AI data center infrastructure and finance, explained.

The data-center capital stack is the set of distinct assets that sit at a single data-center address: land and interconnection, the shell and fit-out, generation and power equipment, and the compute inside. They share a location and almost nothing else — their economic lives differ by an order of magnitude, their dominant risks are unrelated, and each is naturally held and funded by a different kind of capital. Reading the stack layer by layer, rather than as one building, is the starting point for every structuring decision in the sector.

Powered land is a parcel whose value rests on a documented, time-bound path to the electricity a data center requires. The label alone proves nothing: a financeable position has to state how much capacity is available, when it can be energized, whether service is firm or interruptible, what facilities and network upgrades remain, what they cost, and whether the rights survive a sale or change of control. Secured power is therefore a diligence conclusion about documents and obligations, not a synonym for land near a substation.

A powered shell is a completed or near-completed building with secured power delivered to it and connectivity pathways in place, but without the tenant-specific IT fit-out. The tenant is buying three things: a schedule it did not have to originate, a power position it did not have to secure, and design control over the part of the facility that differentiates its operation. The commercially decisive question is where the demarcation line between landlord and tenant scope sits, because that line decides which capital funds which half of the building.

Speed-to-power is how fast a data center can go from a signed site to live, reliable megawatts. With grid interconnection now running beyond five years, developers increasingly buy speed with on-site and behind-the-meter generation — turning time-to-power into the variable that wins deals.

A data center gets power one of three ways — through a grid interconnection, through on-site generation built behind the meter, or through a hybrid of the two — and in the AI buildout the choice is driven less by cost than by time. Grid connection is the cheapest firm power and the slowest to secure, with large-load interconnection now running beyond five years in the busiest markets. On-site generation, usually natural-gas turbines, buys speed at the price of fuel, permits, emissions and running a power plant. Most large projects end up combining them, bridging to first power on site while a grid position or permanent generation is built behind it. Which mix a project can actually reach, and how firm each leg of it is, is what decides when the facility energizes — and therefore when it can earn.

Energization is the point at which a facility can first draw firm power at a defined capacity, as distinct from being built, connected, or commissioned. It is the pivot date of a data-center structure: construction financing is sized to it, equipment deliveries are sequenced against it, leases and offtake commence from it, and revenue starts after it. Because so many obligations key off a single date, the substantive question in most structures is not when energization is forecast but who carries the cost if it moves.

The commercial operation date is the contractually defined milestone on which a facility — or a phase of one — is accepted as ready to perform against agreed criteria, and from which the commercial obligations run. It is a documentary event rather than a physical one: it occurs when the tests are passed and the certificate is issued, not when the lights come on. Energization is a precondition of it, sometimes months earlier. The distinction matters because different milestones release different money: energization typically ends the construction period a facility was sized against, while commercial operation is what starts rent, starts offtake, and lets equity see a return.

An interconnection queue is the ordered process a utility or system operator uses to study and connect new load or generation to the grid. A position in that queue is site-specific, capacity-specific and conditional, and it is the input every other layer of a data-center project waits on. Because the queues in the major markets run years long, the position — not the megawatts — is the scarce asset, and most of the work of holding one is avoiding the ways it can be forfeited.

A will-serve letter is a written statement from the utility that would actually serve a site, saying it currently expects to be able to provide a described quantity of capacity to a described load at a described location. It is worth having and worth reading closely. What it generally is not is a reservation of that capacity against other applicants, a commitment to a date, or an allocation of who pays for the network upgrades that serving the load would require. A power claim resting on one is a claim that a utility has looked at the site and not said no — which is a diligence milestone, not yet a position capital can be arranged against.

An interconnection agreement is the executed contract between an interconnecting customer and the party that owns the network it is connecting to. It is the last rung of the power-document ladder and the first one that creates obligations rather than expectations: it allocates the scope and cost of the interconnection facilities and network upgrades, sets milestones and the security that backs them, and states what each side must do and by when. What it does not do is deliver electricity on a date — construction still has to happen — and what a lender reads in it is mostly the size of the obligations it puts on the customer, not the rights it gives them.

