Introduction: The Five-Year Promise in a One-Year Technology Race
On Monday, October 26, 2026, Space Exploration Technologies Corp. filed a Current Report on Form 8-K with the U.S. Securities and Exchange Commission that, on its surface, read like a routine piece of corporate housekeeping, and that, beneath its surface, described one of the most consequential financial commitments of the artificial intelligence era. Two days earlier, on October 24, the company had signed a binding commitment letter to enter into a master finance lease with an entity managed by Apollo Global Management. Under that contemplated lease, the Apollo-managed lessor would acquire graphics processing unit pods and related equipment, built around NVIDIA’s next-generation Vera Rubin platform, and lease them to SpaceX for use at data centers selected by the company. The equipment would arrive in tranches, each tranche governed by a separate lease schedule running approximately five to five and one-quarter years from its commencement date. Rent under each schedule would be payable quarterly in advance and would generally consist of one-twentieth of the financed cost of the applicable tranche, together with an amount calculated at a fixed rate per annum on the outstanding lease balance, plus certain ongoing costs and expenses, with the fixed rate to be determined at pricing before SpaceX actually entered into the master lease.[1] The company expected aggregate principal payments under the lease schedules, over their respective terms, to total approximately $40 billion, and it expected the master lease to close before the end of 2026, with the first lease schedules commencing during the first half of 2027 upon delivery of the GPUs.[1]
Apollo and Goldman Sachs led the committed financing, with Goldman Sachs serving as sole structuring agent, and affiliates of BlackRock, Goldman Sachs Asset Management, and KKR provided anchor financing commitments.[2] SpaceX described the arrangement as the largest committed GPU financing ever assembled, a claim that was true by a meaningful margin, since it exceeded the roughly $36 billion structured notes transaction that Apollo and Blackstone had finalized earlier in the year to purchase Google tensor processing units and lease them to Anthropic, a deal in which Broadcom had provided residual value support on the senior tranches.[2][71] Within forty-eight hours of the SpaceX announcement, Fitch Ratings indicated that the lease would leave SpaceX’s BBB+ investment-grade credit rating unchanged, a judgment that signaled how the rating agencies had come to understand these structures: not as a hidden mountain of debt, but as an asset-backed financing with its own amortization logic, its own collateral, and its own risk allocation.[4] The commitment was binding on both parties, but the execution of the definitive lease remained subject to specified closing conditions, and the 8-K cautioned, in the formulaic language that securities counsel insist upon, that the transaction might not close on the anticipated terms or timeline, or at all, and that actual equipment deliveries, lease terms, or aggregate principal payments might differ from current expectations.[1]
The deeper story of this transaction lies not in its size, remarkable as that size is, but in the relationship between the duration of the financial contract and the velocity of the underlying technology. A lease schedule of five to five and one-quarter years, commencing in the first half of 2027, runs into mid-2032. During that same span of time, if NVIDIA holds to the annual product cadence it has publicly committed to, the semiconductor industry will have introduced Rubin Ultra, which NVIDIA expects to begin delivering in the second half of 2027, and at least two further architectures beyond it.[27] AMD will have shipped the MI450 series inside its Helios rack-scale architecture, the custom MI450-based accelerator it is co-engineering with Meta, and the MI500 series that its collaboration with Meta is already informing.[28] Google, Amazon, Microsoft, and Meta will have iterated through multiple generations of their own custom accelerators. Software developers will have found new ways to compress models, distill them, quantize them, and route them so that a given quantity of useful intelligence requires fewer floating-point operations than it did the year before. Hyperscalers will have shifted the economics of training and inference in ways that cannot be fully anticipated in November 2026. A Vera Rubin accelerator that is extraordinarily valuable at its delivery in 2027 may remain fully productive in 2030, and may still be generating revenue in 2032, while commanding a substantially different market valuation than the one against which it was financed.
