security taken - or given - up front. Each favors whoever drafted it.
A GPU-capacity agreement is a contract to reserve or supply GPU compute at scale - through colocation, bare-metal rental, or compute-as-a-service - usually billed as a monthly recurring charge (MRC) over a fixed term.
GPU compute is capital-intensive in a way CPU compute never was. A model company that raises $20 million USD can commit the entire raise to a single month of GPU MRC, where a traditional software startup might spread that spend over years. That asymmetry changes the stakes on both sides of the table: the buyer is exposed if the cluster underperforms, and the seller is exposed if the buyer cannot keep paying. The five clauses below are where that exposure hides.
This article is written for operators on either side of the deal - buyers reserving capacity and providers supplying it. Each clause notes who it usually bites and what to negotiate.
Clause 1: Service credits as your "sole and exclusive remedy"
In most GPU-capacity contracts, service credits are the "sole and exclusive remedy" for downtime - which means a credit worth a few percent of a monthly invoice is the most you can collect, even when a failed cluster costs you far more. The clause that actually protects a buyer is not a bigger credit. It is a right to leave when the outages do not stop.
Standard credit ladders are modest and come with conditions. Credits are usually issued as future account credit rather than cash, often expire if unused within about 90 days, and carry short claim windows - some providers require written notice within 24 hours or the credit is forfeited (Spheron, 2026). Uptime is also measured at whichever layer favors the provider: node-level guarantees near 99% sit alongside rack-level figures closer to 95%, and "a node going down for six minutes and a rack losing power for six hours can both get reported as one incident" (Spheron, 2026).
Sole-remedy comparison - published SLA credit ladders
| Provider | Stated SLA | Credit tiers | Form of credit |
|---|---|---|---|
| AWS EC2 (single instance) | 99.5% | 10% / 30% / 100% | Future account credit |
| Oracle OCI | 99.99% | 10% / 25% / 100% | Future account credit |
| DigitalOcean | 99% | 10% / 25% / 50% | Future account credit |
Source: Spheron GPU Cloud SLA Guarantees, 2026.
The quiet cost is the "sole and exclusive remedy" language itself. It converts every failure into a capped credit and forecloses any other path - including the door out. One analysis puts the cost of downtime at roughly $23,750 USD per minute for a large organization, a figure no standard credit approaches (Compute Law Blog, 2026). Reviews of current GPU-cloud SLAs find that almost none include a right to terminate for chronic or repeated failure.
What to negotiate (buyer side): a chronic-outage exit. Add a termination right triggered by persistent breach - for example, a set number of failed SLA months in a rolling window, or a hard availability floor over a quarter - carved out of the sole-remedy clause, with a refund of prepaid and unused fees on the way out. If the provider cannot fix constant downtime, service credits are not a remedy; they are a subscription to a broken service.
Clause 2: Who actually owns the GPUs?
Before you sign a GPU-capacity deal, confirm the provider owns the GPUs free and clear - many do not. The hardware is frequently pledged as collateral to the lender that financed it, and a provider default can put a financier's repossession rights ahead of your access to the cluster you are paying for.
GPUs depreciate quickly, so operators finance against them aggressively. In August 2023, CoreWeave borrowed $2.3 billion USD using Nvidia H100s as collateral in a facility led by Magnetar Capital and Blackstone - reportedly the first time H100-based hardware was used as collateral at that scale (Quartz, 2026). When hardware secures a loan, the operator does not hold it free and clear: if the operator defaults on its debt, the secured lender - not the customer - can control the equipment mid-term.
What to run down (buyer side), on both sides of the border:
Lien search. A UCC-1 search in the US, or a PPSA (Personal Property Security Act) search in Canada, against the operator and the specific equipment, to see who already holds a security interest.
Title representation. A representation and warranty that the provider owns the hardware and that there are no undisclosed encumbrances, with notice obligations if that changes.
Lender non-disturbance. A recognition or non-disturbance agreement from any secured lender, so a foreclosing financier honors your access rather than pulling the GPUs.
Step-in and continuity rights. Step-in rights to the underlying colocation lease on provider default, plus a right to relocate or continue using the hardware if the provider fails.
This is the clause most buyers do not know to look for, and the one most likely to detonate at the worst moment - when the provider is already in trouble.
Clause 3: Who underwrites the power-price risk?
In power-heavy deals, the clause that quietly moves the most money is the one that decides who absorbs rising electricity costs. An uncapped pass-through hands the buyer an open-ended bill; a fixed price with no escalator hands the seller the risk. Neither is right or wrong - but someone is underwriting power, and the contract should say who.
Colocation and large compute deals are increasingly priced on power (dollars per kilowatt), and the underlying electricity supply may be a pass-through, a fixed tariff, or a power purchase agreement. Fuel, utility, and electricity cost-escalation clauses are standard contract language, which means the mechanism is negotiable rather than fixed (Law Insider clause library).
What to verify and negotiate (both sides):
Read the underlying supply contract. Confirm its term, its pricing certainty, and whether the provider has actually locked power for the length of your deal - not for this year alone.
Decide who underwrites an increase. Negotiate a cap or collar on escalation, tie any escalator to a named public index rather than the provider's discretion, and require transparency into how power is calculated and billed.
Watch how power is measured. In colocation specifically, whether you are billed on total facility load or on IT load changes the number materially. That measurement question is deep enough to deserve its own treatment - see our pre-signing checklist for a colocation MSA.
Clause 4: What happens when the buyer can't keep paying?
The seller's largest exposure in a GPU-capacity deal is a buyer who stops paying the MRC. If the exit for non-payment is slow or soft, the provider is left carrying idle, financed hardware while an insolvent customer sits on reserved capacity. Take-or-pay language secures revenue on paper - but against an insolvent buyer, take-or-pay is only an unsecured claim.