Curtailment is a right, held by a utility or system operator or created by contract, to reduce the power a facility may draw under defined conditions. It is usually read as a defect. It is better read as a price: a curtailable load can frequently be connected faster and on terms a firm load cannot obtain, and the consideration for that is availability the facility can no longer promise. The question capital asks is never whether curtailment exists but where it lands — whether the facility absorbs it, the tenant accepts it, or on-site capacity covers it — because unallocated interruption is the version that reprices everything above it.

Entitlement is the set of land-use approvals — zoning, special or conditional use permits, site plan approval, air and water permits, building permits — that make the intended use of a parcel lawful. It is not administrative overhead; it is a condition precedent, because until it is resolved no other layer of the project can be built and none of the capital behind them can be drawn. The distinction that governs pricing is whether the use is permitted as of right or requires a discretionary approval, because a discretionary approval is an option held by someone who is not a party to the deal.

Behind-the-meter generation is electricity produced on the customer's side of the utility revenue meter and consumed by the facility directly, without passing through the transmission system. It is used in data-center development mainly to escape the interconnection timeline, and it works. What it also does is replace a tariff relationship — where the utility carries reliability, fuel and dispatch — with an operating business the project now owns, which is a credit change as much as an engineering one.

A power purchase agreement, or PPA, is a contract between an electricity producer and a buyer covering volume, price and term. In a data-center context it runs in two directions: the facility buys power under one, and the facility's own credit is frequently what makes new generation financeable under another. What a PPA actually transfers is set by its shape rather than its existence — a fixed-price, firm, baseload contract is a different instrument from an as-generated one at an indexed price, and only one of them supports a facility that runs continuously.

A data center is financed not as one building on one loan but as a sequence of stages and a stack of separable assets. The sequence runs from development and land equity, through a construction loan that is often a mini-perm, across a stabilization period, and out into permanent capital — a term loan, a sale-leaseback, private credit, or a securitization. The stack is financed layer by layer, because the land, the shell, the power and the compute have different lives and attract different capital. The through-line under all of it is that the credit being underwritten is the contracted tenant cash flow — the lease or the offtake — far more than the concrete, which is why data-center capital structuring looks closer to corporate credit than to commercial real-estate lending.

AI compute — the GPUs and accelerators, not the building around them — is financed separately, on its own short cycle, through four main routes: equipment leases, vendor and OEM paper, asset-backed bilateral debt, and, increasingly, the capital markets. What each route underwrites is not the hardware invoice but the compute contract behind it and the credit of the party paying, because accelerators lose value on the product cycle and are a weak second line of defence. That is why compute is kept out of the long-dated structures that fund the real estate: it turns over several times inside the life of the building, so it is repaid on its own schedule, against its own contract, by capital that priced for a short-lived asset.

Project finance is capital advanced to a standalone project entity and repaid from that project's own cash flow, secured on its own assets and contracts, with limited or no recourse to the sponsor behind it. It is not a cheaper way to borrow — it is a different question being asked. Corporate borrowing tests a company; project finance tests a set of contracts, and it is only available where those contracts allocate every material risk to a party able to bear it.

A special-purpose vehicle is an entity created to hold one asset or project and carry on no other business, so that the asset's cash flow can be financed without the rest of its owner's affairs reaching it — and without the asset's own risks reaching the owner. The separation is only as good as the covenants and documentation that create it: an entity labelled an SPV while sharing accounts, staff and obligations with its parent is a name, not a ring fence. In a data center it is the mechanism that lets four layers with incompatible economic lives be financed separately.

