Project Finance & Capital Formation (Mechanics)

Chapter 1.8 scored the asset: does this specific build, financed this specific way, earn its cost of capital over its economic life? This chapter is the layer beneath that scorecard — the deal mechanics that decide whether the capital can be raised at all, on what terms, and against which collateral. The financing strategy lives in 1.8; the financing machinery lives here. The two are chained: the contracted-vs-merchant split that 1.8 treats as a revenue-quality input is, in this chapter, the variable that literally sizes the debt tranche, sets the DSCR covenant, and fixes the advance rate against the hardware.

We build the project-vs-corporate finance fork and the SPV security package; we walk the underwriting model — the multi-year pro-forma from capacity ramp to debt service to levered cash flow, and the WACC/DCF/IRR/NPV/DSCR machinery that turns it into a financing decision; we examine anchor-tenant and offtake underwriting and the counterparty, concentration, and residual-liquidity risks that sit under every structure; we treat the PPA as the second underwriting pillar; we lay out lease-vs-own, tax-equity/ITC for on-site renewables, construction-loan-to-perm mechanics, GPU-backed ABS and its ratings treatment, and the developer/operator/capital-provider deal archetypes. The through-line: the 2026 build-out has outgrown self-funding, and the structures invented to bridge the gap are only as sound as their most-contested assumption — usually the residual value of a depreciating chip.

The first fork is also the most consequential, because it decides who is on the hook when a project fails. Corporate finance raises debt against the whole enterprise: the parent's balance sheet, cash flows, and credit rating back the borrowing, and a single project's underperformance is absorbed by the firm. It is fast, covenant-light relative to the asset, and the natural mode for an investment-grade hyperscaler that can fund a campus out of operating cash flow at the lowest cost of capital in the industry. The price is that risk is not ring-fenced — leverage piles onto the consolidated balance sheet, and rating agencies watch the aggregate.

Project finance does the opposite: it carves a single asset into a bankruptcy-remote special-purpose vehicle (SPV), finances it on a non-recourse or limited-recourse basis, and lets the project's own contracted cash flows service the project's own debt. If the project fails, lenders' claims stop at the SPV; the parent is insulated. This is the structure that built merchant power plants, toll roads, and LNG terminals for decades, and it is now the template for capital-intensive AI campuses whose sponsors are not investment-grade — most visibly the neoclouds. The cost of the insulation is rigor: lenders who cannot reach the parent demand a complete security package, hard covenants, and contracted revenue they can underwrite, because the asset is all they have.

The fork is not binary in practice. The dominant 2026 pattern is a hybrid: a hyperscaler-grade anchor lease provides the contracted cash flow, an SPV ring-fences the campus, and the sponsor contributes equity while non-recourse debt does the heavy lifting. CoreWeave's landmark March-2026 facility is the canonical example — non-recourse, ring-fenced to a dedicated entity, repaid solely from that vehicle, secured by the GPU clusters and the Meta contract backlog inside it. → the operating-archetype economics in Chapter 1.8; the contract instruments that populate the security package in Chapter 2.4.

If you choose project finance, the SPV is not a formality — it is the legal machine that makes the asset financeable, and its security package is the list of things lenders can seize. A complete package in an AI-campus deal pledges, at minimum: a first-lien on the physical assets (the GPUs, the cooling and power plant, sometimes the shell); an assignment of the offtake/lease contracts and their revenue; a pledge of the project accounts (a waterfall of revenue, opex, debt-service-reserve, and distribution accounts that pays lenders before equity); a pledge of the SPV's equity (so lenders can take the whole company, not just its parts); step-in and direct-agreement rights with the EPC contractor and the offtaker (so lenders can cure a default and keep the contracts alive); and completion support from the sponsor during construction. The Chapter 2.4 commercial instruments — the EPC contract, the equipment-supply and slot-reservation agreements, the interconnection agreement, the PPA, the anchor lease — are precisely the documents that get assigned into this package.

The package is also where the residual-value problem becomes structural. A first-lien on a depreciating, deflating asset is only worth what the secondary market will pay for it on a foreclosure — which is why the contracts in the package, not the hardware, usually carry the credit. → the contract-instrument detail in Chapter 2.4; insurance requirements that lenders bolt onto the package in Chapter 2.6.

Every financing decision in this chapter reduces to one artifact: a multi-year pro-forma model that a credit committee can stress. It is the same DCF discipline used everywhere in infrastructure, with two AI-specific pressure points: the depreciation schedule and the revenue deflation curve. The chain runs: capacity ramp → revenue → opex → EBITDA → debt service → levered/unlevered cash flow, and each link is a decision the lender re-prices.

The capacity ramp is the top line: MW energized and GPUs racked over time, generation by generation. It matters because revenue does not begin at financial close — it begins at energization, and a six-month interconnection slip pushes the entire revenue curve right while debt service waits for no one. The ramp feeds revenue (contracted $/kW-month or $/GPU-hr times utilization), which nets against opex (power as the largest line, plus staff, maintenance, and the reliability overhead that is 6-21% of TCO) to produce EBITDA. From EBITDA the model subtracts debt service to get levered free cash flow, and the ratio of the two — DSCR, debt-service coverage — is the single covenant that governs whether the deal survives a bad year.

