Everyone runs the engineering checklist. Tier III concurrent maintainability, liquid cooling capacity, fiber diversity, rack density. That list tells you whether the building can run your GPUs. It does not tell you whether you will still be in the building in year four. Many of the sites being pitched to neoclouds today are first builds by developers who are new to data centers, financed by lenders who have never underwritten a GPU tenant, on power that was secured eighteen months ago under rules that have since changed. The contracts those sites were built around were written for $5 million enterprise colocation deals, not $100 million plus GPU deployments where a 100-hour outage costs more than the rent. This is the diligence a data center lawyer would want a neocloud to run before racking.

The 14-item legal due diligence checklist for a GPU colocation deployment

Each item is a question with a document behind it. An operator who answers with an assurance instead of a document has told you something.

  1. Title and liens. Title commitment on the land and building; searches for mortgages, construction and mechanic's liens, and tax liens. New builds carry unpaid-contractor risk.
  2. The operator's own rights. Deed, recorded ground lease, or recorded option. If the operator leases, its term must outrun yours and the landlord must consent to your occupancy.
  3. SNDA from the lender. A signed subordination, non-disturbance and attornment agreement from every mortgage lender on the property.
  4. Lien waivers on your GPUs. Written waivers from the landlord and its lender of any lien or security interest in your equipment, and a clear characterization of your GPUs as removable personal property.
  5. UCC-1 or PPSA searches. Against the operator and the property owner, for blanket liens and collateral assignments of contracts, including the power agreements.
  6. Interconnection agreement. Executed, with utility queue position, firm energization date, and the MW that are firm rather than requested.
  7. PPA continuity. Term at least as long as yours plus renewals, with the assignment, collateral assignment, and lender step-in clauses read in full.
  8. Reserved capacity. A contractual MW allocation to your suite, with an oversubscription prohibition and a priority rule among tenants.
  9. Tariff exposure. Which utility tariff applies, its minimum demand charge, its contract term, and who bears a mid-term tariff change.
  10. Curtailment regime and plan. The curtailment rules the site is subject to, any demand-response program the operator has enrolled in, and the site's written contingency plan with hour budgets.
  11. Backup generation. Air permit run-hour limits, fuel supply contracts, and whether backup is sized for your load or the operator's share.
  12. Permits and moratoria. Zoning or conditional use permit, building permits, certificate of occupancy, and whether the jurisdiction has a pause in force.
  13. Operator and lender. Comparable build history, who runs O&M, construction loan maturity against your term, and whether the lender has consent rights over your agreement.
  14. Access and removal rights. Guaranteed physical access and a fixed removal window after termination, including during an operator default.

The rest of this guide takes each group of checks and explains what goes wrong when it is skipped.

A lien you did not find is a creditor who can take the building with your GPUs inside it

The first colocation red flag is a facility whose public record does not match the pitch. Run a title commitment on the land and the building shell and search the county records for recorded mortgages, construction and mechanic's liens, and tax liens. New-build sites are the highest risk because the general contractor, the electrical subcontractor, and the switchgear vendor may each hold lien rights for unpaid work, and in many US states a mechanic's lien takes priority from the start of construction rather than from the date it is recorded. A lien ahead of your agreement is a party who can foreclose on the ground your GPUs sit on.

Operator insolvency is the scenario that turns a lien search into a survival question. When Cyxtera filed for Chapter 11 in 2023, it rejected the lease on its Amsterdam facility and told customers to relocate by August 31, with operations ceasing September 6, a window of roughly twelve weeks for a site that was only 6% occupied. A full room of GPUs at a 50 MW site does not move in twelve weeks. The estate kept the sites that made money and dropped the ones that did not; a customer's agreement had no vote.

