Project Finance & Capital Formation (Mechanics)

Chapter 1.8 compared GPU buy/rent and retained-fleet economics over an assumed earning 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. Project funding, debt sizing, CFADS and coverage live here; GPU buy/rent remains in 1.8. The two are chained. The contracted-vs-merchant split that 1.8 treats as a revenue-quality input is, in this chapter, one input to debt sizing and the DSCR covenant, alongside costs, reserves, collateral advance limits and sponsor support.

The mechanics, in order: the project-vs-corporate finance fork and the SPV security package; the underwriting model, a multi-year pro-forma from capacity ramp to debt service to levered cash flow, with the WACC/DCF/IRR/NPV/DSCR machinery that turns it into a financing decision; anchor-tenant and offtake underwriting, with the counterparty, concentration and residual-liquidity risks under every structure; the PPA as the second underwriting pillar; then 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. Running underneath all of it: Morgan Stanley’s July 2025 scenario allocates $1.5T of 2025–2028 global data-center funding to external capital, and the structures invented to bridge that gap are only as sound as their most-contested assumption, usually the residual value of a depreciating chip. A debt-coverage pass answers the lender’s test. The sponsor still needs a separate equity return after construction draws, debt service, reserves and retained losses.

The first fork is 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 one project's underperformance is absorbed by the firm. It can avoid a separate asset-level security package 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, recourse stops at the SPV only to the extent the financing documents provide: completion support, guarantees, indemnities and negotiated carve-outs can still reach the sponsor. Legal isolation is also not accounting deconsolidation; VIE/control analysis determines whether the parent consolidates the SPV. Project finance 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. One structure 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 March 31, 2026 facility is the canonical example: non-recourse, ring-fenced to a dedicated entity, repaid solely from that vehicle, and secured by identified GPU infrastructure and a customer contract. Before attributing that collateral to Meta, obtain the DDTL 4.0 customer-contract schedule or an issuer disclosure identifying the customer; CoreWeave’s release does not name it. → the operating-archetype economics in Chapter 1.8; the contract instruments that populate the security package in Chapter 2.4. The same sponsor’s August 7, 2026 DDTL 5.5 agreement, announced August 10, instead discloses an unconditional parent guarantee. An SPV and a DDTL label do not determine recourse.

Meta's Hyperion JV, announced October 21, 2025, shows the hyperscaler version. It is 80% owned by Blue Owl-managed funds and 20% by Meta. Beignet Investor LLC raised $27.3B of A+-rated debt to finance the Blue Owl side. Meta's initial four-year leases have renewal options, and a capped residual-value guarantee covers qualifying non-renewal or termination during the first 16 operating years. Meta's June 2026 filing reports a declining guarantee threshold of about $28B and $46.03B of maximum venture exposure, including its equity, leases, funding commitments and guarantee. Meta does not consolidate the VIE because it lacks power over the remarketing and sale decisions that most affect economic performance. Deconsolidation can coexist with substantial sponsor exposure, so assess the retained commitments separately from legal ownership and accounting treatment.

If you choose project finance, the SPV 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 and 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 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 worth only what the secondary market will pay for it at 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 → CFADS → debt service → levered cash flow, and the lender re-prices each link.

The capacity ramp is the top line: MW energized and GPUs racked over time, generation by generation. Revenue does not begin at financial close. It begins when contracted capacity is accepted and starts billing, which means energization under a capacity lease and the accepted reference run on a self-operated cluster (Chapter 2.1). A six-month interconnection slip pushes the whole 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, staff, maintenance and goodput engineering, priced against measured losses rather than the April 2026 model’s ~6–21% goodput-expense scenario) to produce EBITDA. From EBITDA the model makes the financing agreement's cash adjustments, including cash taxes, working-capital changes and maintenance capex, to reach cash flow available for debt service (CFADS). DSCR = CFADS / (cash interest + scheduled principal). After debt service and any reserve funding not already deducted in CFADS, the remainder is levered free cash flow.

Keep two cash-flow series and discount each at its own rate: unlevered project cash flow at the WACC for project NPV, and equity cash flow at the equity hurdle rate for equity NPV. Unlevered and levered IRR are solved from their own series rather than produced by discounting at a chosen hurdle. MOIC is equity distributions over equity invested, and minimum/average DSCR is read period by period 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. The sensitivity tornado ranks which assumption moves the IRR most, and in an AI deal two bars almost always top it: 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.

Follow cash through construction, amortization and reserve release before approving equity. The case below holds the earning life fixed and changes only full-use receipts.

A lender pass can still destroy sponsor value. Base debt = capex × leverage. Construction interest = debt × annual rate × average construction exposure. Operating debt includes that capitalized interest. After grace, annual payment = operating debt × rate / [1 − (1 + rate)^{−amortizing years}]. Reserve = payment × reserve months / months per year. Debt plus capex equity plus reserve equity must equal capex plus construction interest plus reserve.

