You signed a multi-year GPU reservation, prepaid a slice of it, and assumed the contract was written to protect you: the step-in rights, the offtaker language, the exit clause. Now the provider’s financing has come apart, and you are asking one question: what actually happens to my workloads and my data, and in what order?
This is a continuity question. On enforcement, the lender preserves the cash flows pledged against the loan, and your contract is one of those cash flows. That is the mechanism behind what failure looks like in practice, and the sequence below is what it produces.
The trigger is the maturity wall that precedes enforcement, which arrives before any formal filing. What that means for your workloads and your data, in order, is what the three sections below walk through.
What should you look for in step-in, assignment and change-of-control clauses if a lender enforces?
On enforcement, read step-in rights for who may step in, how quickly, and whether service levels survive the transfer. Check whether the provider can assign the contract without your consent, and whether a change of ownership triggers your termination right.
These clauses read this way because your contract is part of the lender’s security package: the revenue it generates is what repays the debt. Read the failure-allocation clauses and you are reading the lender’s asset map.
Group them by what they allocate. Control: step-in and transition rights, assignment, change of control. Money: prepayment, most-favoured-customer and repricing, SLA credits. Data: egress, escrow of GPU inventory. Time: termination for convenience, transition windows, cross-default. All four groups add up to counterparty risk in a multi-year inference commitment.
A step-in right with no service-level continuity only entitles the transferee to keep billing you; nothing in it obliges the transferee to keep you running.
Assignment and change of control are where buyers most often miss a lever. Change-of-control clauses in AI compute contracts commonly treat a change of ownership as an assignment requiring consent. The question that protects you is the reverse: can the provider assign your contract to a third party, including a competitor, without your consent, and does a change in the provider’s ownership trigger your termination right?
Escrow of GPU inventory is where a claim becomes an asset. Where your prepayment bought physical capacity, escrow or title retention converts it into something you can point at; without it, you are an unsecured creditor of a bankrupt operator. In Australia that title interest has to be registered on the PPSR or it ranks as unsecured.
Continuity cannot be read from the contract alone; it sits inside the clauses in their credit context. The public record on neocloud clause sets is thin, so this section reasons from credit structures.
Does “investment-grade offtaker” language protect your continuity, or only the lender’s repayment?
“Investment-grade offtaker” identifies the creditworthy customer whose take-or-pay contract supports the lender’s borrowing base. It is a lender’s representation about revenue quality. It protects the repayment schedule. Nothing in the definition obliges the lender to keep serving you, and your remedies sit in the service contract.
Lambda’s $926 million term loan B, at SOFR plus 300, rated Baa2 and amortising to 2030, was justified by “an investment-grade offtaker”. The term was written to protect the lender’s repayment schedule, and the rating agencies are pricing the offtaker’s balance sheet.
The asymmetry is worth stating plainly. A take-or-pay commitment requires you to pay whether or not capacity is delivered, while the provider’s delivery obligation is capped by service credits that are themselves an unsecured claim in an insolvency. You underwrite the provider.
There is one point where your interests align. A lender holding a functioning offtaker contract has an incentive to keep the cluster running, because an interrupted contract impairs the collateral. That is an incentive, and it lasts only while the contract is worth more running than enforced.
The prepayment shows the asymmetry most directly. Offtakers prepay 15 to 25 per cent of contract value, a first-loss cushion at the bottom of the stack. Without escrow, title retention or a refund-on-insolvency term, the prepayment is gone at enforcement. It is credit support you provided with no matching protection, which is why the prepayment needs contractual protection.
If your GPU provider fails mid-contract, what actually happens to your workloads and your data?
A missed refinancing or covenant breach cross-defaults the facilities; the lender or administrator takes control and operations stop. Reserved capacity stops being schedulable, in-flight runs die at the last checkpoint, and your data sits on the provider’s storage until the egress terms and the administrator decide otherwise.
Walk the sequence in order. The lender enforces, or an administrator is appointed. Operations stop: staff, power contracts, supplier terms and NVIDIA allocation all halt at once. Notice comes last, if at all, while the status page still says “investigating”. The chain runs from GPU depreciation through tenant default to data-centre SPV distress, and the standstill is what kills workloads.
Weights, checkpoints, datasets and logs sit on the provider’s storage, and a bankrupt provider may have no funds, no personnel and no interest in performing. Contractual approaches to data recovery then have little effect in insolvency. Data gravity means the exit is measured in days to weeks.
Portability is decided by egress and architecture. A contractual right to leave is worthless against 40TB of checkpoints and a powered-down cross-connect. At a throttled 1 Gbps, moving that volume takes roughly 89 hours; on a 10 Gbps pipe the floor is about nine hours. The clause says you may leave. The arithmetic decides.
The same question sharpens in Australia. Beneath a foreign lender and a foreign governing law, can step-in and transition rights be exercised locally, and which jurisdiction’s administrator answers the phone? In the US, a Chapter 11 stay means the GPUs stay in the racks and the lender cannot touch them. Australian voluntary administration and receivership look different, and the PPSR decides who ranks on the hardware.