That mismatch, between a five-year financial promise and a one-year technology race, is the subject of this paper, and it is the reason I have chosen the title “Collateral Half-Life.” The phrase joins two disciplines that are becoming inseparable in the artificial intelligence economy. Collateral is the language of finance: it denotes the equipment and the associated recovery rights that support a financial obligation, the thing a lender can take and sell if the borrower stops paying. Half-life is the language of physics: it denotes the time required for a measured quantity to decay to half of its initial value, a metaphor that has migrated from radiochemistry into pharmacology, into information science, and now, I would argue, into asset-backed finance. I define collateral half-life, for the purposes of this paper, as the time required for the independently estimated net recoverable market value of a specific GPU system to decline to half of its initial financing-date value. It is an analytical metric that I propose, not an established industry standard, and I propose it precisely because the industry does not yet have one. The useful-life assumptions that hyperscalers disclose in their financial statements, the amortization schedules that lessors embed in their contracts, and the resale quotations that circulate in secondary markets all measure different things, and none of them measures the single quantity that a lender most needs to know, which is how quickly the thing it could repossess is losing the value it was underwritten at.
The title is appropriate for a further reason. Technological, financial, and economic timelines do not necessarily align, and the direction in which they diverge is not predetermined. A GPU may remain technically operational for many years after newer architectures have made it less competitive for frontier training workloads, which is the pessimistic reading, in which the asset decays faster than the loan amortizes. But the same GPU may be rescued by forces that the pessimistic reading ignores: the explosion of inference demand, the proliferation of smaller and more specialized models, the arrival of robotics and autonomous agents that need large quantities of modest compute rather than small quantities of frontier compute, and the simple fact that a datacenter already energized, cooled, and networked is a scarce asset in a world where the grid, not the chip, has become the binding constraint. In October 2026, median H100 rental rates had rebounded to between $3.30 and $3.60 per GPU-hour, roughly 30 percent above their 2025 lows, for a chip that NVIDIA first shipped in 2022 and that two full architectural generations have since superseded.[52] Older systems, in other words, can sustain value longer than the obsolescence narrative predicts. The paper therefore examines how these competing forces jointly determine collateral recovery, debt-service capacity, institutional risk, and the financial sustainability of what I call the Five-Layer AI Economy: energy, chips, datacenters, models, and applications.
A word on method and on the evidentiary base. This essay draws on the primary documents of the transactions it discusses, principally SEC filings and issuer press releases; on the financial stability assessments published in 2026 by the Federal Reserve, the International Monetary Fund, the European Central Bank, and the Financial Stability Board; on the public disclosures of CoreWeave, SpaceX, NVIDIA, AMD, Meta, and other companies through their most recent reporting periods; and on the published commentary of academic economists and finance scholars at New York University, the Massachusetts Institute of Technology, Stanford, Harvard, and Princeton, among others, whose views I quote where their words are more precise than my paraphrase. Where numbers are illustrative rather than observed, I say so. Where the evidence is contested, I present the contest rather than resolving it by assertion. The paper proceeds in six sections. Section 1 traces how GPUs became financeable assets and anatomizes the SpaceX–Apollo structure. Section 2 develops the concept of collateral half-life and situates it against four competing measures of equipment life. Section 3 turns to underwriting, lease design, and the allocation of risk. Section 4 examines the institutional market that has formed around AI collateral and the financial-system questions it raises. Section 5 looks forward to 2027–2030 and proposes a Collateral Half-Life Dashboard. Section 6 distills seven pillars of what we have learned, and the Conclusion returns to the question with which we began: what it means to write a financial contract against a machine whose frontier it will outlive.
Section 1: The Financialization of Artificial Intelligence GPUs
1.1 From semiconductor purchases to institutional financing assets
For most of the history of the computing industry, a server was a corporate expense that happened to be capitalized. A company bought it, placed it on the balance sheet as property and equipment, depreciated it over three or four years, and replaced it when it broke or when the software it ran demanded something faster. Banks did not lend against servers in any meaningful way, because servers were not scarce, were not standardized in a way that supported resale, and lost value too quickly to serve as credible collateral. The capital markets financed the companies that owned the servers, through corporate bonds and equity, and the servers themselves were an afterthought in the credit analysis, a line item in the capital expenditure forecast rather than a security interest in a loan agreement.