Your right to stay only runs as deep as the operator's own lease and the lender's non-disturbance promise

Most colocation operators do not own the building. Confirm whether the operator holds a deed, a recorded ground lease, or a recorded option, and if it is a ground lease, read it. The operator's lease term has to outrun your 5 to 7 year deployment, the landlord has to consent to your occupancy if the lease requires it, and you should know what an operator default on the ground lease does to you, because as a rule a sublease falls with the head lease unless a recognition agreement from the landlord says otherwise.

Then get the SNDA. A subordination, non-disturbance and attornment agreement is the lender's written promise that if it forecloses on the landlord or operator, it will not disturb your possession as long as you perform. Torys describes SNDAs as a standard feature of data centre finance, and reports that lenders now ask for notice of all customer-agreement defaults, an extended cure period for landlord defaults, a tolling of your termination right while the lender forecloses, and a qualified-operator requirement so that an experienced operator always runs the site. Those lender asks are negotiable; the non-disturbance covenant is not optional. Without it, a foreclosure can end your agreement regardless of what the colocation contract says.

In Texas and Ontario the landlord can hold a claim against GPUs sitting in the room, so get waivers before racking

A lien waiver on your equipment is the check that most colocation customers skip because the GPUs are obviously theirs. Ownership is not the issue; the landlord's and lender's statutory and contractual claims are. Under Texas Property Code section 54.021, a person who leases a nonresidential building has a preference lien on the property of the tenant or subtenant in the building for rent due, and for rent coming due over the following twelve months. In Ontario, the Commercial Tenancies Act lets a landlord distrain the chattels of subtenants and occupants on the premises for a head tenant's arrears, unless the subtenant files the statutory declaration under section 32(2) and pays what it owes the head tenant to the landlord. In both jurisdictions, a secured lender's blanket lien over the operator's "equipment and fixtures" can be drafted broadly enough to invite an argument that your racks are covered.

Three documents close the gap: a landlord lien waiver, a lender lien waiver or collateral access agreement, and contract language that characterizes your GPUs, servers, and networking gear as your removable personal property that never becomes a fixture. The flip side matters as much. If you financed the GPUs, your own lender will require a landlord waiver before funding; lenders to neoclouds are already demanding stronger protections, with spreads of 275 to 325 basis points above investment-grade benchmarks. A first-time operator who cannot get its landlord and lender to sign a waiver has stalled your financing before you have racked a server.

Power is an asset someone else may already have pledged, so verify the interconnection, the PPA term, and who stands behind both

Power continuity is the second red flag in the checklist, and the one most likely to be quietly encumbered. Start with the interconnection: an executed interconnection agreement with the utility, a queue position, a firm energization date, and the number of MW that are firm as opposed to requested. In ERCOT, Senate Bill 6, signed June 21, 2025, applies to loads of 75 MW and above, requires proof of site control and a flat study fee of at least $100,000, and requires disclosure of duplicate interconnection requests elsewhere in Texas. ERCOT forecasts 138 GW of large loads by 2030, up from 87 GW in 2025, and the same reporting cites experts who expect 80% to 90% of proposed US data centers in the queue to never be built. A queue position is not a guarantee of energization.

Then the PPA. Confirm its term covers your full deployment plus any renewal options, and read three clauses: assignment, collateral assignment, and lender step-in. A PPA is an assignable asset that operators pledge to raise construction financing, and if the operator defaults on that financing the lender can enforce its collateral assignment, step into the contract, or terminate it. Search UCC-1 filings in the United States, or PPSA filings in Canada, for security interests against the operator's contract rights, not only its real property. If a senior creditor already holds the PPA as collateral, your power allocation sits behind that creditor on the day the operator fails. The data center counterparty diligence checklist covers the entity-level searches in more depth.

Reserved MW is a covenant, not a brochure, so check oversubscription and minimum-demand tariffs before you price the deal

The power red flag that does not show up in any public record is oversubscription. Ask for a written MW allocation to your suite, a covenant that the operator will not sell the same capacity twice, and a priority rule that says whose load is cut first when the site is short. An operator who will not commit a number has not reserved one.

