When AI Credit Eats Its Tail

$100 Billion Meets Its Own Purchase Order

AI infrastructure has entered a phase where capital is part of the product.

NVIDIA can invest in a model company that buys NVIDIA systems. Google can earn Marvell equity as Google-linked purchases create Marvell revenue. Broadcom can support the financing vehicle that buys Broadcom-designed racks for an AI customer. Private-credit firms can turn compute leases into long-duration assets. The cash, chips, equity and demand begin to form a loop.

The largest announced number is $100 billion. In September 2025, NVIDIA and OpenAI announced a letter of intent covering at least 10 gigawatts of NVIDIA systems, with NVIDIA intending to invest up to $100 billion as capacity was deployed. In February 2026, OpenAI announced a $110 billion financing round that included $30 billion from NVIDIA, alongside dedicated inference and training capacity on NVIDIA systems.

Another loop carries a different shape. Google’s warrant on Marvell covers as many as 58,970,907 shares at $206.58, a gross nominal exercise amount of roughly $12.18 billion. Almost all of the award depends on $500 million increments of qualifying Google-linked revenue, across a potential $120 billion arc.

Then credit enters directly. Broadcom, Apollo and Blackstone established a $35 billion AI infrastructure platform built around more than one gigawatt of capacity. Broadcom disclosed a maximum $29 billion backstop tied to the initial structure. Later reporting described discussions around a much larger senior and junior financing package. Executed terms for that later proposal remained unknown at the research cutoff.

These transactions share a visual resemblance. Their economics differ.

The investor’s job is to follow the cash before following the excitement.

Three Degrees of Circularity

I use three degrees.

Degree One: alignment after commercial proof.

The Google–Marvell warrant belongs here. Google receives equity upside as qualifying revenue accumulates at Marvell. The bulk of the warrant vests through performance milestones. Google still has to direct purchases into defined commercial programs. Marvell still has to earn recognized revenue.

The loop shares upside. It does not create the purchasing cash by itself.

The main investor questions are program scope, margin quality, dilution and Marvell’s role beside Broadcom in Google’s TPU ecosystem. Performance vesting would provide evidence. Time-based vesting without meaningful performance vesting would send a weaker signal.

Degree Two: ecosystem investment alongside customer demand.

NVIDIA and OpenAI sit here. NVIDIA supplies the platform, invests in the customer, and benefits if the customer deploys more NVIDIA systems. OpenAI gains capital and capacity. NVIDIA gains strategic alignment, platform adoption and a potential equity return.

The demand can be entirely real. OpenAI runs one of the world’s largest AI services. The loop still changes the quality of evidence. Revenue created by a customer that received supplier capital deserves a different annotation from revenue funded by independently generated customer cash.

The relevant question is incremental: how much deployment would occur at the same pace and price without the equity support? Public sources do not fully answer it.

Degree Three: supplier-supported purchasing capacity.

The Broadcom AI infrastructure vehicle moves closest to the credit engine. Third-party lenders fund a special-purpose vehicle. The vehicle buys AI racks. Customers lease the compute. Lease payments service the debt. A supplier backstop can reduce the credit risk borne by senior capital and increase the financing capacity available to buy the supplier’s systems.

The loop can accelerate real deployment. It also moves risk toward the supplier if utilization, lease coverage, refinancing or residual value disappoints.

This degree demands the strongest underwriting.

The Tail Is Useful Until It Carries the Body

Circular structures are common in industrial expansion. Aircraft manufacturers support customer finance. Equipment companies lease machines. Export-credit agencies help buyers fund strategic goods. Semiconductor vendors invest in the ecosystems that make their platforms more valuable.

The structure earns concern when financing begins carrying more of the demand than end-user economics can support.

AI creates special pressure because the capital needs arrive before the revenue model has matured. A data center can cost billions. Chips age quickly. Power and networking require long commitments. Model prices continue falling. Customer revenue may grow fast while gross cash requirements grow faster.

This produces a timing gap. Suppliers have cash and strategic urgency. Model companies have demand and enormous compute needs. Infrastructure investors have long-duration capital and want contracted yield. A special-purpose vehicle can connect them.

The loop becomes productive when each participant earns a return from useful compute.

The loop becomes fragile when a supplier repeatedly expands support to protect shipment growth, customer leases depend on fresh financing, and residual chip value receives a heroic assumption.

The same transaction can move from one condition to the other over time.

A Five-Line Scorecard

Investors need a scorecard that works across structures.

1. Where did the purchasing cash originate?

Independent customer cash provides the cleanest demand signal. Third-party debt with strong lease coverage can remain economically sound. Supplier equity and credit deserve explicit attribution. A commitment is not funded cash.

2. What must happen before value transfers?

Marvell’s performance warrant has visible revenue milestones. An infrastructure loan may draw as racks deploy. An equity investment may arrive before the customer commits to a specific purchase. The sequence determines which party finances uncertainty.

3. Who absorbs failure?

Equity absorbs valuation loss. Junior credit absorbs losses before senior credit. A guarantee transfers some loss toward the supplier. Lease remedies and asset sales may reduce the gap. The contract, rather than the press release, answers the question.

4. Can the asset move?

General-purpose compute with a broad customer base has stronger recovery value. Highly customized racks tied to one customer, software environment or site may have weaker portability. Useful life, energy efficiency and chip-generation cadence matter.

5. Does operating cash eventually replace financing cash?

Every productive loop should mature. Utilization rises. Customers generate revenue. Lease coverage expands. Suppliers receive cash without adding fresh support. If each turn requires a larger capital intervention, the tail is carrying the body.

NVIDIA Turns Compute Into an Asset Class

The direction became even clearer in August 2026 when NVIDIA announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR aimed at mobilizing more than $500 billion of third-party capital over time.

