Financing on-site generation

The default treatment of on-site power is to build it into the facility budget: the developer buys the units, installs them, and the cost disappears into the project's overall capital requirement. That works, and it is the wrong answer more often than it looks, for the same reason folding compute into a building facility is the wrong answer.

Generation equipment has a life measured in decades. The building around it lasts a generation; the compute inside turns over several times within either. Financed as part of the facility, the generation layer is funded on the facility's terms and its own long life is wasted. Financed separately, it attracts energy-infrastructure capital that is comfortable with power assets, prices them on their own merits, and holds them for the horizon they actually last. That is the whole argument of [the data-center capital stack](/learn/the-data-center-capital-stack), applied to the layer that most often gets absorbed into the one below it.

Separating it means answering, deliberately, the questions common ownership would otherwise answer silently:

• Which entity owns the plant, and does it own the site rights it stands on.
• What it sells, to whom, and under what contract. A generator inside the same entity as the building sells nothing; a generator in its own entity sells power under an agreement that can be read, assigned and underwritten.
• What happens if the data center stops taking power, which is the question the whole structure turns on.
• Who operates it, and what the availability obligation is.

The reason this matters commercially, rather than just structurally, is speed. On-site generation is used because it takes the interconnection queue off the critical path — the problem described in [speed-to-power](/learn/speed-to-power). A project that has decided to build its own power has already accepted a substantial capital commitment on a schedule; deciding late how that commitment will be held and funded is how the layer ends up on the wrong balance sheet at the wrong tenor.

The models differ in who holds the asset and what the data-center entity actually contracts for. All three are ordinary; none is universally right.

Once the plant sits in its own entity, the security package looks like an ordinary small power project rather than like part of a building.

The components:

• The equipment, identified at unit level. Serial-level schedules matter for the same reason they matter on any equipment financing, and the mechanics are covered in [equipment finance for turbines and gensets](equipment-finance-for-turbines-and-gensets).
• Site rights. A lease, easement or licence over the land the plant stands on, with access rights, and with a term that outlasts the financing. A generator on land the generation entity does not durably control is an asset that can be locked out of its own site.
• The power services or supply agreement, assigned. This is the revenue, and without it the equipment is a set of machines with no customer.
• The fuel arrangements, including supply contracts and any pipeline or delivery infrastructure rights. Fuel is the plant's largest recurring input and its availability is a condition of performing the supply contract at all.
• Permits and consents, and the ability to transfer them to a successor. Emissions and operating permits are frequently the constraint on whether a plant can run at the hours the contract assumes.
• The operating and maintenance arrangements, including long-term service agreements, which are what stand behind an availability obligation.
• Equity in the generation entity, pledged, so enforcement can deliver the whole project rather than a set of separate assets that have to be reassembled.

What the package is testing is whether a successor could step in and keep the plant running: reach it, fuel it, operate it, and keep selling its output. Where any one of those depends on a consent nobody sought, the position is weaker than the documents suggest — the same test applied to compute in [collateralizing compute](/compute/collateralizing-compute), on a slower-moving asset.

The plant lasts a long time. The question is what it is contracted to do, and for how long, because that is what the capital is actually matched to.

The supply agreement between the generation entity and the facility does the work a power purchase agreement does in a conventional project: it converts a machine into a cash-flow stream. What it has to establish is familiar credit territory — whether the obligation is firm or consumption-based, how long it runs, what the availability requirement is and what happens when it is missed, how fuel cost is passed through, and what termination rights exist on either side.

Three relationships have to hold, and they are the practical test of whether the structure is coherent:

The supply agreement must outlast the financing behind the plant. Otherwise the plant's capital depends on a recontracting that may not happen.

The supply agreement must outlast, or at least match, what the facility has promised its own customers. A hosting agreement committing capacity for longer than the power behind it is a commitment the facility cannot keep — the seam described in the capital-stack apex, and the one that most often goes unchecked because the two contracts are negotiated by different teams.

The site rights must outlast both. A generation entity whose lease over its own footprint expires before its supply obligation is a countdown, not a position.

Where the offtaker is an affiliate — the common case in the sponsor-owned separate-entity model — the contract is still read as a contract. An intra-group agreement drafted thinly, on the assumption that both sides are the same people, is exactly the document that fails to support financing at the generation entity and has to be renegotiated later with a third party's counsel involved.

The distinctive risks of this layer are concentration and stranding. The rest are ordinary project risks.

• One customer, one meter. A plant built behind the meter of a single facility has exactly one buyer. If that facility fails, is not built, or stops taking power, the plant's revenue does not decline — it stops. This is the risk that most distinguishes on-site generation from a merchant or grid-connected plant, and it is the reason the offtaker's credit is examined as closely as the equipment.
• No route to anyone else. The follow-on question is whether the plant can sell to the grid or to a neighbouring load if its customer disappears. That depends on an export interconnection, the applicable rules in that market, and whether the equipment is physically and contractually able to serve another buyer. A plant with an alternative has a residual; a plant without one has a scrap value and a relocation problem.
• Fuel tenor shorter than the plant's obligation. Supply arrangements that run for a period shorter than the power contract leave a repricing risk in the middle of the structure, and the generation entity carries it unless the contract passes it through.
• Permits that cap the run hours. Operating consents can limit how a plant may run. Where the supply contract assumes availability the permit does not support, the obligation and the licence disagree, and the obligation is the one with damages attached.
• Delivery and commissioning slipping past the date the project needed. The layer exists to compress the schedule, so a delay here defeats the reason it was built — which is where it meets [bridge to energization](bridge-to-energization).
• The layers were never actually separated. Generation inside the facility entity, selling nothing, under no contract, financed on the building's terms. The structure that was supposed to attract energy capital never became visible to it.

Continuum structures and arranges around these assets and coordinates the parties in them. It does not own, operate or supply generation equipment, does not sell power, and is not a bank, a broker-dealer or a direct lender; it does not hold client funds. Equipment specifications and generation engineering are covered by Pantheon, not here.