Tariffs are the related exposure and they are changing fast. AEP Ohio's data center tariff, effective July 23, 2025, applies to loads of 25 MW and above, sets a minimum monthly demand charge of 85% of contract capacity, runs for a ramp period of up to four years plus eight years, and requires collateral equal to 50% of total minimum charges for the full term unless the customer carries an S&P rating of at least A- and a Moody's rating of A3. In Québec, Hydro-Québec has proposed a new rate for data centres of 5 MW and above averaging about 13 cents per kWh, roughly double the current large-power rate, effective November 1, 2026, alongside an underuse charge that applies when peak demand over 12 billing periods is below 60% of reserved capacity. Your colocation agreement should state which tariff the site is on, who pays the minimum demand charge when your ramp is slower than planned, and who absorbs a tariff change during the term.

Budget for 85 to 177 curtailment hours a year and write down who decides which racks go dark

Grid curtailment is the third red flag in the checklist, and the one where the math is most often skipped. The Duke University Rethinking Load Growth study equates a 0.25% annual curtailment rate to 85 hours a year, a 0.5% rate to 177 hours, and a 1% rate to 366 hours, and finds that ERCOT could integrate about 10 GW of new load at the 0.5% rate. Utilities and grid operators are designing large-load programs around those numbers. Under Texas SB 6, noncritical loads of 75 MW and above must be curtailable during firm load shed, shutoff equipment is a condition of interconnection, and ERCOT may direct loads with backup generation to deploy it or curtail once market-based services are exhausted. The voluntary program carries a minimum 24-hour notice; the mandatory program does not state one.

Do the revenue math for your own deployment. At an illustrative 1.3 kW of all-in facility power per high-end GPU, including cooling and overhead, a 50 MW IT load supports on the order of 38,000 GPUs. At a neocloud on-demand rate of about $2 to $3 per H100 GPU-hour, 100 curtailment hours is roughly $7.7 million to $11.5 million of lost billing, and 177 hours is $13.6 million to $20.4 million. SLA credits in a typical colocation agreement are a percentage of monthly fees and do not come close. Winter Storm Uri in February 2021 produced a continuous 70.5-hour ERCOT load shed order with up to 20 GW cut, so a single bad week can consume most of an annual hour budget.

Diligence here is four questions. What curtailment regime is the site subject to, mandatory and voluntary? Has the operator enrolled the site in a demand-response or reliability program that obligates it to shed your load for a payment it keeps? What is the written contingency plan, with hour budgets by scenario? And in the contract, who decides which tenant's load is cut first, and what does the operator owe you per curtailed hour beyond the standard SLA?

Diesel backup is capped at 100 nonemergency hours a year, so a contingency plan built on generators has a ceiling

Backup generation is the usual answer to the curtailment question, and it has a regulatory limit. Under the federal Clean Air Act rules for stationary engines, emergency generators may run without an hour limit during a true emergency but are capped at 100 hours a year of nonemergency operation, with at most 50 of those hours under a demand-response arrangement, and only when dispatched by the local balancing authority or transmission operator under five listed conditions. EPA's May 1, 2025 guidance loosened the reading of "imminent" emergency but did not change the 100-hour cap. A demand-response contract with a third-party aggregator counts as nonemergency use.

For diligence that means three documents: the air permit for the generator fleet with its run-hour limits, the fuel supply contract with delivery commitments during a regional event, and the operator's allocation of backup capacity among tenants. A site with N+1 generators sized for the operator's original 20 MW enterprise build does not back up your 50 MW GPU hall. Ask for the single-line diagram and the generator nameplate total, and compare it to your contracted load.

