Your GPU Prepayment Is an Unsecured Loan: Six Ways B300 and GB300 Offtakers Lose $20 Million, and the Documents That Stop Each One

The sequence below is a composite. Each step is a failure LegalLayer sees offtakers exposed to in GPU capacity MSAs, and each legal consequence is how US commercial and bankruptcy law ordinarily works.

1. Day 0. You sign a 3-year MSA for 512 B300 GPUs at $5 per GPU-hour, a $67.3 million contract, and wire a 30% prepayment of $20.2 million. You sign your own inference customers at 99.49% uptime.
2. Month 3. The GPUs are not live. Your provider used your deposit to place its GPU order and is still waiting on energized colocation space. Your customer launches slip.
3. Month 7. The cluster is live. Your provider's largest customer stops paying, and your provider misses a payment on its GPU loan.
4. Month 8. The lender declares a default. In this composite, the GPU loan is secured by all of the provider's GPUs, including the 64 nodes serving you, and no lender has signed a non-disturbance agreement. Under UCC § 9-609, a secured lender may take possession of its collateral after default, without going to court if it does so without a breach of the peace. The lender never signed your MSA, so your MSA does not bind it.
5. Month 8, the same day. Your inference goes dark. A 99.49% SLA allows about 44.7 hours of downtime a year. A repossession uses that up in under two days, and every hour after that is a service credit you owe your own customers.
6. Month 9. Your provider files for Chapter 11 and asks the court to reject your MSA. The prepayment you have not used becomes a general unsecured claim, paid after the secured lender that already took the GPUs.

Every step in that sequence could have been caught by a document the offtaker could have demanded before wiring. The rest of this guide takes each failure in turn.

Prepayments in GPU capacity contracts are eight-figure sums, and public deals show the range. IREN's $9.7 billion, 5-year GB300 contract with Microsoft includes a 20% prepayment. Soluna's 36-month, $34.0 million GPU services agreement with HPE required $10.3 million up front, about 30%. Axe Compute's $260 million, 36-month contract for 2,304 B300 GPUs is secured with a deposit, a prepayment and monthly payments in advance. In the GPU capacity MSAs LegalLayer works on, buyers are typically asked for 20% to 30%, and sometimes 40%.

A bank lending $20 million would demand collateral, financial statements, a lien search and covenants. Most offtakers wire the same amount against a sales deck and an MSA. Underwrite the provider the way a lender would, because economically that is what you are.

Some providers sell capacity they do not have. They sign the MSA, take your deposit and use it to place the GPU order, then still have to receive the hardware, find colocation space with enough power, and bring up the cluster. Each step can slip, and a provider without experience running large clusters slips more often. NVIDIA describes the GB300 NVL72 as a "fully liquid-cooled, rack-scale architecture" with 72 Blackwell Ultra GPUs per rack, which needs a liquid-cooled facility and a team that has installed it before.

Delays reach the top of the market. In November 2025, CoreWeave cut its revenue forecast after delays with a top data center partner, and the affected customer agreed to extend its contract (Reuters). A smaller provider holding your deposit and no hardware has less room to recover, and every month it slips is a month of your customers' revenue you do not book.

Demand before you wire: the OEM purchase order or delivery confirmation, a serial-number schedule for the nodes assigned to you, the signed colocation agreement for energized space, and named references for a comparable cluster the provider has brought up. Make delivery by a fixed date an obligation, with an outside date that lets you terminate and recover the full prepayment.

The company that signs your MSA is often not the company that owns the GPUs. A typical chain looks like this:

• You, buying capacity to serve your own customers
• A broker or reseller, who signed your MSA
• The neocloud operator, who runs the cluster
• The GPU financier, who may hold title to the GPUs or a first-priority lien on them
• The colocation provider, who supplies space, power and cooling
• The colocation landlord and its lender, who can foreclose on the building
• The utility, which supplies the power

Your MSA binds only the party who signed it. If the reseller collapses, the operator owes you nothing under your MSA and can resell your capacity to the next buyer. If the operator defaults to its financier, the financier never agreed to anything with you. Each layer is another company whose failure ends your access, and another margin on your price.

