The Contract Stack & Commercial/Legal Framework

Everything in Part 2 so far has been about getting the project built — the schedule (Chapter 2.1), the owner's organization and delivery model (Chapter 2.2), and the long-lead supply chain (Chapter 2.3). This chapter is about the instrument that binds all of it together and decides who pays when something goes wrong: the contract stack. An AI data center is not procured under one master agreement. It is procured under a dozen of them — an interconnection agreement with the utility or ISO, a power-purchase agreement, an engineer-procure-construct contract for the shell and power chain, equipment supply agreements and slot reservations for the long-lead gear, a colocation or master-services agreement if you are renting space, operations-and-maintenance contracts, and GPU-cloud service-level agreements if you sell compute. Each one allocates a slice of risk. The project's commercial integrity is the sum of those allocations — and, critically, the seams between them.

This chapter applies that frame to legal risk. We map the stack end to end; we trace how schedule, performance, and force-majeure risk are allocated across it; we work the mechanics of liquidated damages, performance guarantees, warranties, and bonding that make those allocations enforceable; we examine the equipment-supply and slot-reservation contracts and the owner-furnished-equipment interface where warranties most often gap; we treat the interconnection and grid-side agreements as the commercial instruments they actually are; and we close on the tenant-facing terms that turn a built facility into a revenue contract. The through-line: risk does not disappear when you sign it away — it moves to whoever the contract names, and the expensive mistakes are the ones where it lands on a party that cannot control the underlying event.

Picture the stack as a vertical column from the grid down to the customer. At the top sits the interconnection agreement (an LGIA, ISA, or the utility's large-load service agreement) and the power-purchase agreement — the two instruments that secure the megawatts and price the energy, and the two that are hardest to unwind once signed. Below them, the EPC contract (or the set of trade packages that replaces it) builds the physical asset, fed by equipment supply agreements and slot-reservation contracts for the transformers, switchgear, chillers, and gensets that have 2-to-5-year lead times. If the developer is not the end user, a colocation agreement or master-services agreement (MSA) sits between the facility owner and the tenant; once live, O&M contracts govern operations and a GPU-cloud SLA governs what the compute customer is owed. Each layer is a separate counterparty, a separate negotiation, and a separate risk-allocation regime.

The reason this matters more for an AI campus than for a legacy data center is the concurrency and the seam count. A traditional facility could be built sequentially — secure power, then build, then fit out, then lease. An AI build runs all of these in parallel under a speed-to-power clock: the interconnection study is still open while the EPC is mobilizing while the GPUs are being allocated while the anchor tenant is in diligence. The contracts are therefore signed before their dependencies are resolved, which means the interfaces between them — the handoff dates, the owner-furnished-equipment delivery commitments, the conditions precedent — carry far more risk than the body of any single agreement. The recurring failure mode is not a bad clause; it is a gap between two contracts that no clause covers.

The governing principle of risk allocation is older than data centers and rarely followed: assign each risk to the party best able to control or absorb it. Schedule risk on the concrete belongs to the EPC; schedule risk on the GPU allocation belongs to whoever holds the OEM relationship; schedule risk on energization belongs to nobody the contract can name, because it sits with the utility and the queue. The art is matching the allocation to the control, and the recurring error is allocating a risk to a party who has no lever over the event — at which point they either price it punitively or default on it, and either way the owner pays.

Schedule risk is the dominant one because the speed-to-power clock dominates the economics. It splits three ways: construction schedule (the EPC's, capped by delay LDs), equipment-delivery schedule (the OEM's, often weakly capped or excused), and energization schedule (the utility's, almost never backed by enforceable damages). The bind is that the three are serially dependent: a slipped transformer delays energization, which strands a finished building full of depreciating GPUs. Yet the contracts that govern them are independent, and the party suffering the loss (the owner) holds recourse against none of the others for the full amount.

Performance risk — does the facility deliver the contracted power, PUE, and availability at completion — is the EPC's to carry, backed by performance guarantees and performance LDs. This is the cleaner allocation, because the contractor controls the design and build. The complication is the owner-furnished-equipment line: when the owner furnishes the GPUs, the CDUs, or the switchgear, the EPC's performance guarantee carves them out, and a facility that underperforms because the owner's chillers cannot reject the heat is the owner's problem, not the contractor's. The performance guarantee is only as good as the equipment it actually covers.