The artificial intelligence boom has inverted that relationship, and it has done so with a speed that would have seemed implausible as recently as 2022. Three conditions converged to make the GPU a financeable asset in its own right. The first was scarcity. From late 2022 through 2024, demand for NVIDIA’s H100 so dramatically exceeded supply that the chips traded at premiums to list price, carried allocation waiting lists, and generated rental rates as high as eight dollars per GPU-hour, levels at which the hardware could pay for itself in well under two years of operation.[53] Scarcity gave the asset a visible market price, and a visible market price is the first precondition of collateral. The second condition was standardization. Unlike the heterogeneous servers of the prior era, AI accelerators arrived in a small number of configurations, dominated by one vendor, deployed in rack-scale systems whose specifications were public, and ran a software stack, CUDA, that made a GPU in one datacenter functionally interchangeable with the same GPU in another. Standardization is the second precondition of collateral, because it is what allows a lender to believe that the thing it repossesses can be sold to someone else. The third condition was the sheer size of the capital requirement. When hyperscalers alone are projected to deploy more than $650 billion of capital expenditure in 2026, and when Goldman Sachs estimates roughly $7.6 trillion of capital across compute, datacenters, and power over the coming years, no corporate balance sheet, however fortress-like, can absorb the burden through retained earnings and unsecured bonds alone.[22] The demand for capital created the supply of financial engineering.
What emerged from these conditions was a new category of asset-based finance in which the GPU, or more precisely the installed GPU system, became the primary security for an obligation, and in which the credit analysis shifted, at least in part, from the borrower’s general creditworthiness to the asset’s expected cash flows and recoverable value. The shift is partial rather than total, and that partiality matters. In the SpaceX–Apollo structure, as in the CoreWeave delayed-draw term loans that preceded it, the lessee or borrower remains a party of substance, and its contracted revenue, its cash reserves, and its rating all inform the pricing. But the architecture of the financing is built around the equipment: the lessor owns it, the lease schedules are tied to its delivery, the amortization is tied to its financed cost, and the lender’s recovery in a default scenario depends, in the first instance, on what the equipment is worth. That is the structural meaning of financialization. The chip has become a capital-market instrument.
1.2 The SpaceX–Apollo financing architecture
To understand what the SpaceX–Apollo transaction is, it helps to be precise about what it is not, because the headline figure invites a number of misreadings. It is not a $40 billion loan, in the sense of a single disbursement of cash to SpaceX. It is not a completed transaction, in the sense that equipment has changed hands or rent has been paid. And it is not a measure of the total financing cost that SpaceX will bear, in the sense of an all-in figure inclusive of interest. Each of these distinctions bears on how the deal should be analyzed, and each is drawn explicitly in the 8-K.
The first distinction is between equipment ownership and lease-use rights. Under the contemplated master finance lease, the Apollo-managed entity, which the filing calls the Lessor, acquires the GPU pods and related equipment and holds title to them. SpaceX obtains the right to use that equipment at datacenters of its choosing, and it pays rent for that right. At the level of legal form, this is an equipment lease of the kind that has financed aircraft, railcars, and shipping containers for decades; at the level of economic substance, because the rent is structured to return the full financed cost over the term plus a fixed return on the outstanding balance, it functions as a secured financing in which the lessor’s ownership of the equipment is the collateral. The filing describes the arrangement as a master finance lease and does not discuss how SpaceX will account for it, but the “finance lease” characterization implies that, under ASC 842, SpaceX will record a right-of-use asset and a corresponding lease liability on its balance sheet as each schedule commences, which is why Fitch could evaluate the arrangement as a form of leverage while still concluding that it left the rating unchanged.[3]
The second distinction is between expected principal payments and total financing cost. The 8-K states that the company expects aggregate principal payments under the lease schedules, over their respective terms, to total approximately $40 billion.[1] That figure is the sum of the one-twentieth-of-financed-cost components of the quarterly rent across all tranches, which is to say that it is the aggregate financed cost of the equipment, repaid over twenty quarterly installments per tranche. It excludes the fixed-rate component, which is the lessor’s return and which has not yet been priced, and it excludes the ongoing costs and expenses that the lease passes through. The all-in cost of the financing therefore depends on a rate that will be set shortly before closing, in a market in which, as reporting in the final week of October 2026 indicated, the recent turmoil in debt markets for AI-related companies had pushed borrowing costs higher than they had been when the deal was first discussed in early October.[8] If one assumes, purely for illustration, a fixed rate in the range of six to eight percent on a declining balance that averages roughly half of the financed cost over a five-year term, the interest component alone would add on the order of $6 billion to $8 billion to the $40 billion of principal, and the arithmetic of quarterly rent for a fully drawn facility would approach $2 billion of principal per quarter plus interest.[72] These are illustrative figures, not disclosures, but they convey the scale of the cash-flow commitment.
