The marginal build-out of artificial intelligence is no longer financed exclusively by internal cash generation. That sentence is deliberately narrow, because broader versions of it are false. Meta generated 31.86 billion dollars of operating cash flow in the second quarter against 31.08 billion of capital expenditure; Microsoft still produced 19.6 billion of free cash flow in its latest quarter. These companies are not running out of money. What has changed is the margin: Alphabet reported negative free cash flow of 5.9 billion dollars in the second quarter, its first since listing in 2004, and Amazon's has turned negative over twelve months; both are reported results. Back in May, forecasts compiled by Visible Alpha already put the combined third-quarter free cash flow of the four largest spenders near four billion dollars. Operating cash flow remains enormous. What the numbers establish is narrower: external capital has become structurally material to the marginal build-out.
The gap is being closed with external capital, and it is worth separating the two forms this takes, because they are different objects that the coverage tends to merge. The first is ordinary corporate borrowing: the hyperscalers had issued 194 billion dollars of investment-grade debt by 7 July, incremental debt has gone from nine per cent of capex in the 2024 financial year to roughly a third by mid-2026, and Goldman Sachs expects debt to fund over a third of AI investment next year. The second is something else: a machinery that converts part of that funding need into asset- and contract-backed credit built around the compute: the chips, the data centres and the twenty-year lease payments of those who will use them. The first sits on corporate balance sheets and is boring. The second is building an asset class at their edges, and is not.
The machinery
Photo by ObjectType RAW / Unsplash
The architecture was already visible in February, when the Financial Times described the growth of chip-backed lending. A special-purpose vehicle, formed jointly by a technology company and an investment firm, buys the chips and leases them back; the stated purpose was to keep the borrowing off the technology company's balance sheet. The leases carry "hell or high water" clauses that bar early termination, so obsolescence risk stays with the lessee for as long as the lessee is solvent. Moody's rates the paper on structures that repay in full within the first lease term, so that repayment does not depend on any terminal residual value. And part of the market declined to participate at all, on the ground that nobody knows what a three-year-old GPU will be worth: there is no price history for second-hand AI chips, and current valuations may reflect shortage rather than durable worth.
The documented case of what this looks like at scale is the roughly 200 billion dollar web of contracts the FT reconstructed around Anthropic's compute in August: the first tranche of TPU hardware passes from Google through Broadcom into a special-purpose vehicle organised with Morgan Stanley; the vehicle borrows the money, with Apollo and Blackstone among its major investors, buys the hardware and leases it to Anthropic; Broadcom provides residual-value and credit support to part of the structure. The distinction between what is reported and what is disclosed matters here: the FT reconstructs the contractual chain around Anthropic, while Alphabet's 10-Q independently confirms the scale and accounting treatment of the underlying class of backstops, without naming Anthropic, disclosing a maximum potential exposure of 43.785 billion dollars on its data-centre credit derivatives and a liability carried at a fair value of 815 million. Meta, which provides a particularly large example of the same logic for its own sites, states in its own 10-Q, against an aggregate guarantee threshold of roughly 28 billion dollars, that RVG payments are "not probable", and that no liability has therefore been recorded. On 15 August, before the Ohio guarantee was signed, Bloomberg counted roughly 70 billion dollars of such shadow backstops across Nvidia, Broadcom and Meta, and a JPMorgan strategist gave the layer they form beneath the visible balance sheets a name: phantom leverage, a backlog of leases, purchase commitments and guarantees he expects to stretch into the trillions. Two days later, a single filing would add a contingent cap larger than the entire amount Bloomberg had just counted.
On 10 August the machinery acquired its institutional form. Nvidia announced financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, with the stated aim of mobilising more than 500 billion dollars of third-party capital. For now these are memoranda of understanding; final agreements are still pending. The release describes the goal as turning Nvidia compute into an investable asset class, and asserts the two properties lenders need before they are willing to lend against an asset: the compute, it says, is "fungible and transferable across customers and operators". The following day Nvidia confirmed the principle that had until then circulated as reconstruction: it may provide, case by case, residual-value support for up to a quarter of an opportunity, while stressing that the financing institutions must still underwrite customer credit, demand, utilisation, cash flow and residual value themselves. What that support is worth remains unknowable from the outside; up to 25 per cent of an opportunity is not a floor of 25 per cent under the chips' value, and neither the triggers, nor the ranking, nor the legal form have been disclosed. What can be said precisely is what the support does: it takes the uncertainty that was, in February, the stated reason part of the market stayed away, the terminal value of the hardware, and makes it more financeable. Private capital firms reportedly believe the resulting debt can now be split into layers carrying different levels of risk and different ratings, and placed with the insurers whose assets they manage; CoreWeave already carries a non-recourse, investment-grade facility of 8.5 billion dollars from March. The market is being engineered so that the paper can migrate into insurance and pension portfolios.
