Take-or-Pay Contract
Definition
A take-or-pay contract requires a customer to pay for a minimum committed amount of capacity, power, compute, or service whether or not the customer actually uses it. In AI infrastructure, take-or-pay economics can appear in GPU capacity reservations, data center leases, power contracts, managed compute agreements, and long-term hosting commitments.
Why it matters
Take-or-pay terms can convert volatile usage into contracted revenue, which is why lenders and infrastructure investors care about them. A GPU provider with spot-market demand has uncertain revenue; a provider with creditworthy customers committed to minimum monthly payments can finance hardware and facilities more easily. But take-or-pay is not magic. The revenue is only as strong as the customer's credit, termination rights, service-level protections, and willingness to keep paying if model economics, hardware generations, or AI demand change.
Common misconceptions
- •Take-or-pay reduces volume risk but does not eliminate counterparty credit risk.
- •A reservation is not always take-or-pay; cancellation rights and termination fees matter.
- •Committed revenue can still be exposed to service-level failures or force majeure provisions.
- •A contract headline value can overstate financeable revenue if ramp periods, free months, credits, or step-downs reduce early cash flow.
- •Take-or-pay does not protect against technology obsolescence if customers can migrate to newer hardware or renegotiate under commercial pressure.
Technical details
What Makes It Financeable
Minimum payment amount: The contract should specify a real floor, not merely an expected usage level.
Term: A three-year commitment supports more leverage than a three-month reservation, especially when hardware payback depends on long utilization.
Counterparty credit: A take-or-pay agreement with an investment-grade or well-capitalized customer is different from one with a venture-backed startup burning cash.
Termination rights: Convenience termination, hardware-refresh rights, benchmark pricing clauses, and broad service-out provisions reduce financeability.
Remedies: Late-payment rights, deposits, letters of credit, parent guarantees, and termination payments determine whether the provider can actually collect.
Numerical Example
A provider buys a $30 million GPU cluster and signs a customer to a 36-month take-or-pay contract for $1.2 million per month. Headline committed revenue is $43.2 million.
If operating costs are $350,000 per month and debt service is $550,000 per month, the contract appears to cover both. But if the customer has a right to terminate after 12 months with a three-month fee, the financeable revenue may be closer to the first year plus termination payment, not the full $43.2 million.
Service-Level and Delivery Risk
Customers usually require service-level commitments for uptime, latency, networking, security, support response, and hardware availability. If the provider misses those standards, the customer may receive credits, termination rights, or reduced payment obligations.
This matters in AI compute because bottlenecks are not limited to GPUs. Power availability, cooling, fiber, storage, orchestration software, and cluster networking can all affect deliverability.
Difference From Pure Usage Pricing
Usage pricing exposes the provider to demand volatility: revenue rises when customers run workloads and falls when they stop.
Take-or-pay shifts some volume risk to the customer. The provider still bears operating risk, customer credit risk, renewal risk, and residual-value risk after the contract ends.
Diligence Questions
What is the true minimum monthly payment after credits, ramp periods, and discounts?
Can the customer terminate for convenience, benchmark pricing, model changes, or hardware refresh?
Are service-level credits capped, or can they materially reduce monthly revenue?
Is there a parent guarantee, deposit, letter of credit, or other credit support?
Does the contract match the debt amortization schedule and hardware useful life?
What happens if newer GPUs make the contracted fleet less competitive?