The third distinction is between binding commitment and completed transaction. What SpaceX signed on October 24 was a commitment letter, which obligates the parties to enter into the master finance lease subject to specified closing conditions. The definitive documentation had not been executed as of the filing, the fixed rate had not been priced, and no equipment had been delivered. The lease schedules commence only upon delivery of the GPUs, expected in the first half of 2027, which means that SpaceX does not receive $40 billion of equipment on day one and does not begin paying rent on $40 billion on day one. Instead, the facility draws down as tranches arrive, and each tranche’s five-year clock starts when it is delivered. This staging is one of the structure’s principal virtues from the lessee’s standpoint, because it aligns the financing obligation with the arrival of the revenue-generating asset, and it is one of the reasons the structure is preferable, for SpaceX, to another large unsecured bond offering so soon after the $25 billion bond issue the company completed in June 2026 around the time of its initial public offering.[8]
The fourth distinction, and the one that organizes the rest of this paper, is between GPU collateral and the borrower’s broader creditworthiness. SpaceX is, by any measure, a strong lessee. At September 30, 2026, it reported approximately $51.3 billion in cash and cash equivalents and $38.6 billion in marketable securities, roughly $89.9 billion combined, against $38.8 billion of debt and finance lease liabilities.[9] Its third-quarter 2026 revenue of $14.6 billion was up 161 percent year over year, its net income was $4.4 billion, and its Super Intelligence segment, which includes the former xAI and the compute-leasing business, generated $7.56 billion of revenue, more than the Starlink-led connectivity segment for the first time.[10] A lessor financing equipment for such a counterparty is, in the ordinary course, relying on the lessee’s ability to pay rent from operating cash flow, not on repossessing and reselling GPUs. But the structure is nonetheless built so that the equipment stands behind the obligation, and the terms on which institutional investors committed $40 billion were shaped, in part, by what they believed that equipment would be worth if they ever had to rely on it. The question of collateral value therefore matters even when the lessee is strong, because it determines the price of the financing, the advance rate, the reserves, and the covenants, and because the lessee’s strength in 2026 is not a guarantee of its strength in 2031.
1.3 The evolution of specialized compute financing
The October 2026 transaction did not emerge fully formed. It is the latest and largest iteration of a template that Apollo and a small number of other alternative asset managers have been refining since at least 2024, and the January 2026 financing for Valor Compute Infrastructure and xAI is the most instructive precedent, because it involved the same manager, substantially the same ultimate lessee, and the same vendor.
On January 7, 2026, Apollo announced that Apollo-managed funds and affiliates had led a $3.5 billion capital solution for Valor Compute Infrastructure L.P., a fund managed by Valor Equity Partners, to support its $5.4 billion acquisition and lease of datacenter compute infrastructure, including NVIDIA GB200 GPUs, to a subsidiary of xAI Corp.[5] The transaction was structured as a triple-net GPU lease financing to support xAI’s second datacenter, and it closed one day after xAI announced a $20 billion Series E round in which Valor, NVIDIA, Fidelity, the Qatar Investment Authority, and Abu Dhabi’s MGX participated.[5] Apollo’s partner Christopher Lahoud characterized the deal in language that is worth preserving, because it captures the institutional investor’s self-understanding of these structures.
“This transaction represents a hallmark, downside-protected investment for Apollo”
— Christopher Lahoud, Partner, Apollo Global Management [73]
Valor’s founder and chief executive Antonio Gracias described the vehicle from the equity side, noting that it offered investors exposure to compute infrastructure with recurring distributions and an equity claim on the hardware.