The language matters: compute is becoming an investable asset class.

That evolution can broaden access and match long-lived capital with long-lived usage contracts. It can also move AI demand into structures whose risks resemble project finance, equipment leasing and private credit more than traditional software sales.

Semiconductor analysis therefore needs new muscles. Unit shipments and gross margin remain important. So do lease duration, customer concentration, debt service, guarantee caps, draw schedules, asset recovery and refinancing.

The financial architecture around the chip may become as consequential as the chip roadmap.

The Investment Judgment

The contrarian view is constructive: circularity can be a feature of a capital- intensive transition. It coordinates suppliers, customers and long-duration capital before the market has settled.

The discipline lies in degrees.

Google–Marvell begins with commercial milestones and contingent dilution. NVIDIA–OpenAI combines a large strategic equity relationship with enormous infrastructure demand. The Broadcom vehicle adds supplier-supported credit and asset-recovery risk. NVIDIA’s financing partnerships aim to scale the entire category through third-party capital.

Each loop has a different cash origin, proof gate and failure absorber.

My posture is WATCH, TRACE THE CASH, WAIT FOR OPERATING PROOF.

Track funded cash, deployed capacity, utilization, customer revenue, lease coverage, guarantee exposure, vesting, margin conversion and refinancing cost. Keep unknown terms unknown. Give commitments no credit until they fund. Give revenue full credit when independent end demand and cash conversion appear.

AI credit can eat its tail and keep moving forward. The decisive question is whether the circle produces useful intelligence or merely another turn of financing.

Decision appendix: what the ledger should carry

A credible circularity ledger needs more than deal size. It should separate legal commitment, funded cash, delivered equipment, useful capacity, customer revenue and supplier cash conversion.

For NVIDIA–OpenAI, the ledger begins with two distinct records. The September 2025 announcement was a letter of intent covering at least 10 gigawatts and up to $100 billion of progressive investment. The February 2026 announcement described a completed financing round with $30 billion from NVIDIA and a capacity plan involving 3 gigawatts of dedicated inference and 2 gigawatts of training on Vera Rubin systems. Those numbers belong in different columns. Intent, investment and deployed capacity are different evidence states.

The next useful observations are cash funded, systems ordered, systems delivered, power available, capacity operational, utilization, OpenAI revenue served by the capacity and NVIDIA revenue collected from the same expansion. Without that sequence, the market can celebrate one announcement several times.

For Google–Marvell, the contract offers a cleaner performance meter. The time-based portion covers 1,360,867 shares. The performance portion covers 57,610,040 shares. Up to 240 revenue milestones, each representing another $500 million of cumulative qualifying revenue, define the commercial path. The measuring period runs from August 2026 into early 2033.

The ledger should track time shares vested, performance shares vested, cumulative qualifying revenue when disclosed, actual Google purchases, the scope of Marvell’s programs and the accounting treatment of the warrant. Actual purchases and Marvell’s share of the TPU stack relative to Broadcom remain unknown until filings or issuer disclosures make them visible.

For Broadcom’s infrastructure vehicle, the confirmed June baseline is a $35 billion platform, more than one gigawatt of capacity, a five-year customer lease and a supplier backstop capped at $29 billion. The later reported proposal around a larger senior and junior structure requires its own status. Discussion is not closing. A reported range is not funded cash. A possible guarantee is not an executed obligation.

The ledger should wait for final tranche size, cash draws, guarantee cap, customer lease obligations, deployment schedule, debt-service coverage, refinancing terms, residual-value assumptions and customer concentration.

The same controls apply to NVIDIA’s announced $500 billion financing ambition. “Mobilize over time” describes capacity to assemble capital. It does not mean $500 billion has been committed, funded or deployed. Investors should measure platform formations, fund closes, drawdowns, projects and operating capacity in sequence.

The ledger’s greatest value is preventing unit confusion. A press release can contain equity dollars, debt capacity, equipment value, potential revenue and market capitalization in adjacent paragraphs. They are not additive.

My preferred dashboard uses five rows for every deal:

Evidence row	Question
Contract	What obligation is legally signed?
Cash	What amount has actually funded?
Asset	What equipment has been delivered and energized?
Use	What capacity is utilized by paying customers?
Return	Which participant has collected operating cash and who still carries risk?

When all five rows advance together, circularity can be a financing technology. When contract headlines race ahead while cash, assets and use remain still, the loop is a story awaiting proof.

For an investment committee, the next quarterly review should update those five rows before debating valuation. That sequence forces press-release scale to meet cash-flow evidence. It also makes disagreement productive: one investor can accept more sponsor concentration while another requires utilization proof, yet both are discussing the same loop.

The decision transfer

The framework applies to every capital-intensive growth market. Score the loop before pricing the equity. Identify the party that supplies cash, the party that buys the asset, the proof required for revenue and the balance sheet that absorbs failure. Then update cash, asset delivery, utilization and return on separate clocks.

This turns circularity from an accusation into a measurable underwriting variable. A strong loop can accelerate real deployment. A weak loop can delay recognition of demand risk. The same scorecard can distinguish them early.

Investors can disagree on price after agreeing on the plumbing.

That shared map makes every later valuation debate more productive.

Sources

• OpenAI and NVIDIA 10-gigawatt partnership
• OpenAI’s February 2026 financing and NVIDIA partnership
• NVIDIA fiscal 2026 Form 10-K
• Marvell Google warrant filing
• Executed Google–Marvell warrant
• NVIDIA compute-financing platform announcement

#AIInfrastructure #AICircularity #NVIDIA #OpenAI #Marvell #Broadcom #CapitalMarkets