Force-majeure risk is where 2026 has rewritten the playbook. A force-majeure clause excuses performance for events outside a party's control — and the question of what counts as 'outside control' has become the single most negotiated definition in the stack. A transformer factory fire, an HBM allocation shortfall, a tariff or export-control change, a grid-upgrade delay: a decade ago these read as genuine acts of God. In 2026 they are foreseeable, recurring features of the supply chain, and a contract that lets the OEM or contractor excuse them as force majeure hands the owner an uninsured, unrecoverable loss. The sophisticated owner narrows the definition aggressively and carves supply-chain and change-in-law events out of the excusable set.

Risk allocation is a promise; liquidated damages, guarantees, warranties, and bonds are what make the promise collectable. Liquidated damages (LDs) are a pre-agreed sum payable per day of delay (delay LDs) or per unit of performance shortfall (performance LDs), set at contract signing so the owner need not prove actual loss in court. They come in two flavors and both are capped. Delay LDs in data-center EPC contracts are commonly structured per-day with an aggregate cap; performance LDs (for missing a guaranteed PUE, power capacity, or capacity-test milestone) are typically capped at 10-15% of contract price, sometimes as wide as 5-25%, with an overall aggregate liability cap often in the 10-20% band (Womble Bond Dickinson; EPC-market practice, 2025). The number that matters is the gap between that cap and your real loss.

A one-month energization slip on a 100 MW AI campus strands a building whose GPUs are depreciating at a multiple of any LD a contractor will accept — the LD cap recovers a fraction of the loss and the residual sits on the owner, by design. LDs are a risk-sharing instrument, not a make-whole one; the contractor will not price unlimited delay exposure, and if you force it, the bid price absorbs the risk premium anyway. The correct posture is to treat the LD cap as the contractor's contribution, then own the residual explicitly — through schedule buffer, delay-in-startup insurance (Chapter 2.6), and a financing structure that survives a slip (Chapter 2.5).

Performance guarantees are the technical promises the LDs enforce: a guaranteed PUE at a defined load and ambient, a guaranteed power capacity delivered to the rack, a guaranteed availability demonstrated at integrated systems testing. They are proven at commissioning — which is why the commissioning program (Chapter 13.1) and the acceptance test plan (Chapter 13.2) are commercial documents as much as technical ones: the test scripts define the conditions under which the guarantee is met, and a loosely-written acceptance test is a guarantee the owner cannot enforce. Warranties then cover defects after acceptance — typically 12-24 months on workmanship, longer on specified equipment — and the warranty's value depends on it not being voided by the O&M contractor's maintenance or by an OFE interface the warranty excludes.

Bonding and security backstop the whole regime. A performance bond (commonly 10-100% of contract value depending on jurisdiction and counterparty), a parent-company guarantee, retention withholding, and letters of credit ensure that an LD or warranty claim is actually collectable against a contractor who may be thinly capitalized at the project level. The security package is the difference between a contractual right and a recoverable dollar — and on a build where the EPC is an SPV or a JV, it is the only thing standing between the owner and an uncollectable judgment.

The long-lead equipment that dominates the schedule (Chapter 2.3) is procured under its own commercial regime, and three contract types govern it. An equipment supply agreement is the conventional purchase: spec, price, delivery date, warranty, and LDs for late delivery or non-conformance. A slot-reservation agreement is the newer and more interesting instrument — a distinct contract that buys a place in the manufacturer's production queue for transformers, switchgear, or accelerators, typically with a substantial non-refundable deposit and price-and-spec terms to be fixed later. In a supply-constrained market, the slot itself is the scarce asset, and the reservation contract is how you secure it before you have finalized the design. The deposit is the price of optionality on a delivery date you cannot otherwise guarantee.

The commercial subtlety in a slot reservation is that it inverts the usual leverage. A normal supply agreement lets the buyer hold delivery risk over the seller via LDs; a slot reservation in an allocation-driven market lets the seller hold the buyer, because the deposit is sunk and the alternative is the back of a multi-year queue. That is the GPU and HBM allocation game in contract form, and it bleeds into financing — vendor-financed and prepay structures exist precisely because the slot is worth more than the cash. The owner's defense is to negotiate delivery-date firmness and a remedy for slippage into the reservation itself, rather than accepting a deposit that buys a place in line but no enforceable date.