The capital intensity that makes these deals large also makes default a real scenario: a buyer can outrun its runway in months, not years. Reported offtake deals lean heavily toward the provider on paper - a breaching customer that fails to cure typically owes the full remaining balance for the term (American Compute, 2026) - but paper protection is worth little if the counterparty is gone.
What to build (seller side):
A clean default exit. Prompt suspension and then termination on non-payment, with a short, defined cure period; acceleration of the remaining term; and no-offset, no-withholding language so a service dispute cannot be used to stop payment.
Repossession and access mechanics. Clear rights to deny access, reclaim capacity, and re-let it, so the provider is not frozen while the meter runs.
Force majeure that says who pays. Payment treatment during an outage varies widely across reported MSAs - from "payments delayed but not excused" to a customer right to exit after 30 consecutive days (American Compute, 2026). Decide it explicitly rather than inheriting a template.
The real protection, though, is upstream: verify the buyer's committed funds before signing, which is where Clause 5 comes in. A buyer should push the mirror image of this clause - the chronic-outage exit from Clause 1.
Clause 5: Payment security up front
Payment security up front is where the counterparty risk actually gets priced. Providers protect themselves with deposits, prepayment, letters of credit, and parent guarantees; buyers protect themselves by refusing to pay for capacity that has not passed acceptance testing. Both belong in the same clause, and both sides should expect to give something.
Up-front security in reported deals is substantial. Structures have included roughly 30% of contract value paid before a single GPU-hour is delivered, and nine-figure letters of credit posted as customer security on top of monthly advance payments (American Compute, 2026). For a leveraged provider, buyers on the other side negotiate parent-company guarantees, standby letters of credit, and step-in rights to the provider's leases on default (Compute Law Blog, 2026).
What to negotiate:
Seller side. Deposits or prepayment, a standby letter of credit or parent/sponsor guarantee sized to the exposure, and security that survives a buyer's financial trouble.
Buyer side. Tie payment obligations to an acceptance and burn-in gate, so the meter starts when the cluster passes node and fabric checks - not on a paper delivery date. Scope force majeure so you are not prepaying straight through an outage.
The five clauses at a glance
| Clause | Who it usually favors as drafted | What to negotiate |
|---|---|---|
| Service credits as sole remedy | Seller | A chronic-outage termination right and refund of unused fees |
| GPU ownership / security interests | Seller (and its lender) | Lien search, title rep, lender non-disturbance, step-in rights |
| Power-price risk | Whoever didn't draft it | Cap or collar, named index, transparency into power billing |
| Buyer-default exit | Seller on paper, no one in insolvency | Clean suspension/termination, acceleration, repossession mechanics |
| Payment security up front | Seller | Deposits, LC, guarantees - vs. an acceptance/burn-in gate for the buyer |
Key takeaways
Service credits are a cap, not a cure. The "sole and exclusive remedy" line is the quiet cost; the fix is a chronic-outage exit, not a bigger credit.
Confirm the provider owns the GPUs. Financed hardware can carry a lender's security interest that outranks your access - run a UCC-1 or PPSA search before signing.
Name who underwrites power. An uncapped pass-through and a fixed price with no escalator both move real money; decide the mechanism deliberately.
Sellers need a clean default exit. Take-or-pay is only as good as the buyer's solvency, so pair it with fast termination rights and up-front security.
Payment security cuts both ways. Providers secure revenue with deposits and letters of credit; buyers secure value with an acceptance and burn-in gate.
Frequently asked questions
What is a GPU-capacity agreement?
A GPU-capacity agreement is a contract to reserve or supply GPU compute at scale - through colocation, bare-metal rental, or compute-as-a-service - usually billed as a monthly recurring charge over a fixed term.
Are service credits really the only remedy for GPU downtime?
In most GPU-capacity contracts, yes. Service credits are typically labeled the 'sole and exclusive remedy,' capping recovery at a small share of the monthly fee and often paid as future account credit rather than cash. The protection worth negotiating is a chronic-outage termination right carved out of that clause.
How do I confirm a GPU provider owns its GPUs free and clear?
Run a UCC-1 lien search in the US or a PPSA search in Canada against the operator and the specific equipment, require a title-and-no-undisclosed-encumbrances representation, and negotiate a lender non-disturbance agreement plus step-in rights so a financier's security interest cannot cut off your access mid-term.
Who pays when electricity prices rise in a colocation or GPU deal?
It depends on the pricing clause. An uncapped pass-through puts rising power costs on the buyer; a fixed price with no escalator puts the risk on the seller. Verify what the underlying electricity supply contract says, then negotiate a cap, a defined index, and transparency into how power is billed.
What protects a compute provider if the customer can't keep paying?
A clean default exit: prompt suspension and termination on non-payment with a short cure period, acceleration of the remaining term, no-offset language, and clear repossession or denial-of-access mechanics - backed by up-front payment security and diligence on the buyer's committed funds before signing.
Move fast without missing these clauses
Every idle GPU hour spent waiting on redlines has a cost, and every one of the clauses above is easier to fix before signing than after. LegalLayer is a tech-native legal team built for compute infrastructure: in 2026 we have negotiated more than $266 million USD in client contract value; data-center work makes up the bulk, and drafts and redlines are back within six hours, guaranteed. We work both sides - buyers reserving capacity and providers supplying it - as technical peers, licensed in the US (California and Nevada) and Canada (Ontario).
This article is general information, not legal advice. For advice on your specific situation, consult a qualified attorney. LegalBooks provides counsel for compute infrastructure deals across the US and Canada.