Demand before you wire: the bill of sale or financing documents showing your counterparty owns the GPUs or is the borrower on them, and written disclosure of every company between you and the hardware. Prefer the owner-operator. If you must buy through a reseller, require a direct agreement with the operator so your capacity survives the reseller, and SLAs that flow back-to-back down the chain.

Take a GPU capacity MSA with a 99.49% uptime commitment as an example. It allows about 44.7 hours of downtime a year, roughly 3.7 hours a month, before the provider owes you anything. For comparison, AWS commits to 99.5% for a single EC2 instance and DigitalOcean to 99% for GPU Droplets, according to Spheron's 2026 survey of GPU cloud SLAs. When the provider misses, the usual remedy is a service credit against a future bill, not cash, and it does not cover what you owe your own customers.

The power, cooling and network your provider's SLA depends on live in the colocation facility. A GB300 NVL72 rack is liquid-cooled as a single system of 72 GPUs, so a cooling fault can take a whole rack offline at once. If the provider's colocation contract gives it a weaker SLA than the one it gave you, the gap is your downtime.

Demand before you wire: the colocation agreement, with a term longer than your MSA; the power and cooling redundancy serving your racks; the facility's outage history; the colocation SLA your provider receives, which must be at least as strong as yours; and a non-disturbance agreement from the colocation landlord's lender. Our guide to colocation red flags for AI infrastructure covers the full 14 checks.

In LegalLayer's experience, many small and mid-sized neoclouds finance GPUs with debt secured by the GPUs themselves, often across several lenders and entities, so one large customer default can push the provider into a liquidity crisis. Whether its lender can then take the GPUs serving you depends on how they were financed:

• Owned outright with no lien: no lender has a right to the GPUs, though they remain the provider's assets in its bankruptcy.
• Financed with a loan secured by the GPUs: after default, UCC § 9-609 lets a secured party "take possession of the collateral" through a court, or without one if it proceeds without a breach of the peace.
• Leased from a financier that holds title: on default, the financier can generally recover its equipment under the lease.
• A blanket lien with cross-defaults: a default on another customer's cluster can reach the nodes serving you. A lien limited to your cluster in its own SPV cannot.
• A lender that signed a non-disturbance agreement: the lender has agreed not to disturb you while you keep paying.
• A provider in Chapter 11: the automatic stay pauses repossession until the court lifts it.

Your MSA does not change that list. Most GPU capacity MSAs are services contracts, not equipment leases. The UCC rule that lets an ordinary-course lessee of goods keep its lease free of the lessor's lender, § 9-321(c), protects a leasehold interest, which a services customer typically does not have.

Demand before you wire: a UCC-1 search (PPSA in Canada) against the provider and every entity holding the GPUs, the default and remedy sections of the GPU financing, and a non-disturbance agreement signed by every GPU lender. If a lender will not sign one, you are prepaying for GPUs that lender can take.

Bankruptcy stops the repossession clock, not the loss. A Chapter 11 filing triggers the automatic stay, but the lender can ask the court to lift it, and the provider can ask the court to reject your MSA. Rejection is treated as a breach, and the unused prepayment generally becomes a general unsecured claim, paid from whatever is left after secured creditors.

Compute North, a bitcoin mining host, shows how this plays out. It was Marathon's largest hosting partner. When Compute North filed for Chapter 11 in September 2022, Marathon disclosed $81.3 million of exposure: $50 million in operating deposits and prepayments, mainly for the King Mountain and Wolf Hollow sites in Texas, a $21.3 million unsecured promissory note and $10 million of preferred stock.

The secured lender took the sites. Generate Capital, which Compute North owed about $101 million, was the only bidder for the Wolf Hollow site in Granbury, Texas and the Kearney, Nebraska site, and bought them for $5 million plus the debt it was already owed. The whole asset sale raised $14.7 million in cash, against about $39.4 million of unsecured debt.