Alberta, Ontario, and Québec each gate grid access differently, and the gate decides whether your power exists

Canadian sites carry the same checklist with a different power regime. In Alberta, the AESO's interim large-load limit of 1,200 MW has been fully allocated, and the Data Centre Regulation (AR 117/2026) released June 9, 2026 moves new projects of 75 MW and above to a bring-your-own-generation model in which a project is "tethered" to committed new generation and cannot draw above that generation until it is in service. Projects without full tethering can receive "bridged" temporary access for up to three years, and the regime states that large and bridged data centres may be prioritized for curtailment during supply constraints. A colocation site in Alberta that is on bridged access has a three-year clock and curtailment priority against it; your 5 to 7 year term needs to know that.

In Ontario, Bill 40, passed December 2025, removed the non-discrimination principle for grid access, and under the draft framework the Minister of Energy and Mines would approve data centre connection requests, with eligibility to be set by regulation after the final Data Centre Playbook expected in fall 2026. A connection the operator says it has "applied for" in Ontario is a ministerial discretion, not a queue position. In Québec, data centres requesting 5 MW or more need authorization from the Ministère de l'Économie under Bill 69, adopted June 2025, through a comparative review that weighs economic benefit per megawatt and can refuse or defer. In all three provinces, ask for the actual authorization or allocation letter, not the application.

Thirty-two jurisdictions had active data center pauses in October 2026, so a site mid-approval is a site at risk

Permits are a red flag because the pause can arrive after the deposit. As of early October 2026, the dcmap.us policy tracker lists 32 active data center pauses, five statewide and 27 local. New York's Executive Order 62 of July 14, 2026 holds incomplete permits for facilities of 50 MW and above pending an environmental impact statement. Pennsylvania issued a statewide executive pause on August 18, 2026. Local moratoria range from 45 days in Gilroy, California to two years in Prince George's County, Maryland, and Indianapolis has paused approvals through the end of 2027. Ohio stopped considering new sales-tax exemptions for data centers on May 27, 2026 while a legislative study runs.

Confirm the zoning or conditional use permit, the building permits, and the certificate of occupancy are issued, not pending. Check whether any tax abatement the operator is relying on carries a clawback that passes through to occupants, and whether an equipment sales-tax exemption that affects who should buy the GPUs is still open in that state. A site where the operator is "finalizing" entitlements is a site whose power, building, and economics can all change by ordinance.

A first-time operator's lender, loan maturity, and other tenants all sit between you and your uptime

The operator itself is the last group of checks, and the most important for a new entrant. Ask for named, verifiable projects it has energized at comparable scale, and if the answer is none, ask who is actually running operations and under what O&M contract. Then look at the capital stack: the construction loan's maturity date against your term, whether there is a completion guaranty, who the senior lender is, and whether the lender holds consent rights over your agreement, including a deemed-assignment-on-change-of-control clause. Mayer Brown notes that colocation agreements can deem a change of control to be an assignment requiring consent, which means an acquisition on either side of the deal can trigger a consent you do not control.

Other tenants are the exposure nobody prices. Ask how many tenants share the site, what the aggregate contracted MW is against the interconnection, and whether a default by one tenant cross-defaults the operator's loan. JLL counts more than 190 operators in the neocloud segment, some signing 15-year leases against customer contracts that run for hours or a few years. If the tenant next door is one of them and misses rent, the operator's lender is looking at your room too.

Write removal and access rights now, because the day you need them the operator may be in default

The last red flag is an agreement with no exit mechanics. A standard colocation contract gives you badge access while the agreement is in good standing. It rarely says what happens when the operator is in default, the lender has stepped in, or the lease above the operator has been rejected. Draft the following: a guaranteed physical access right that survives termination and operator default, a fixed removal window of at least 60 to 90 days after termination with power and cooling maintained at a stated minimum, a prohibition on the operator or any successor holding your equipment for amounts in dispute, and a step-in or direct agreement with the lender so the access right binds whoever ends up owning the site.