The model is then discounted at the project's WACC / hurdle rate to a DCF, which yields the headline metrics: levered and unlevered IRR, NPV, the equity multiple (MOIC), and minimum/average DSCR across the loan tenor. The contracted-vs-merchant split sets debt capacity directly — lenders advance generously against take-or-pay cash flow and grudgingly against merchant exposure — and the sensitivity tornado ranks which assumption moves the IRR most. In an AI deal the tornado is almost always topped by two bars: utilization and residual value / depreciation life. The companion levered-IRR model and the input register live in Appendix C; the contested figures bind to the dated forecast register in Appendix D; the firm-level ROI scorecard this model feeds is Chapter 1.8.

The lender's first question is never about the GPUs — it is who pays the rent, and can they be made to. An AI campus is financeable in proportion to the credit quality, tenor, and structure of its offtake. The gold standard is a long-dated, take-or-pay anchor lease with an investment-grade hyperscaler: the tenant pays whether or not it uses the capacity, the term matches the debt, and the obligor's credit is unimpeachable. That single contract can carry an entire non-recourse financing — it is exactly what let CoreWeave's March-2026 facility, secured by GPU clusters and the ~$14.2B Meta master service agreement, achieve the first investment-grade rating on GPU-backed debt (A3 from Moody's, A(low) from DBRS).

The lender's view of GPU-cloud demand is where underwriting gets adversarial. Lenders discount three things developers tend to take on faith: that the anchor's contract will be renewed (most backlogs are a few large multi-year deals, not a diversified book); that the tenant's own demand is durable (a tenant financing its growth on the same contested AI thesis is not an independent credit); and that the contract's termination economics are real (a take-or-pay with a soft exit is not take-or-pay). The underwriting discipline is to read the offtake as the credit and the hardware as the recovery — and to price the gap between them. → the anchor-lease and offtake instruments in Chapter 2.4; tenant-facing productization in Chapter 10.9.

If the anchor lease is the revenue pillar, the power contract is the cost pillar — and lenders treat a firm-power arrangement as a financing precondition, not a nice-to-have. The logic is the same as the offtake logic, inverted: energy is the largest controllable opex line (~$0.6B/yr in the 1 GW reference model), so an uncontracted power cost is an uncontrolled hole in the pro-forma. A long-dated PPA, an on-site generation contract, or a firm utility supply agreement converts that hole into a known, bankable number.

The decision that matters mechanically is tenor matching: the power-contract tenor must cover the debt tenor. A 7-year term loan under a 5-year PPA leaves a two-year merchant tail — a window where the project is exposed to spot power prices with debt still outstanding — and a lender will either price that tail punitively, require a reserve against it, or refuse the structure. The mirror risk is curtailment: a flexible/curtailable interconnection buys speed-to-power and cheaper rates, but in the lender's downside it is lost revenue, so the model must show coverage even under the contract's worst-case curtailment hours. The cheapest power on paper can be the least financeable if its firmness does not match the debt. → energy-supply strategy and PPA structures in Chapter 3.4; the interconnection agreement as a commercial instrument in Chapter 2.4.

Chapter 1.8 priced build-vs-own-vs-lease as an NPV with an option premium; here we cover the mechanics of the lease structures themselves, because the accounting treatment changes what shows up on the balance sheet and in the covenants. The pivotal distinction is operating vs finance lease. A finance lease (ownership-like: the lessee bears substantially all risks and rewards, the term covers most of the asset's life) sits on the balance sheet as an asset and a liability, and the asset depreciates on the lessee's books — so a GPU finance lease re-imports the entire depreciation debate. An operating lease keeps the asset off the lessee's books as a usage right, converting capex into a rental opex line — the treatment that makes colocation and build-to-suit attractive to a tenant who wants megawatts without the silicon's residual risk.

Two sale-leaseback variants recur. A shell sale-leaseback sells the long-lived, real-estate-like building and power plant to an infrastructure investor and leases it back — monetizing the stable, financeable part of the asset at a real-estate cost of capital while the operator keeps the IT. A GPU sale-leaseback does the same with the silicon, and it is far more fraught: it transfers a depreciating, deflating asset to a lessor who must underwrite the residual, so the lease rate carries that risk premium and the structure lives or dies on the secondary-market assumption above. The clean version of the decision: sell the shell (long-lived, cheap to finance), keep or short-lease the GPUs (short-lived, expensive to finance against). → build-vs-lease economics and the option premium in Chapter 1.8.

Where an AI campus pairs on-site renewables or storage with its load — increasingly common as operators chase firm, low-carbon power — the federal tax credits become a real piece of the capital stack, and a piece the engineering team is usually surprised to find on the financing diagram. The Investment Tax Credit (ITC) applies to solar, battery storage, fuel cells, and geothermal; the Production Tax Credit (PTC) applies per-MWh to generation. Both have value only to a taxpayer with enough tax liability to absorb them — which a build in net operating losses, like most early-stage developers, does not have.