That points to the alternative worth weighing first: shorten the term rather than negotiate clauses. A shorter commitment is optionality, and no enforcement can take that back.
The contract was the lender’s cash-flow-preservation document, and you, prepayment, take-or-pay obligation and all, are one of the assets it preserves. Your real exit was always an architecture decision: the checkpoints you can copy out at a rate you can afford outweigh any termination right.
Failure-allocation clauses tell you who controls what. “Investment-grade offtaker” is the lender’s language, and a shorter term beats a better clause.
Frequently Asked Questions
What should I do in the first 24 hours after my GPU provider fails?
Treat it as an incident, not a negotiation. Freeze new writes, export the smallest set of artefacts you cannot recreate (latest weights, tokeniser, evaluation sets), rotate credentials and preserve logs of what was running. Notify your counsel and check whether the administrator has published a data-access process. Do not wait for a formal notice, and do not assume the storage layer will still be readable tomorrow.
Will I get notice before the platform goes dark?
Assume you will not. In practice, customers usually learn from a status page still saying “investigating”, a failed API call or a missed support ticket, not a contractual notice letter. An administrator’s first obligations run to creditors and to keeping the estate intact, and a lender enforcing security has no duty to brief tenants. A 30-day notice clause is worth little once the power and staff have stopped.
Does a bankruptcy or administration filing stop me from accessing my own data?
It can, at least temporarily. In a US Chapter 11, the automatic stay blocks actions against the debtor, and if the data is treated as property of the estate you need court relief to retrieve it. Australian voluntary administration brings an ipso facto stay that restricts terminating a contract over the administrator’s appointment. Contract language saying you own your data helps, but it does not guarantee access on demand.
How long does it take to move 40TB of checkpoints off a failing provider?
Longer than the notice you will get. At a sustained 1 Gbps, 40TB takes roughly 89 hours before protocol overhead, and a throttled or shared egress pipe will be slower. At 10 Gbps the theoretical floor is about nine hours, which still assumes the cross-connect is up, the storage is readable and nobody is competing for the same pipe. Plan for days.
What will egress actually cost during an emergency exit?
Treat it as a ransom, not a line item. At $0.09 per GB, 40TB costs about $3,700 before any premium for priority or restored bandwidth, and rates vary widely by provider. The bigger cost is time: an administrator may insist on payment before restoring a throttled pipe, so a zero-fee disaster egress clause and a pre-agreed transfer window matter more than the headline rate.
What happens to my prepayment if the provider fails?
It usually becomes an unsecured claim. If you prepaid 15 to 25 per cent of contract value and hold no escrow, title retention or refund-on-insolvency term, you rank behind the secured lender and recover cents in the dollar, often after years. The prepayment was credit support you gave the provider. Without matching security, it is the first money lost and the last returned, and your remedy is an insolvency claim, not a route to the front of the queue.
Can I stop paying or set off my losses once the provider defaults?
Not automatically. A take-or-pay obligation survives non-delivery, and a set-off right is only as good as the contract clause that creates it. Administrators routinely pursue unpaid invoices as assets of the estate, while your service credits remain an unsecured claim. Suspend only where the contract expressly permits it, and get legal advice before withholding payment, because a wrongful suspension can hand the administrator a cross-claim.
What is cross-default, and why does it matter to a GPU customer?
Cross-default is the clause that lets one missed payment or covenant breach trigger default across every facility the provider has. It compresses the timeline: a breach in a single lender’s agreement can pull the whole capital structure into enforcement at once, so the operational standstill arrives sooner than any single facility’s maturity suggests. Read it as the mechanism that turns a refinancing problem into a platform outage.
Can my data be sold as an asset in the provider’s insolvency?
It can be claimed, which is why ownership language matters. If the provider’s terms are silent, an administrator or trustee may treat stored data as part of the estate and seek to sell it with the business, subject to privacy law and any court supervision. A clause stating that you retain all right, title and interest, and that the provider holds the data only as bailee, gives you a stronger argument.
Is a hyperscaler actually safer than a neocloud?
Safer as a counterparty, not necessarily safer as an exit. A hyperscaler carries investment-grade credit and a balance sheet built to survive the cycle, but its egress fees, proprietary storage and rigid contract terms can make leaving just as slow. The relevant question is not who is bigger. It is whether your checkpoints sit somewhere you can copy them out of at a rate you can afford.
What is the difference between voluntary administration and receivership in Australia?
They are different control regimes. In voluntary administration, an independent administrator takes control of the company, the ipso facto stay limits termination for the appointment itself, and creditors decide the company’s fate. In receivership, a secured lender appoints a receiver to realise the assets, with no ipso facto stay and no duty to keep your cluster running. Which one applies decides who answers the phone.
What early warning signs suggest my GPU provider is heading for failure?
Watch the financing, not the marketing. A refinancing wall inside your contract term, covenant waivers, supplier disputes, stalled capex and sudden changes in NVIDIA allocation all matter more than utilisation claims. The provider’s own lenders are pricing the same signals. If the next maturity is close and the debt is trading at a distressed yield, assume the operational standstill is already on the table.