The contract terms that cure what diligence finds

Diligence tells you what the site is; the contract fixes what it is not. Seven terms carry most of the protection for a GPU deployment. An SNDA and a landlord-lender lien waiver, executed before the first server ships. A reserved-capacity covenant with an oversubscription prohibition and a tenant priority rule. A curtailment protocol that names the regime, allocates the hour budget, and sets a per-hour payment above the standard SLA. A tariff pass-through clause that assigns minimum-demand charges and mid-term changes. Representations and a lien-free warranty on the land, building, interconnection, and PPA, with a schedule of disclosed encumbrances. A termination right if energization or a stated MW is missed by a defined number of days, with a return of prepayments or escrow. And the access and removal rights above, drafted to survive everything.

An operator who intends to perform for seven years will sign these. One who resists every enforceable term is telling you where the diligence should have looked harder.

Frequently asked questions

What are the biggest colocation red flags for AI infrastructure?

The biggest colocation red flags for AI infrastructure are legal rather than technical: a recorded lien or mortgage on the facility that no one disclosed, an operator whose own lease is shorter than your term, no non-disturbance agreement from the operator's lender, a power purchase agreement or interconnection right that has been pledged to a senior creditor, no cap or protocol for grid curtailment, a first-time operator with no comparable build history, and no written right to access and remove your GPUs after termination or default.

What happens to my GPUs if my colocation provider goes bankrupt?

If a colocation provider goes bankrupt, the provider or its creditors can reject the lease on the building your GPUs sit in and give you a short window to move. In Cyxtera's 2023 Chapter 11 case, customers at its Amsterdam site were told to relocate by August 31, with operations ending September 6. Your GPUs remain your property, but without a lien waiver from the landlord and lender, a non-disturbance agreement, and written access and removal rights, you may be negotiating for physical access from a creditor rather than exercising a contract right.

Can a creditor take over the power allocation at a colocation site?

Yes. A power purchase agreement or interconnection right is an assignable asset that operators commonly pledge as collateral for construction financing. If the operator defaults, the lender can enforce its collateral assignment, step into the agreement, or terminate it, and the capacity you contracted for goes with it. Diligence should include a UCC-1 search in the United States, or a PPSA search in Canada, against the operator's contract rights, plus a read of the assignment and lender step-in clauses inside the PPA itself.

How many hours of grid curtailment should a GPU deployment plan for?

A GPU deployment in a flexible-load regime should plan for between 85 and 177 curtailment hours a year, depending on the program. Duke University's Rethinking Load Growth study equates a 0.25% annual curtailment rate to 85 hours and a 0.5% rate to 177 hours. In ERCOT, Texas Senate Bill 6 requires loads of 75 MW and above to be curtailable during firm load shed, and Winter Storm Uri in February 2021 produced a continuous 70.5-hour load shed order. The contract should state who decides which load is cut first and what the operator owes you for each hour.

Does a colocation customer need an SNDA?

Yes, for any multi-year, high-value GPU deployment. A subordination, non-disturbance and attornment agreement (SNDA) is the lender's written promise not to disturb your possession if it forecloses on the operator or the landlord. Torys describes SNDAs as a standard feature of data centre finance, and lenders now ask for notice of defaults, an extended cure period, and a qualified-operator requirement in return. Without an SNDA, a foreclosure can end your agreement regardless of what the colocation contract says.

Run the checklist before the deposit, not after the first curtailment

A 5 to 7 year GPU deployment is a bet that the site, the power, and the operator all survive longer than the term. The fourteen checks above are how you find out before the money moves whether that bet is on a building or on a slide deck.

LegalLayer drafts, negotiates, and reviews high-value infrastructure and enterprise agreements, with a licensed lawyer accountable for the work. It runs the title, lien, PPA, and permit diligence above, negotiates the SNDA, lien waivers, reserved-capacity covenant, and curtailment protocol with the operator and its lender, and gives your deal team legal agents tailored to how it works. If you are evaluating a colocation site for a GPU deployment, start with LegalLayer at legallayer.ai.

This article is general information, not legal advice. Statutes, tariffs, and grid rules cited here change frequently and vary by jurisdiction. For advice on a specific transaction, consult a qualified attorney.

Sources