A colocation contract is mostly boilerplate, but five provisions reliably determine how the deal performs once the easy part is over and the parties have a real disagreement. We've seen deals go sideways and deals work fine across roughly the same operator base — and the difference is almost always in these five clauses. This is what we look for, what to push on, and what's actually negotiable.
1. Curtailment terms
Curtailment is the operator's right (or the grid's right) to interrupt your power supply. The provision matters because it converts a stated power rate into an effective rate, and the effective rate can be meaningfully worse than the headline number if the curtailment terms are loose. We look at four sub-points:
- Frequency and duration limits. Annual cap on curtailment hours, single-event maximum duration, and rolling-window limits.
- Notification requirements. How much notice are you owed, in what form, and what's the operator's obligation if they fail to give it?
- Compensation structure. Service credits, rate rebates, or no compensation? The default for many mining-era contracts is 'no compensation' on grid-initiated curtailment; that's a meaningful AI buyer concern.
- Force majeure interaction. Is grid curtailment force majeure (no compensation) or operational interruption (compensation)? The answer varies and matters.
2. Pause and ramp-down provisions
This is what protects you when your business strategy changes or when external conditions force a temporary reduction in load. Without explicit pause provisions, the default is 'pay the contract minimum regardless.' For long-dated contracts in a volatile demand environment, that's a real risk.
What we negotiate for: a defined right to pause for a fixed maximum period (typically 30-90 days per year) with stated notification requirements, modified minimums during the pause period, and a clear restart obligation on both sides. Operators reasonably want notice and bounded exposure; customers reasonably want flexibility.
3. Default, cure, and termination
This is the framework for what happens when one side stops performing. It matters more than people initially think because the default version of these provisions is usually unbalanced in the operator's favor, and the asymmetry compounds over a 5-10 year term.
Key sub-points:
- 01What counts as a default by each side? Specific monetary thresholds, specific operational failures, specific SLA breaches. Vague language ('material breach') leans toward whoever has more lawyer hours to dedicate to litigation.
- 02Cure periods. How long does the defaulting party have to fix the problem? Asymmetric cure periods (e.g., 30 days for the customer, 5 days for the operator) are common and worth pushing on.
- 03Termination rights and reciprocity. Can the customer terminate for cause, with what notice, and what's the financial settlement? Symmetric rights are the goal; partial-term termination buyouts are often the negotiation lever.
- 04What survives termination? Service credits earned but unpaid, data return obligations, equipment return logistics. Stretch deals fail here.
4. Assignment rights
Assignment provisions govern who can transfer the contract to whom. This matters more for the customer than the operator typically realizes. If you're an AI company that might be acquired, the assignment provision determines whether your hosting contract survives the acquisition or triggers renegotiation.
We push for permitted-transferee language that covers parent/subsidiary transfers, sale-of-business transfers, and change-of-control transfers without operator consent. The operator's legitimate interests are protected by financial covenants and creditworthiness thresholds; full consent rights are usually overreach.
Symmetrically, the operator's right to assign matters too. If the operator gets acquired or restructured, you want the right to consent to material changes in the entity providing your service.
5. Dispute resolution
Where do disagreements get resolved, and how? The answer affects how aggressively each side will pursue claims and how realistic settlement is. We look at:
- Venue and choice of law. Operator-favorable venue (their headquarters jurisdiction) versus a neutral venue. For deals where the operator is a startup, neutral venue matters more.
- Arbitration versus court. Arbitration is faster and quieter; court is slower and more public. Pick deliberately, not by default.
- Mandatory mediation or step-up clauses. Forced settlement attempts before litigation reduce costs and preserve relationships.
- Discovery scope. Arbitration discovery can be limited by agreement; sometimes you want full discovery available and shouldn't trade it away.
- Carve-outs for injunctive relief. Both sides usually want the ability to seek emergency relief in court regardless of the arbitration clause.
What we don't lose sleep over
Several provisions get a lot of negotiation attention but matter less than people think. Service level definitions matter, but the SLA credit cap usually limits the practical impact; what matters more is the termination right that triggers when SLA failures accumulate. Insurance certificates and indemnification matter for catastrophic events; for normal operations, the SLA structure does more work.
Confidentiality clauses matter, but only as much as either side actually keeps confidential. Most operators routinely disclose customer names; if confidentiality matters to you, push for it specifically rather than relying on the boilerplate.
How to use this
If you're negotiating a hosting contract right now, run these five provisions through a specific scenario: what happens if the grid curtails for 200 hours next year? What happens if your AI customer cancels their commitment in year three? What happens if the operator's parent company gets acquired by a competitor? What happens if you and the operator disagree about whether an SLA breach occurred?
If the contract's answers to those questions are clear and acceptable, the rest of the agreement is probably fine. If the answers are ambiguous or one-sided, those are the clauses to redline before signing.
