Who keeps the risk in hardware financing?

Five claims on the same hardware cash

Dated entry guide · October 26, 2026. These instruments can coexist. A guarantee changes recourse; it does not supply a second pot of deployment cash. Customer prepayments arrive with a delivery obligation.

What the current edition adds

The CoreWeave debt bridge reconciles $35.551 billion before unamortized costs to $35.068 billion of debt carrying value at June 30, 2026. It separates the $8.5 billion DDTL 4.0 commitment from $2.837 billion outstanding, and the September $4.2 billion convertible closing from its earlier $3.7 billion pricing announcement. June filing · September closing

The GPU stress model changes accounting life independently of earning life, utilization and resale. The project waterfall follows contract acceptance, construction and power through to equity distributions; use the AI megawatt calculator before turning a facility-power claim into a GPU count.

The Waymo update connects dated financing to manufacturing, vehicles and paid service. The airline comparison explains recurring partner receipts and cash control, then identifies the limits of applying that collateral logic to GPU contracts.

Original July investigation and later DDTL / humanoid additions

CoreWeave’s August 2026 GPU facility matures in about five years, while its initial customer contracts average about three. I’d underwrite the second contract before crediting the collateral as protection. The $2.6 billion subsidiary facility also carries an unconditional parent guarantee, so a weak asset pool can reach the wider company. Filed terms.

Borrowing, leasing, and advance payments let hardware companies deploy with less equity per machine. They add rent, debt service, and delivery claims to the same service cash. A lender can have a clear claim on equipment that few solvent buyers want when the original operating thesis fails.

The company’s funder may never own the machine

A venture round funds a company’s future possibilities. An equipment loan finances a narrower pool of assets and repayments. A customer prepayment brings forward cash for service that still has to be delivered. One fleet can use all three, but the money arrives in exchange for different rights.

The capital stack becomes most revealing when it is read against the same operating timeline. The machine arrives, acceptance or billing begins, payments fall due, the first customer contract ends and replacement eventually becomes necessary. Financing is useful when those dates fit the claim being funded. A corporate funding headline leaves most of that fit unexplained.

The payment survives the growth plan

Equity absorbs a disappointing deployment through a lower return. Debt and rent introduce a payment calendar. The operator must meet it even when paid use falls, unless the contract gives it a way out.

Customer and public funding have their own clocks. A customer who pays early is buying delivery; a grant may require jobs, spending or milestones before the company can keep the benefit. Cash arrives with a job to do.

The signed promise gives creditors legal recourse. Commercial incentives can go further: a parent might support a legally nonrecourse vehicle to protect customers or future funding. The analyst must identify that possibility as discretionary support and keep it apart from the guarantee. A credible analysis keeps the enforceable obligation and the incentive to rescue it on separate lines.

Waymo buys time with parent equity

Waymo’s 2024 and 2026 rounds total $21.6 billion of visible equity, with Alphabet supplying a significant majority in 2026. In the July comparison, that is a patient parent backstop without an identified large fleet facility setting the refinancing calendar.

The cars still consume capital. Charging and depots have to be built, vehicles maintained, and autonomy hardware refreshed before a growing fleet can pay its way. Nothing public shows those obligations moving to a lender, so operating losses and impairment remain with Waymo equity and, economically, Alphabet.

That is flexibility purchased by concentration. Alphabet can tolerate a longer ramp than a lender demanding scheduled repayment, but it also carries the opportunity cost and can change its willingness to fund the next one. The cited disclosures leave vehicle titles, depot leases, liens and hardware ownership unresolved.

Serve’s robot has to earn its equipment payment

Serve layers Farnam secured equipment financing and a Magna supplier warrant onto corporate equity. It can deploy a robot with less purchase cash from shareholders, while the lender receives collateral and the supplier receives equity economics.

The July model exposes the early mismatch. FY2024 cash capex was 5.66× revenue. At modeled base utilization of 48.2%, fleet revenue is about 0.28× the Farnam payment, before field operations, maintenance or replacement. That ratio tests revenue against the payment; even on that narrow comparison, the fleet needs more paid work.

A supplier warrant can help secure production and align incentives, but its economic cost includes dilution. Concentration and minimum-volume dependencies also belong in the assessment where the agreement imposes them. Financing the next robot becomes attractive when its service cash grows faster than the fixed claims, rather than merely because another robot can be financed.

CoreWeave borrows against work it still has to perform

CoreWeave’s FY2025 model records $21.615 billion of debt principal, $8.195 billion of operating-lease obligations, $254 million of finance leases and $8.185 billion of deferred revenue. Selected asset financing equals 78.6% of gross property and equipment; capex is 2.01× revenue.