“quarterly cash distributions and upside through ownership of the compute assets”
— Antonio Gracias, Founder and CEO, Valor Equity Partners [5]
Three features of the January transaction carried forward into October, and one did not. What carried forward was the basic tripartite structure, in which an asset-owning vehicle acquires the equipment, a senior capital provider finances the bulk of the purchase, and the operating company leases the equipment under a long-term contract; the vendor’s participation in the ecosystem, with NVIDIA an anchor limited partner in the Valor fund and the supplier of the equipment in both cases; and the explicit framing of the investment as asset-based and downside-protected rather than as corporate credit. What did not carry forward was the intermediary. In January, Valor’s fund stood between Apollo and xAI as the equipment owner and lessor; in October, the Apollo-managed entity itself is the lessor and SpaceX, which absorbed xAI in a February 2026 merger, is the direct lessee.[11] The elimination of the intermediate layer reflects both the maturation of the template and the scale of the October deal, which at roughly seven times the January financing exceeded the sum of all of SpaceX’s existing debt and lease liabilities as of the end of September in a single commitment.[72]
Between January and October, the template also absorbed a third variant, the Anthropic TPU financing, which added a feature that neither the Valor deal nor, so far as the public record shows, the SpaceX deal contains: explicit vendor residual value support. In that transaction, Apollo and Blackstone structured roughly $36 billion of debt, reported as approximately $6 billion of A1 notes, $25 billion of A2 notes, and $4.5 billion of B notes, to be issued by a special purpose vehicle that would purchase Google tensor processing units and lease them to Anthropic for deployment at sites in New York, Texas, Louisiana, and Indiana.[6][75] Broadcom, which helps Google develop the TPUs, provided a residual value support agreement on the A1 and A2 tranches, under which, if Anthropic defaulted on its lease obligations and a chip resale failed to cover what was owed, Broadcom would make the senior noteholders whole.[7] The practical effect was to align the senior tranches’ risk profile with investment-grade exposure, and the structural lesson was that the industry had found a way to transfer the collateral half-life risk, the risk that the chips would be worth less than the outstanding debt at the moment of default, from the lenders to a party with a commercial interest in the chips’ continued relevance. Whether and how the SpaceX–Apollo documentation allocates that risk is not disclosed in the 8-K, and the absence of a disclosed vendor backstop is one of the open questions that Section 3 takes up.
1.4 The emerging institutional ecosystem
The participants in the SpaceX–Apollo financing form a cross-section of the institutional ecosystem that has grown up around AI collateral, and each brings a distinct interest to the table.
Apollo’s interest is the most straightforward and the most strategic. The firm has deployed more than $40 billion into next-generation infrastructure since 2022, including compute capacity, digital platforms, and renewable energy, and it has positioned asset-based finance as a core growth engine for its credit business.[5] Its insurance affiliate, Athene, with roughly $785 billion of assets under management, provides a perpetual pool of long-dated liabilities that must be matched with long-dated, investment-grade-like assets, and a five-year amortizing lease to an investment-grade lessee secured by equipment is, from an insurer’s asset-liability perspective, a nearly ideal instrument.[74] Apollo has also been explicit about its ambition to create secondary trading markets for these exposures, which would allow it to originate at scale, retain the portions it wants, and distribute the rest.[67] For Apollo, the GPU lease is not a one-off; it is a product line.
Goldman Sachs’ interest, as sole structuring agent and co-lead, is in fees and in franchise. The investment bank that structures the largest GPU financing in history establishes itself as the house to call for the next one, and the structuring expertise developed in aircraft and equipment finance translates directly. Goldman Sachs Asset Management’s participation as an anchor investor reflects the same asset-liability logic that motivates Athene: insurance and pension capital seeking yield above comparable corporate bonds with the protection of a hard asset.
BlackRock’s and KKR’s interests are those of large diversified managers with substantial infrastructure and private credit platforms, for whom AI compute is both an investable theme and, increasingly, a client demand. Public pension commitments to private markets jumped 51 percent in a single quarter in early 2026, much of it directed toward infrastructure and private credit strategies that market AI exposure as a differentiator.[22] For these managers, participation in a landmark deal is as much about demonstrating capability to limited partners as it is about the returns on the specific tranche.
The commercial banks, though not named as anchors in the SpaceX deal, are present throughout the ecosystem as providers of warehouse lines, delayed-draw term loans, and the asset-backed securities that increasingly refinance equipment leases. The IMF reported that U.S. bank loans to nonbank financial intermediaries rose 18 percent in 2026 and that bank exposures to private equity and private credit firms rose 30 percent, a reminder that the private credit that finances GPUs is, several layers down, often bank credit.[19]
The chip manufacturers’ interests are the most complex. NVIDIA benefits from a $40 billion committed deployment channel for hardware that has not yet shipped at scale, which provides demand visibility and supports its own guidance; it has also, through its investments in customers and its anchor positions in financing vehicles, become a participant in the capital structure of its own demand. The same week that SpaceX announced the Apollo lease, Meta launched a Class A equity offering in which NVIDIA indicated interest in purchasing up to $3 billion of shares and AMD up to $1 billion, as both vendors expanded their deployments at Meta.[78] This circularity, in which vendors finance the customers who buy their products, is a cushion in good times and a potential channel of contagion in bad ones, and Section 6 returns to it.