Marathon recorded a $55.7 million impairment on Compute North. In February 2023 it accepted a $40 million general unsecured claim in return for accepting the plan, with recoveries left to a litigation trust. We found no public report that Marathon has recovered any of it. In December 2023, Marathon paid Generate Capital $178.6 million in cash for the Granbury and Kearney sites, where 64 MW of its own miners were already running.

The secured lender ended up with the sites. The customer wrote off most of its deposits and then paid the lender to buy them.

Demand before you wire: the prepayment in escrow or backed by a standby letter of credit from a bank, released against hardware delivery and cluster acceptance and applied to the final months of the term, so the provider is never holding your money before it has earned it.

Service credits are the remedy for an ordinary outage. They do not let you leave. If the cluster underperforms month after month and credits are your only remedy, you keep paying the monthly fee, $1.87 million on a $67.3 million 3-year contract, collecting credits against future bills, while the provider holds the unused prepayment.

A material breach clause defines, in numbers, when the provider's failure lets you walk away with your money. In the MSAs LegalLayer negotiates for buyers, the triggers are:

• Chronic underperformance: availability below a negotiated threshold for 4 to 6 consecutive months.
• Severe outages: a single outage, or a set of outages, longer than a negotiated number of hours.
• Loss of the GPUs or the facility: repossession by a lender, loss of the colocation space, or any other failure that leaves the contracted GPUs unable to meet your service demands.
• Missed delivery: the GPUs are not delivered and accepted by the outside date.

Demand before you sign: those triggers written into the MSA, no billing for capacity that is not delivered, the right to terminate on notice when a trigger occurs, and a refund of the unused prepayment within a fixed number of days.

A bankruptcy-remote special purpose vehicle (SPV) is a single-purpose company that holds one deal's GPUs, GPU financing, colocation agreement and customer MSA, with separateness covenants and an independent manager designed to keep it out of its parent's bankruptcy. If the provider's other customers default or the parent files for Chapter 11, your GPUs and your MSA sit in a separate entity whose lender's collateral is your deal only. CoreWeave finances large customer contracts through single-purpose borrowers its CEO calls "the box".

A GPU non-disturbance agreement is a signed agreement in which the provider's GPU lender promises that, as long as you are not in default under your MSA, it will not take the GPUs away from you, it will honor your MSA or have a replacement operator honor it, and it will credit the prepayment you have already made. It works like the subordination, non-disturbance and attornment agreement a commercial tenant gets from its landlord's lender. It is not automatic, so it has to be negotiated before the provider's financing closes.

With these in place, a provider failure means a change of operator, not a lost cluster. You keep the GPUs, your prepayment is protected, and you keep the commitments you made to your own customers.

1. OEM purchase order or delivery confirmation for the GPUs, with a serial-number schedule for your nodes.
2. Bill of sale or GPU financing documents showing your counterparty owns the GPUs or is the borrower on them.
3. UCC-1 search results (PPSA in Canada) against the provider and every entity holding the GPUs.
4. A non-disturbance agreement signed by every GPU lender.
5. The colocation agreement, with a term longer than your MSA, the power and cooling redundancy serving your racks, the facility's outage history and the SLA your provider receives.
6. References for a comparable cluster the provider has brought up.
7. An org chart showing which entity signs your MSA, which owns the GPUs and which owes the debt, and the provider's revenue concentration in its largest customer.
8. An escrow agreement or standby letter of credit covering your prepayment, with release tied to delivery and acceptance.

Then write the protections into the MSA: a hard delivery date with a full-refund outside date, ownership and lien representations against the serial-number schedule, no resale or subcontracting without consent, a 99.49% SLA that flows back-to-back to the colocation facility, a material breach clause with a prepayment refund, no billing for undelivered capacity, and the right to move the contract to a replacement operator if the provider fails. A provider that intends to deliver will sign these. One that refuses is telling you how this ends.