This stack finances capacity before all of its service revenue is earned. Lenders have secured claims, landlords have rent due, and customers who prepaid expect working capacity. Deferred revenue brings cash forward without releasing CoreWeave from the cost of delivery.

The model’s operating-cash-flow debt-service coverage is 0.70×. Removing the deferred-revenue increase leaves negative $1.116 billion of operating cash flow. These analytical adjustments show how much the funding depends on customer cash arriving early; they use no governing covenant definition.

A lender is underwriting more than the customer’s credit. Equipment can start incurring financing costs before it is accepted for service; a contract can end while principal remains; a new hardware generation can change the next customer’s price. The same asset may be well matched to its first contract and poorly matched to the debt’s final years. The maturity and renewal assumptions therefore belong beside the initial revenue commitment.

A five-year loan outlasts a three-year customer contract

DDTL 5.5 makes that timing visible. The August 2026 facility provides $2.6 billion of committed delayed-draw capacity to CoreWeave Financing DDTL V-V, LLC, an indirect subsidiary, primarily for GPU servers and related infrastructure needed to perform customer contracts. Committed capacity is the credit made available; the amount drawn at closing needs its own source. CW1

The loan matures on September 1, 2031, about five years after closing, while customer contracts average roughly three years, leaving an interval in which the initial buyer can have paid every promised dollar and the equipment must still find another paying use to retire its financing. The second contract is part of the credit case. CoreWeave’s closing release includes renewal or re-leasing in repayment.

The June filing describes DDTL collateral generally as assets underlying contributed customer contracts and pledged contractual cash flows, often involving investment-grade counterparties. That gives a lender a defined customer and asset pool to underwrite. The remaining years require a second judgment about what the equipment can earn after the first contract. CW2

The 8-K summarizes minimum 1.35-times debt-service coverage after a specified trigger. The agreement’s cash-flow numerator, service denominator, and start date control that test. I’d use those definitions before treating any revenue-to-interest ratio as protection.

Delayed draws postpone interest on money the project does not yet need. The assumed example below still charges for an unused commitment; once drawn, the balance costs the benchmark plus the margin. Amortization, reserves, hedging and transaction costs remain outside that narrow comparison.

The collateral has to survive the state in which it is needed

CoreWeave’s June filing describes special-purpose consolidated subsidiaries holding financed infrastructure and related customer contracts. Accounting consolidation does not make every pledged asset available for other corporate uses. The filing distinguishes nonrecourse facilities, with specified carve-outs, from parent-guaranteed debt; the later DDTL 5.5 carries an unconditional parent guarantee. “SPV debt” is too broad a label to describe that difference. June filing; August guarantee.

The reserves show what protection costs. For DDTL 4.0, the June filing describes restricted cash based on forward-looking interest, principal, swap settlements and operating expenses. That buffer supports the particular lender and operating pool. It is also liquidity unavailable to finance the parent’s next deployment. A well-protected loan and a cash-hungry corporate expansion can coexist. Debt note.

Shleifer and Vishny’s 1992 paper explains a further weakness in treating equipment as independent protection. Specialized assets may need buyers from the same industry. When an industry-wide shock forces sales, those buyers can be financially constrained too; the liquidation price can fall below the value in best use. Buyer financial capacity matters alongside the physical condition of the equipment. Liquidation and debt capacity.

For a GPU pool, I’d ask whether another operator could obtain the power, network, software, and customers needed to earn from it during industry distress. A transferable asset with solvent buyers can recover well. Specialized robots and unfinished plants may need a different buyer and completion plan.

The financing case is strongest when contracted cash can retire the debt conservatively and the residual asset is an additional source of protection. It becomes more speculative when repayment depends on a buoyant resale market precisely after the operating thesis has failed. A lien determines who can enforce a claim; the recovery market determines what that claim can collect.

Anduril has to earn the right to keep public support

Anduril combines venture capital, customer-linked demand and public support. The July model compares a $310 million JobsOhio grant with $910.5 million of capex, or 34.0%. It estimates $426.1 million of base net public benefit and about 0.88× public support to private capital.

Those figures describe announced or modeled economics. The company benefits when support is funded, collectible, and protected from clawbacks after compliance. A grant, a tax credit and a public guarantee help through different mechanisms and on different dates.

A customer contract can support factory demand without paying the construction bill. A grant can reduce that bill without ensuring a profitable product. Anduril still needs private capital to bridge the dates and qualified output to justify the capacity; a delayed or resized program can weaken the business even after incentives have made the factory cheaper.