Finally, the AI infrastructure operators, of which SpaceX is now one of the largest, have an interest in obtaining compute at the lowest cost of capital and with the greatest flexibility, and the lease structure serves both. It allows the operator to add capacity in step with contracted demand, to keep its unsecured borrowing capacity in reserve, and to match the financing term to a period over which it can reasonably forecast revenue. What it does not do is eliminate the obsolescence risk; it merely allocates that risk between lessee and lessor in a manner that the documentation specifies and the public filings, so far, do not.
1.5 The Five-Layer AI Economy
Throughout this paper I will rely on a framework I call the Five-Layer AI Economy, because it clarifies where the GPU sits in the value chain and why its financing depends on activity far removed from the datacenter floor. Layer 1 is energy: the generation, transmission, and distribution capacity that supplies electricity to compute at the scale and reliability that AI workloads demand. Layer 2 is chips: the accelerators, CPUs, memory, and networking silicon that perform the computation. Layer 3 is datacenters: the buildings, cooling systems, power distribution, and physical infrastructure that house and energize the chips. Layer 4 is models: the trained artificial intelligence systems, whether frontier foundation models or specialized enterprise systems, that transform compute into useful intelligence. Layer 5 is applications: the products and services, from coding assistants to customer service agents to autonomous vehicles, that convert intelligence into revenue that customers are willing to pay.
The SpaceX–Apollo transaction finances Layer 2 assets, the Vera Rubin GPU pods, for deployment inside Layer 3 facilities, the datacenters SpaceX selects. But the rent that services the lease is paid from SpaceX’s operating cash flow, which is generated at Layers 4 and 5: from Grok, from the Cursor coding tool SpaceX acquired in August 2026, and above all from the compute-leasing business in which SpaceX rents capacity to Anthropic, Google, Reflection AI, and others, who in turn generate revenue from their own models and applications.[9][77] The chain is long. A Vera Rubin GPU in Memphis earns its keep only if a customer of a customer of SpaceX finds enough value in an AI application to pay for the inference that runs on it. And the chain depends on Layer 1 at every step, because, as Elon Musk told investors in October 2026, power, cooling, and infrastructure had become greater constraints on deploying AI than chips.[9] The framework’s central implication for collateral analysis is that the recoverable value of a Layer 2 asset is not an intrinsic property of the silicon; it is a derived property of the health of all five layers, and a lender who underwrites the chip without underwriting the economy around it has underwritten a fraction of the risk.
Section 2: Technological Depreciation and the Half-Life of Compute
2.1 Four different measures of equipment life
The single most consequential confusion in the public debate over AI infrastructure finance is the conflation of several distinct measures of how long a GPU “lasts.” The word “obsolete” is deployed as though it named one event, when in fact it names at least four, and the four can be separated by years. Because the financing of a GPU system is a contract written against time, getting the measure of time right is not a pedantic exercise; it is the whole of the credit analysis. I distinguish the following.
Technical life is the period during which the GPU continues to function within specification: the silicon powers on, the memory holds, the interconnects pass data at rated speeds, and the failure rate remains within the bounds that an operator can tolerate. Survival studies and hyperscaler retention data point to a physical useful life of six years or more for datacenter accelerators under ordinary operating conditions, and the fleets of NVIDIA A100s from 2020 that are still generating rental revenue in 2026 are empirical proof that the silicon does not expire on the schedule the obsolescence narrative implies.[40] Technical life is the longest of the four measures, and it is the least relevant to a lender, because a GPU that functions perfectly but that no customer wishes to rent is, from a cash-flow standpoint, indistinguishable from one that has failed.