A construction loan funds the building of an asset rather than its operation: it is drawn in stages against verified progress, carries interest that is usually capitalised rather than paid from cash flow, and matures shortly after completion. It is never repaid by the project's own operations, because the project has none while it is being built. It is repaid by a takeout — a term financing, a sale, or a leaseback — and a build with no identified takeout is an equity position wearing a loan's name.

A mini-perm is short- to medium-tenor debt — typically a few years — that funds a project through construction and its first operating years, then has to be refinanced rather than repaid from its own amortisation. Because it amortises little over its life, most of the principal survives to maturity as a balloon, and the whole instrument is built around getting that balloon refinanced on time. A hard mini-perm makes refinancing mandatory; a soft one makes it merely expensive not to. The same idea — pressure to refinance a short bridge before a long-life asset's debt comes due — reappears in the bond market as the anticipated repayment date.

In a sale-leaseback, the owner of an asset sells it and simultaneously leases it back, so the capital tied up in the asset is released while its use continues without interruption. What has been sold is not the use of the asset — that is retained — but ownership, the end-of-term position, and a measure of control. It is a way of converting an owned asset into a payment obligation, and its price is set almost entirely by what the buyer thinks the asset will be worth when the lease ends.

A letter of credit is an undertaking by a bank to pay a named beneficiary on presentation of specified documents, independent of the underlying contract and of any dispute under it. Its function is substitution: the beneficiary is relying on the bank's credit rather than on the applicant's. In a data-center build it is what makes a utility, a turbine supplier or a landlord willing to deal with a project entity that has no credit history of its own.

Offtake is a contracted commitment by a buyer to take — or to pay for — the output a project produces, for a defined period and on defined terms. It is what converts an input into a project: without it there is an asset and a forecast, with it there is a payment obligation owed by a named counterparty. Every layer of a data center has its own form of offtake, and the layer that has none is being financed on somebody else's.

A take-or-pay contract obliges a buyer to pay for a contracted quantity whether or not it actually takes it, provided the seller stands ready to deliver. Because payment does not depend on the buyer's own demand materialising, the obligation behaves like a debt of the buyer rather than a projection of its behaviour — which is precisely why capital can be sized against it. A consumption-based contract with no minimum is a commercial relationship, not a cash flow, however large the expected volumes.

The debt service coverage ratio compares the cash a project generates in a period against the debt service due in that period. Above 1.0 the project covers its obligations; the margin above 1.0 is the headroom before it does not. Its practical importance is that it sizes the facility rather than merely grading it — capital is advanced in the amount whose service the cash flow covers at the required ratio, which means the sizing runs backwards from cash flow rather than forwards from cost.

Tenor is the length of a financing: the period from drawdown to final maturity, or the term of a lease. It is a separate question from how much capital is advanced and how fast it is repaid, and it is the one most often decided by default. Three durations have to line up — the asset's economic life, the term of the contract behind it, and the maturity of the money — and in a data center those three differ by an order of magnitude between layers.

Residual value is what an asset is expected to be worth when a term ends. It is an assumption, not a fact, and it is embedded in every lease and in every facility that does not repay in full over its life. The question that decides a structure is not what the residual is but who is holding the guess: whoever holds it absorbs the difference between the assumption and the outcome, and across a data center's four layers that difference ranges from negligible to the largest exposure in the deal.

A stranded asset is one that still functions but can no longer generate enough return to support the capital committed against it. It is distinct from an obsolete asset, which no longer works, and from an impaired one, which is an accounting position. Strandedness is relational rather than intrinsic — the same equipment is stranded in one structure and productive in another — which is why most of the work of avoiding it is done when the deal is arranged rather than when the asset is bought.

A neocloud is a specialist compute provider that acquires accelerators at scale and sells access to them, rather than running them for its own workloads. As a counterparty it is neither a hyperscaler nor a colocation provider: it has no diversified business to cross-subsidise a downturn, and it sells compute rather than space and power. The distinctive credit feature is correlation — its revenue, its collateral value and its refinancing capacity all depend on the same variable, so they weaken together rather than in sequence.