Northvolt’s rescue cash could not refill the old stack

Northvolt financed factories with equity, project debt, customer commitments and prepayments, public finance, supplier exposure and bridge capital. When the ramp disappointed, those sources became competing claims on a weaker operating asset.

The July model places a $100 million debtor-in-possession facility beside $5.8 billion of filing debt: 1.7%. Bridge liquidity is 4.22% of debt; modeled base residual exposure is $4.4 billion. Rescue money can preserve the option to operate or sell while leaving a large loss in the existing capital stack.

An offtake agreement promises a buyer. The factory still has to make acceptable batteries at an economic yield and cost. Customer prepayments may become delivery or refund claims, project lenders keep security, and suppliers can have liens or enough operating leverage to shape a rescue.

Recovery then depends on what the unfinished plant can be sold for, what it costs to complete, and which contracts transfer with it. The $4.4 billion measures the modeled exposure; recovery prices remain unappraised. Financial closing bought the chance to execute; it never guaranteed the ramp.

Humanoid equity is financing the experiment and the machine

Apptronik’s February 2026 announcement combined a $520 million Series A-X extension with a $415 million initial Series A, taking the round above $935 million and total capital raised to nearly $1 billion. Figure announced more than $1 billion of Series C commitments in September 2025 at a $39 billion post-money valuation. These are venture and strategic equity rounds carrying technology and deployment uncertainty.

A mature equipment lender can estimate service life, collections and recovery from a record of comparable machines. A pre-scale humanoid may change between generations and depend on proprietary controls, repair and customer integration. Equity is suited to bearing that experiment because it shares the upside if the platform works. Figure’s company valuation does not appraise a foreclosed robot, and a strategic investor does not automatically create a transferable customer contract.

The use of funds explains the scale. Figure said the round would expand BotQ manufacturing, deployment into homes and commercial operations, GPU infrastructure for Helix training and simulation, and data collection. Apptronik described production and deployment of Apollo. The same capital is paying for a factory learning curve and an AI learning curve.

Apptronik’s disclosed investors include Google, Mercedes-Benz, AT&T Ventures and John Deere. A strategic investor can bring manufacturing knowledge or a route to a customer, but an investment is still a different commitment from a paid fleet order. A pilot can end before purchase, and an order can require acceptance before payment.

Eventually, accepted fleets could support debt while equity funds the next generation. I would want a cohort’s installed cost, billable hours, collections and repair history before assigning that debt capacity. Factory capacity and demonstration videos supply neither the collections nor the recovery record.

Financial recourse allocates the loss

A lessor may hold title while the operator guarantees residual value. A first-lien lender may control specialized equipment with a thin resale market. The title, enforceable recourse and recoverable operating value are three different pieces of protection. Each needs its own evidence.

A decline in paid use first reduces the cash supporting fixed claims. A reserve can absorb the shortfall; a guarantee can move it to the parent; foreclosure can transfer the asset. None of those steps necessarily restores the service income. The recovery analysis should test a common shock to usage, renewal prices and the finances of likely buyers, as well as isolated equipment failures.

The indicators below identify where that pressure enters each structure. Serve’s payment diagnostic, a public-support ratio, and a contractual debt covenant each test a different kind of pressure. Their value lies in tracing the claim that can become binding, rather than ranking unlike companies by a single financing percentage.

Replacement can arrive before the old debt is gone

Consider an assumed $100 asset financed with $20 of equity and $80 of debt repaid evenly over four years. After two years, $40 of principal remains. If a technology shock makes the asset unusable with no resale value and a replacement still costs $100, the operator needs the new machine while carrying the old $40 claim, before interest and operating costs.

The debt schedule has not extended the machine’s economic life. Insurance, transferable customers, reserves or a restructuring can change the outcome, but each needs an identified contract or funded source.

Put replacement, remaining principal and committed service cash on the same timeline. Debt can efficiently finance accepted, repeatable cohorts while equity funds the uncertain next generation. The danger is using a stable payment schedule to imply a stable asset life that the operating evidence does not yet support. A good structure lets a reasonable miss be absorbed without turning every replacement into a fresh corporate rescue.

The recovery needs a buyer as well as a lien

The table distinguishes documented financing terms from inferences and assumed recovery paths. Assigning a recovery price still needs titles, enforceable transfer rights, customer continuity, and actual distressed bids. A lien determines who can enforce a claim; the recovery market determines what that claim can collect.

A cheaper deployment can leave the same loss

I prefer financing that follows a proven asset’s collections to financing that asks future growth to rescue old equipment. A good lender can be protected while the operator remains fragile. The attractive hardware business is the one whose capacity pays for service, replacement and its capital claims through a weaker year—not simply the one that deploys the most machines per dollar of new equity.