The guarantor's own filings complete the picture. In its results commentary for the quarter, Nvidia explains why it stepped into infrastructure credit at all: its AI cloud and model-maker customers are growing "faster than their balance sheets and long-term credit profiles can support". Its quarterly report, filed on 26 August, discloses 366 billion dollars of future commitments, with supply and capacity alone more than doubling in a single quarter from 119 to 279 billion, a further 56 billion tied to AI-cloud agreements and third-party data-centre leases, and 108.5 billion of maximum gross guarantee exposure. And in June the company that anchors the ecosystem issued 25 billion dollars of bonds of its own, taking its debt from 8.5 to 33.4 billion. The seller of the silicon has now built an enormous stock of future commitments of its own, and the web of obligations now grows in both directions: downstream towards customers and financing vehicles, upstream towards supply and capacity.
Where the risk went
It is tempting to say the risk changed hands, and it would be wrong. The risk was tranched. The lessee carries obsolescence for as long as it is solvent; senior holders carry its credit; junior holders carry first loss; the vendor carries a slice of the terminal value; Google carries a contingent exposure it books at fair value. Nothing left the system. It was sliced, layered and redistributed, and some of it is contingent, meaning it returns to the guarantor precisely when things go wrong.
Which is where the structure's real weakness sits, and it is the inverse of what a guarantee is for: precisely when the protection is needed, the protection risks being worth less. Credit risk has a name for this, wrong-way risk. Not every future is dangerous to this paper; a collapse in the price of intelligence could, on the contrary, expand usage and keep the leases paid. The dangerous states of the world are specific ones: overcapacity, demand below forecasts, rapid technological substitution, efficiency gains that cut the compute needed per unit of output, or a loss of pricing power among the leading model providers. In those states, the lessee's cash flow and the recovery value of the collateral can fall together. And the parties standing behind the paper, Nvidia and Broadcom, belong to the same cycle: in an industry-wide adverse scenario, their earnings and credit quality would come under pressure at the same time as their support becomes more valuable to lenders. The correlation is structural, not incidental, and the credit market has begun to price the balance-sheet consequences of that support: Broadcom's five-year credit default swaps rose 28 basis points in August, more than Oracle's, as it negotiates a further financing package of more than 60 billion dollars.
The assumptions about future demand are also exposed to forces no contract can contain. China is explicitly pursuing an international open-source AI ecosystem: the action plan its planning commission published on 17 July commits to inclusive compute services, transnational open-source communities and broader sharing of general-purpose AI capabilities, principles Xi restated the same day; party media describes the same strategy more aggressively as "open, low-cost and decentralised". Free frontier-adjacent models can compress the pricing power on which the lessees' ability to pay partly depends, even as they expand inference volumes that favour the silicon seller, who is less exposed to which model captures the rent. The point is not that open source will destroy anyone's revenues. It is that the contracts fix claims twenty years into the future, while neither the industrial structure nor the distribution of rents is remotely fixed for twenty years.
What the market decided in three weeks
Then the market ran the experiment. From late July, Nvidia was discussing backstopping OpenAI's ten-gigawatt Ohio campus for up to 250 billion dollars; its shares fell five per cent when the figure surfaced. On 14 August the Wall Street Journal reported the support cut to less than 120 billion after investor pushback. On 17 August the deal was signed with Nvidia's guarantee capped at 105 billion, an "aggregate payment obligation" disclosed in Nvidia's 8-K, covering an initial 4.25 gigawatts with optional support for approximately 3.8 more, on a twenty-year lease. Over three weeks, the reported exposure fell 58 per cent from the figure initially discussed; the Wall Street Journal attributed the earlier reduction to investor pushback. Support for the additional capacity remains optional and, as of the filing, uncommitted. The documented conclusion is not that the market rejected anything; 105 billion of contingent exposure is enormous. It is that the structure survived, with less than half the guarantee exposure originally discussed, and that in this market the distance between what is discussed and what is signed ran to 145 billion dollars. The 105 billion itself is a maximum gross exposure, phased rather than present: the guarantees take effect lease by lease across nine construction phases, the first expected in Nvidia's fiscal 2029, and decline as payments are made.