For each operating year, revenue follows the disclosed ramp. EBITDA = revenue − opex; tax = max(zero, EBITDA − depreciation − interest) × tax rate; CFADS = EBITDA − tax − sustaining capital. Principal = scheduled payment − interest after grace; DSCR = CFADS / (interest + principal). Unlevered project cash uses tax without the interest deduction. Equity cash subtracts debt service and includes close reserve funding and final reserve release. Discount each cash flow at its time index, and calculate with full precision before displaying the table.

Refuse the base equity investment despite the coverage pass. Reserve funding, reserve release and sale proceeds never enter CFADS or cure that ratio. Higher full-use receipts reverse the decision only after the equity-NPV crossover while the coverage covenant remains satisfied. Price and contract those receipts before treating the flip as bankable; reserve timing cannot supply a missing earning year.

The Microsoft NPV convention separates close cash from end-period flows. The project-finance calculator applies it to this draw and amortization bridge. 1.8 owns GPU buy/rent and earning life; this chapter owns debt, CFADS and funding. Feed the rent/compute acceptance conditions in 2.4 into the ramp.

Lenders assess who pays, enforceability, delivery conditions and recovery if payments fail. 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 has investment-grade credit at underwriting and must meet the agreed credit triggers. That single contract can carry an entire non-recourse financing. It is what let CoreWeave's March-2026 facility, secured by identified GPU infrastructure and one contracted customer agreement the closing release declines to name, 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. First, that the anchor's contract will be renewed: most backlogs are a few large multi-year deals, not a diversified book. Second, that the tenant's own demand is durable: a tenant financing its growth on the same contested AI thesis is not an independent credit. Third, that the contract's termination economics are real: a take-or-pay with a soft exit is not take-or-pay. Underwriters therefore read the offtake as the credit and the hardware as the recovery, and 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, power is the cost pillar, and lenders treat deliverable power as a financing precondition, not a nice-to-have. Energy is the largest opex category in Epoch AI's May 2026 modeled 1 GW GB200 campus, about $0.6B a year. The lender needs evidence of deliverable power and an allocation of delivered-price risk. Physical supply arrangements establish capacity and delivery obligations; a financial PPA is a price hedge and requires separate supply. Model tariff changes, basis, shaping and network charges, tenant reimbursements and remaining hedge exposure.

The decision that matters mechanically is to match hedging to debt service where the project bears energy-price risk. A 5-year hedge beneath 7-year debt leaves a two-year merchant tail, a window where the project is exposed to spot power prices with debt still outstanding, unless tenant reimbursement, another hedge or the lender's downside case covers it. Without that cover, a lender will price the tail punitively, require a reserve against it, or refuse the structure. The mirror risk is curtailment. A flexible or 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. This section covers 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. Under ASC 842, a lease is a finance lease when it meets any classification criterion, such as a term covering a major part of remaining economic life. The lessee recognizes and amortizes a right-of-use asset and records a lease liability, and the lessor can retain unguaranteed residual exposure. A GPU finance lease therefore brings the depreciation debate onto the lessee's books. An operating lease no longer keeps the asset off the lessee's books. Since 2019, IFRS 16 and ASC 842 put essentially every lease longer than 12 months on the balance sheet as a right-of-use asset plus lease liability, and IFRS 16 abolished the lessee operating/finance distinction outright. What survives is the P&L geometry. Under ASC 842 an operating lease runs through earnings as a single straight-line lease expense instead of separate depreciation and interest, the lessee still records a right-of-use asset that is impairable under ASC 360, and the obligation often sits outside legacy debt-covenant definitions. That profile makes colocation and build-to-suit attractive to a tenant who wants megawatts without the silicon's residual risk.

Two sale-leaseback variants recur. Derecognition applies only when the transfer qualifies as a sale under ASC 606 and ASC 842-40; repurchase or fixed-price purchase options routinely fail that test and leave a financing. A shell sale-leaseback sells the long-lived, real-estate-like building and power plant to an infrastructure investor and leases it back. That monetizes 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 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. Choose a shell sale-leaseback when released capital exceeds the value of the new rent, control and exit obligations. Choose GPU ownership or a shorter lease by the workload life, return conditions and residual exposure in 1.8. → build-vs-lease economics and the option premium in Chapter 1.8. Carry maintenance, renewal, purchase options and remaining lease payments into the equity cash series. Legal title, tax ownership and financial-statement recognition are separate tests; none follows from the transaction’s name alone.