Economic life is the period during which the GPU’s use remains commercially worthwhile for its owner: the rental revenue or the internal value of the compute it produces exceeds the marginal cost of operating it, principally electricity, cooling, maintenance, and the opportunity cost of the datacenter space and power it occupies. Economic life is shorter than technical life and is highly workload-dependent. For frontier training, where the newest architecture’s advantage in performance per watt is decisive and where a lab competing to train the largest model cannot afford to use last year’s chips, the frontier-competitive life of an accelerator may be as short as one to two years, as a Princeton Center for Information Technology Policy analysis has noted.[40] For inference, fine-tuning, batch processing, and enterprise workloads, where the relevant metric is cost per token at acceptable latency rather than time to train the largest model, economic life extends considerably, and the “computing cascade” by which hyperscalers redeploy GPUs from training to inference as they age is the industry’s principal defense of its longer depreciation schedules.[37]
Accounting life is the period over which the owner depreciates the asset’s cost in its financial statements. It is an estimate chosen by management within the bounds of generally accepted accounting principles, which prescribe no specific number of years, and it has become the most visible and most contested of the four measures because it directly determines reported earnings. The trajectory of hyperscaler accounting lives over the past decade tells a story in itself: Amazon moved its server depreciation from three years to four in 2020 and to six years by 2023, Alphabet and Microsoft converged on six, and by 2024 the largest cloud operators had normalized on a five-to-six-year schedule for servers and networking equipment, including AI accelerators.[39] Then, in January 2025, the consensus fractured. Meta extended the useful life of most of its servers and network assets to five and a half years, up from a prior range of four to five, a change that reduced its 2025 depreciation expense by approximately $2.9 billion; in the same month, Amazon shortened the useful life of a subset of its servers and networking equipment from six years to five, citing the accelerating pace of AI hardware development, and recorded additional 2025 depreciation of $1.4 billion as a result.[33][34] Two companies, operating substantially the same machines, reached opposite conclusions on the same day, and the Harvard Business School case study on Meta’s decision, published in June 2026, has made that divergence a standard teaching example in the accounting of technological obsolescence.[32] CoreWeave, for its part, disclosed a six-year estimated useful life for its technology equipment in its 2025 annual filing, an assumption at the long end of the industry range for a company whose business is entirely built on the rental value of that equipment.[13]
Collateral recovery life is the measure this paper is concerned with, and it is the one that none of the others captures. It is the period during which a lender or lessor, having taken possession of the equipment following a default, could realize enough value from it, through sale, re-lease, or redeployment, to cover a given fraction of the outstanding obligation. Collateral recovery life is shorter than accounting life whenever the market value of the equipment declines faster than the straight-line depreciation schedule assumes, which in a market with annual architectural generations is nearly always. It is shorter than economic life because the lender’s recovery depends not on the equipment’s value to its current operator but on its value to a third party who must bear the costs of removal, transport, reinstallation, and software integration, and who must find an energized facility in which to deploy it. And it can be shorter than technical life by a very wide margin, because a GPU that functions perfectly may have no buyer at any price that matters if the secondary market is saturated with identical units from other distressed sellers at the same moment, which is precisely the correlated scenario that Section 3.5 examines.
These four measures can diverge substantially, and the divergence is not symmetrical. CoreWeave’s six-year accounting life does not guarantee a six-year collateral recovery life, and in a stress scenario the gap between the two is the lender’s loss. The industry’s defenders have made the reasonable point that investors frequently confuse accounting depreciation schedules with physical end-of-life, and that silicon does not stop functioning at year six.[40] But the industry’s critics have made the equally reasonable point that the question is not whether a GPU powers on in year six; it is whether straight-line depreciation over five to six years matches the period during which the chip earns an attractive return and retains a resale value commensurate with its book value.[35][76] Both points are correct, and the resolution is not to pick one but to measure each life separately, which is what the collateral half-life framework attempts.
2.2 NVIDIA’s accelerating innovation cycle
The reason the four lives have diverged so sharply is that NVIDIA has compressed its architectural cadence from roughly two years to one, and each generation has delivered improvements large enough to reprice the generation before it. The sequence from Hopper to Blackwell to Vera Rubin is the empirical core of the obsolescence debate, and it deserves to be traced with some care.