Two monetization paths bridge that gap. The traditional one is tax equity: a tax-paying investor takes an equity stake structured to absorb the credits and depreciation, providing cash up front in exchange for the tax benefits. The newer, faster path is transferability under IRC §6418: the credits are sold for cash directly to an unrelated buyer, without the partnership machinery — a market that has matured rapidly and now clears confirmed, FEOC-compliant, interconnected projects at close to full value. The 2026 wrinkle is regulatory: the OBBBA's prohibited-foreign-entity rules (operationalized by IRS Notice 2026-15, February 2026) gate eligibility on supply-chain provenance, so a BESS or solar array that beats construction-start safe harbors but trips the FEOC rules loses the credit it was financed on. The mechanic to internalize: tax credits are a financing input, monetizable via tax equity or §6418 transfer, but only if the project clears the foreign-entity and safe-harbor gates — verify eligibility before you underwrite the credit into the model. → on-site generation and storage strategy in Chapter 3.4.

An AI campus is financed in two distinct phases with two distinct risk profiles, and conflating them is a classic mis-structure. The construction phase carries completion risk — the asset is not yet earning and may never finish — so it is funded with a higher-cost, shorter-tenor construction loan, frequently structured as a delayed-draw term loan (DDTL). A DDTL is drawn in tranches against milestones rather than funded in full at close, which matches cash to the build: you draw against energization gates, equipment-delivery gates, and IST/commissioning gates, and you pay commitment fees on the undrawn balance instead of interest on idle cash. The draw schedule is itself a negotiated risk-allocation document — each gate is a point where the lender re-confirms the project before releasing more capital.

At completion the construction loan must be taken out by permanent financing — a long-tenor term loan, a private-credit facility, or an ABS — that refinances the bridge once the asset is operational and de-risked. This is the takeout, and it is the most under-appreciated risk in the structure: the construction loan assumes a takeout exists, and if the capital markets are shut at maturity or the asset has aged into the depreciation problem, the bridge has no landing. CoreWeave's DDTL series is the visible benchmark for how fast the cost of capital can compress as a sponsor de-risks — from ~15% floating on the 2023 DDTL 1.0 to SOFR+2.25% / ~5.9% fixed and an investment-grade rating on the March-2026 DDTL 4.0. The mechanic to hold: never let construction-loan tenor outrun a credible takeout; the gap between bridge maturity and permanent financing is where projects die. → the EPC milestones the draw schedule keys to in Chapter 2.4.

Securitization is how the build-out reaches the deep, rate-sensitive institutional capital that bank balance sheets cannot supply alone — and it is the structure under the most scrutiny in 2026. There are two distinct flavors that are routinely conflated, and the distinction decides how the debt is rated. Data-center ABS / CMBS pledges the real-estate-like asset: the shell, the power plant, and the tenant leases, pooled into a master trust that issues notes (Morgan Stanley projects issuance growing from ~$8B in 2025 toward ~$25B by 2028 as the first wave of refinancings hits; other syntheses put 2026-2027 data-center ABS/CMBS at ~$30-40B/yr). This is conventional, long-lived collateral, and the rating agencies treat it like infrastructure real estate. GPU-backed financing pledges the silicon and its contracts — and that is the contested frontier. The March-2026 CoreWeave facility was the first such structure to earn an investment-grade rating (A3 / A(low)), achieved not because the GPUs are good collateral but because the contracted backlog inside the SPV carried the credit.

The ratings logic is the tell. Agencies rate GPU-backed structures primarily on the quality and tenor of the contracted cash flow — the offtaker's credit, the take-or-pay terms, the ring-fencing — and treat the hardware as recovery value haircut hard for residual uncertainty. That is the right way to read every GPU-backed deal: the rating is on the contract, not the chip.

The final mechanic is who sits in which seat. The 2026 market has disaggregated into three roles that used to be one company — the developer who secures land, power, and permits and de-risks the project to shovel-ready; the operator who runs the live facility and holds the tenant relationship; and the capital provider who funds the build. The deal archetypes are different ways of splitting risk, control, and return across those seats.

A joint venture (JV) pairs a developer/operator with a deep-pocketed capital partner (a private-equity infrastructure fund, a sovereign vehicle, a pension) who funds the build in exchange for the majority of the equity return, leaving the operator with a promote and the operating contract. Build-to-suit (BTS) has a developer construct a campus to an anchor tenant's spec under a pre-signed long lease — the lease is the financing, because its credit and tenor are what the construction debt underwrites against. Forward-funding goes one step earlier: the capital provider funds the development itself, paying as construction progresses against an agreed return, so the developer carries less balance sheet and the funder captures the development margin. The choice among them is a choice about where the equity return and the development risk land: a developer short on capital but long on pipeline takes forward-funding or a JV and trades return for execution; a developer who can fund itself keeps the build and the upside. The common thread with everything above — the anchor lease, the PPA, the SPV — is that the contracted offtake is the asset that makes any of these structures bankable. → the build-to-suit / forward-funding commercial terms in Chapter 2.4; the operating-archetype economics in Chapter 1.8.