Where an AI campus pairs on-site renewables or storage with its load, the federal tax credits become a real piece of the capital stack, and one the engineering team is usually surprised to find on the financing diagram. As of October 5, 2026, section 48E covers qualifying zero-emissions electricity generation and energy storage, while legacy section 48 remains available for qualifying grandfathered property. These sit in the Investment Tax Credit (ITC) family. Fuel-cell eligibility follows construction dates: legacy section 48 requires construction before 2025, and a project beginning in 2025 must satisfy ordinary section 48E eligibility, including the emissions test. Under Public Law 119-21, sections 70513(f) and (g)(4), qualified fuel-cell property beginning construction after December 31, 2025 has a flat 30% section 48E credit without the zero-emissions test or bonus adders, subject to the remaining eligibility rules. The Production Tax Credit (PTC) applies per MWh to eligible generation. Direct use depends on tax capacity: a developer in net operating losses must assess carryforwards, an eligible transfer or elective pay before turning a credit into construction cash.

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. That market has matured rapidly, and a transfer needs a priced transfer agreement, an eligible credit and an allocation of recapture and indemnity exposure. As of September 8, 2026, the IRS Form 7211 instructions for the production credit, Form 3468 instructions for the investment credit, including storage and Notice 2026-15 set the dated eligibility tests. 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. Two further gates differ by technology and decide a hybrid's schedule. Under the OBBBA wind-and-solar termination, a project beginning construction after July 4, 2026 generally must be placed in service by December 31, 2027 to keep the credit, while energy storage sits outside that accelerated termination. IRS Notice 2025-42 tightened how beginning of construction is established, limiting the five-percent safe harbor for the affected wind and solar facilities. Underwrite the solar leg and the storage leg on their own dated eligibility tests, never on one. In short: 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 phases with different risk profiles, and conflating them is a classic mis-structure. Separate construction risk from operating risk. The construction phase carries completion risk, since the asset is not yet earning and may never finish, so the bridge is typically higher-cost and shorter-tenor. A construction bridge may be drawn against milestones, and a delayed-draw term loan (DDTL) specifies delayed funding and can also provide amortizing operating finance. 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, because each gate is a point where the lender re-confirms the project before releasing more capital.

A construction bridge is repaid or converted under its agreed terms, often through permanent financing (a long-tenor term loan, a private-credit facility or an ABS) after completion and operating conditions are met. Refinancing risk arises when an unpaid balance depends on capital not already committed. CoreWeave's DDTL series shows how much terms can differ across transactions. At March 2026, CoreWeave reported a 15% effective interest rate on DDTL 1.0 and approximately 5.9% for DDTL 4.0's closing fixed tranche; these are different transactions and rate measures. The series kept extending through 2026. DDTL 5.0 ($3.1B) was the first publicly syndicated delayed-draw facility backed by HPC infrastructure. DDTL 5.5 ($2.6B; agreement dated August 7, 2026, announced August 10) pushed the template into renewal risk: a ~5-year facility drawn against customer contracts averaging ~3 years, so lenders are now financing the re-lease assumption itself. CoreWeave's August 10, 2026 release reports Ba2 / BB+ ratings and SOFR + 5.50% for DDTL 5.5, versus DDTL 4.0's A3 / A(low) ratings and SOFR + 2.25% floating tranche at its March closing. Different guarantees, collateral and contract terms prevent a like-for-like savings inference. Match the bridge maturity to a documented repayment or conversion plan; the gap between bridge maturity and permanent financing is where projects die. → the EPC milestones the draw schedule keys to in Chapter 2.4.

Compare the disclosed commitments and interest bases only after separating guarantees, tranches, draw conditions and maturity. A lower quoted margin on a different security package is not a like-for-like financing saving.

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. Two flavors are routinely conflated, and the distinction decides how the debt is rated. Data-center ABS and CMBS finance facility-related cash flows through different structures. ABS commonly uses a master trust backed by tenant receipts and, where owned and pledged, property assets; some issuers securitize leasehold and tenant-contract cash flows without owning the real estate. CMBS securitizes mortgage loans, including SASB/large-loan transactions, rather than using the ABS master-trust template. The market printed $11.4B in 2024 ($8.4B of ABS plus $3.0B of CMBS) and roughly $26B in 2025 on preliminary figures (~$15B ABS plus ~$11B CMBS). Morgan Stanley’s July 2025 forecast path for US data-center ABS ran $10B in 2025 toward $37B by 2028 as the first refinancing wave hits, a baseline the ~$15B actually printed already outran. CREFC’s February 2026 primer cites Morgan Stanley’s 2026 US ABS-plus-CMBS forecast of $28B net issuance plus $5.5B refinancing, or $33.5B gross issuance. The rating framework follows the financing vehicle, debt structure, cash flows and property rights. GPU-backed financing pledges the silicon and its contracts, and that is the contested frontier. CoreWeave’s March 2026 DDTL 4.0, the first GPU-backed financing rated investment grade, received A3 / A(low) ratings on financing secured by HPC infrastructure and an associated customer contract.

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 and the ring-fencing. They treat the hardware as recovery value, haircut hard for residual uncertainty. Read the rating against the complete contract, collateral and structural protections of the specific financing.

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 anchor-lease, EPC and equipment instruments these structures assign into the security package in Chapter 2.4; the operating-archetype economics in Chapter 1.8.