Then comes the seam that quietly destroys performance guarantees: the owner-furnished-equipment (OFE) interface. When the owner procures equipment directly — GPUs always, often the switchgear, transformers, and CDUs — and furnishes it to the EPC for installation, the contract must define a clean line where the OEM's warranty ends and the EPC's installation warranty begins. If that line is fuzzy, a failure at the interface falls into a no-man's-land: the OEM says it was mis-installed, the EPC says the equipment was defective, and the owner — who sits on both sides of the handoff — owns a loss neither counterparty will accept. The OFE interface warranty is the single most under-negotiated clause in the AI-build stack, and it is where the performance guarantee you thought you bought silently disappears.

The interconnection and power agreements are usually filed under 'regulatory' and treated as a permitting hurdle. That is a costly mis-framing. The LGIA, ISA, or large-load service agreement is the single most commercially binding instrument in the entire stack — it carries the project's hardest dates, its largest unhedged cost-allocation exposure, and the one obligation that cannot be renegotiated after signing. Read it as a contract, not a permit, and three commercial dimensions dominate: cost allocation, milestones, and security.

Cost allocation answers who pays for the network upgrades the interconnection study identifies. In the large-load tariffs spreading across US states in 2025-2026, the answer is increasingly blunt: the data center pays 100% of the assigned network-upgrade cost, often plus customer-paid distribution upgrades and a surcharge for the largest projects (Oregon, Ohio, and 23 states with approved large-load tariffs as of 2026). This is a direct capex line the project model must carry, and it is identified during the study — meaning the number is uncertain at the moment you must commit. Milestones impose a schedule the owner must hit (and post security against) or forfeit its place; these are the dates the EPC and tenant contracts must be built around, because a missed interconnection milestone cascades through the entire stack. Security is the cash at risk: deposits cited as high as $2.7M per MW, paired with 15-year contract terms and minimum-demand charges, are now a routine condition of large-load service.

The newest commercial wrinkle is the take-or-pay floor. To protect other ratepayers from a data center that contracts for capacity it never uses, state tariffs now impose minimum-payment obligations — Ohio's requires paying for at least 85% of contracted capacity over a 12-year minimum regardless of consumption. That is a fixed cost the project carries through the ramp, when utilization is lowest, and it directly underwrites the lender's view of contracted versus merchant exposure (Chapter 2.5). On the supply side, the December 2025 FERC order directing PJM to create four transmission-service options for co-located load (firm and non-firm contract-demand, interim non-firm, and the traditional NITS) turned a regulatory question into a menu of commercial structures — the developer now chooses its grid-side firmness the way it chooses PPA firmness, trading curtailment risk against cost. → energy-supply strategy and the PPA-vs-BYOP fork in Chapter 3.4; site-side interconnection strategy in Chapter 3.1.

The bottom of the stack faces the customer, and it splits by who the customer is. If you are a colocation operator, the tenant contract is a colo agreement or MSA that sells space, a power-capacity commitment (in kW or MW, not square feet — power is the product), and a facility SLA on uptime and PUE. The critical commercial term is the power-capacity ramp: the tenant's density ramp (Chapter 1.6) must be matched by a contracted capacity-delivery schedule, or the tenant outgrows the contract and the operator either strands capacity or breaches. If you are a GPU-cloud operator, the customer contract is a service-level agreement on uptime, goodput, and support, backed by service credits — and the 2026 market has moved this from a marketing line to a negotiated commercial term, with leading neoclouds offering 99% rack-level uptime backed by credits (SemiAnalysis ClusterMAX 2.0).

The seam to watch at the bottom of the stack is the SLA pass-through: when a facility outage breaches the cloud SLA, who owes the service credit? If the cloud operator leases from a colo, the facility SLA and the cloud SLA must be back-to-back, or the operator owes its compute customers a credit it cannot recover from its landlord. And the SLA metric itself is shifting — a raw uptime percentage understates what an AI customer actually buys, because a cluster that is 'up' but delivering poor goodput (jobs failing, stragglers, fabric congestion) is breaching the spirit of the contract while meeting its letter. The frontier of tenant terms is moving from availability to goodput-based SLAs, which is the same reframing the reliability chapters apply to facility design (Chapter 12.2). The full productization of consumption and SLA tiers — the menu of firm, burstable, and spot compute the customer actually buys — is built out in Chapter 10.9.