And the filing rewards close reading, because the support does not cover chips at all: it is credit support for OpenAI's lease obligations on the physical campus itself: the data centre, power and transmission infrastructure. The machinery has already widened beyond the silicon, down the whole chain that makes the compute exist. The instrument itself, filed with the 10-Q as a Form of Residual Value Guaranty, is built around a defined Guaranteed Minimum Value covering data-centre, power and transmission costs, a contractually defined minimum against which the owner's covered loss is calculated. If the tenant defaults, the security deposit and reserve accounts absorb losses first; Nvidia then covers a defined shortfall against that minimum, and before paying it holds a menu of remedies: assume the lease itself, have the site re-let or sold, allow the lease to terminate, or defer for up to a year. Two clauses reveal the economics of the deal. In exchange for the guaranty, the site exclusively hosts Nvidia infrastructure; and the guaranty terminates early if OpenAI or its parent replaces it with rated bank credit support or itself achieves the rating the contract specifies. Economically, the structure is a credit bridge: Nvidia lends its balance sheet until the tenant's credit is strong enough to replace it. Follow the chain. The developer holds a claim on OpenAI; Nvidia supports part of its value; and if Nvidia has to pay, it can recover only from OpenAI and its parent, and only behind the landlord's remaining claims, a recovery claim likely to be worth least in the same payment-distress scenario that triggers the support. The risk was layered, in contract form. Nothing exits; a new claim is written.
Nor was Ohio the market's only experiment this summer. In April, a 4.6 billion dollar bond for a QTS data centre leased to Microsoft drew orders of nearly three times the deal on the strength of the tenant's triple-A rating; the debt, however, is not fully repaid before the initial lease ends, leaving the structure dependent on finding another tenant afterwards. By August, new QTS bonds were priced to yield more than 7.2 per cent, up from 5.7 in April: the market was now demanding a far larger premium for the structure's risk, despite the tenant's credit quality. The market is learning to read the documents, one repricing at a time.
I argued elsewhere that this capability operates in a munition economy: at once the engine of growth, the centre of gravity of the markets and a national security resource. This summer's financing adds a fourth attribute. The capability has acquired a financial constituency with claims on the continued performance of the build-out already contracted, a population of senior lenders, junior funds, insurers and guarantee counterparties who do not need to believe in artificial intelligence; they need the contracts honoured. Nvidia's own filings show the constituency forming around the vendor itself: 25 billion dollars of equity-investment commitments, subject to contingencies, in what it calls "AI model makers, infrastructure financiers" and other private companies, and revenue-sharing rights in the AI clouds it supplies. Vendor, investor, guarantor and revenue participant, in the same chain. Alphabet can cut its capex tomorrow. What cannot be cut is the stock of fixed claims already written, the irrevocable leases, the take-or-pay contracts, the purchase commitments, the guarantees contingent on default, non-renewal or value shortfalls. The FT estimates that the four largest hyperscalers alone have signed more than 1.5 trillion dollars of leases since the boom began, commitments not yet in effect included.
The distinction matters because of what happened either side of one weekend. On 7 August OpenAI activated its preparedness measures and paused internal Astra activities that did not yet meet the strengthened controls; later reporting described a two-week pause and a halt to Astra training. On 17 August the same company, through an affiliate, entered a twenty-year infrastructure lease supported by contractual claims that can survive any subsequent change of mind. A lab can slow its own model development because a safety threshold may have been crossed. What it cannot do unilaterally is unwrite the claims. Claims can be renegotiated, bought out, impaired or defaulted on, but each of those routes allocates a loss to someone, which is the difference that matters: before the claim is written, slowing down is principally an investment decision; after the claim is written, slowing down becomes an allocation of losses. Debt does not make the build-out unstoppable; it makes stopping it costly to someone. What this summer's financial engineering has done is distribute in advance the rights and obligations that will determine who bears that cost. Who ultimately bears it, and who gets to decide how it is distributed, is the governance question, and it is where this series goes next.