Hopper, embodied in the H100 that began shipping in 2022 and the H200 that followed in 2024, was the architecture that made the generative AI boom possible and that established the GPU as a financeable asset. Its scarcity in 2023 produced rental rates above eight dollars per GPU-hour and secondary prices above $40,000 per unit; its abundance in 2025, as Blackwell supply ramped, produced rental rates below two dollars on some marketplaces and secondary prices that fell, by some accounts, to between $6,000 and $22,000 for used units by mid-2026, before the inference-driven rebound of the autumn lifted both rentals and resale.[50][51] The Hopper price history is the only complete collateral half-life observation the industry possesses, and it is sobering: from peak scarcity pricing in mid-2024 to the trough in early 2026, the secondary value of a used H100 fell by roughly 70 to 75 percent in about eighteen months.[83] It is also encouraging, in a way the pessimists sometimes neglect: the trough was followed by a recovery in rental rates of roughly 30 percent as inference demand absorbed the installed base, and median on-demand H100 pricing in mid-October 2026 stood at $3.49 per GPU-hour.[52]
Blackwell, embodied in the B200 and the GB200 NVL72 rack-scale system that began shipping in volume in 2025, delivered roughly 2.5 times the AI performance of Hopper and dramatically better energy efficiency per token, at a per-GPU acquisition cost roughly 1.6 to 1.7 times that of the H100.[84] The economics of that trade-off, more performance per dollar of capital and far more performance per watt, is what demoted Hopper to a value tier and what compressed its rental and resale pricing. Blackwell Ultra, the B300, followed in 2026.
Vera Rubin is the architecture that the SpaceX–Apollo lease finances, and its specifications, as NVIDIA has stated them, are the reason this paper’s question is urgent rather than academic. NVIDIA announced at CES in January 2026 that Vera Rubin was in full production and would be available through partners in the second half of 2026.[24] The Rubin GPU, manufactured on TSMC’s 3-nanometer process, integrates 336 billion transistors and 288 gigabytes of HBM4 memory at 22 terabytes per second of bandwidth; NVIDIA claims 50 petaflops of NVFP4 inference performance per GPU, five times Blackwell, and 35 petaflops of training performance, 3.5 times Blackwell, achieved with only a 1.6-fold increase in transistor count.[27] The Vera Rubin NVL72 rack pairs 72 Rubin GPUs with 36 Vera CPUs and delivers 3.6 exaflops of NVFP4 inference compute; NVIDIA claims the platform reduces inference token costs by a factor of ten relative to Blackwell and cuts the number of GPUs required to train large mixture-of-experts models by 75 percent.[26] The company confirmed volume production at Computex in June 2026 with more than 350 partners in 30 countries.[85]
Two cautions attach to these figures. The first is that company benchmarks require independent evaluation across workloads, and the gap between NVIDIA’s rack-level claims and the performance an operator realizes on a specific model at a specific precision can be large. The decoder’s observation that “full production” is itself an ambiguous phrase for a product of this complexity is well taken.[25] The second caution cuts the other way and is more important for collateral analysis. If Vera Rubin delivers even a substantial fraction of a tenfold reduction in cost per inference token relative to Blackwell, then the Blackwell systems that were financed in 2025 on five-year schedules face a repricing in 2027 of the same character that Hopper faced in 2025, and the Vera Rubin systems that SpaceX will lease in 2027 face the same repricing when Rubin Ultra, which NVIDIA expects to deliver in the second half of 2027 with 576 GPUs per rack and roughly four times the inference performance of the Rubin NVL144, arrives.[27] Microsoft’s chief executive acknowledged the problem with unusual candor when explaining why the company had chosen not to over-commit to a single hardware generation.
“stuck with four or five years of depreciation on one generation”
— Satya Nadella, Chief Executive Officer, Microsoft [37]
The same company’s chief financial officer offered the counterargument in equally plain terms, noting that the capacity being depreciated over six years was contracted out for the full span.
“already sold for the entirety of their useful life”
— Amy Hood, Chief Financial Officer, Microsoft [38]
Both statements are true, and together they define the terrain. Technological obsolescence arrives on an annual cadence; whether it produces financial loss depends on whether the capacity was sold before the obsolescence arrived.
2.3 Competition from AMD and custom accelerators
NVIDIA’s cadence would be destabilizing enough on its own; the arrival of credible competition multiplies the repricing pressure, because it threatens not only the value of a given NVIDIA generation relative to the next NVIDIA generation but the premium that the NVIDIA ecosystem as a whole commands in the secondary market.
AMD’s MI400 family and its Helios rack-scale architecture represent the most direct challenge. On February 24, 2026, AMD and Meta announced a multi-year, multi-generation agreement to deploy up to six gigawatts of AMD Instinct GPUs, with the first gigawatt powered by a custom MI450-based accelerator co-engineered with Meta, paired with sixth-generation EPYC “Venice” CPUs and running ROCm software inside the Helios rack, which AMD and Meta developed together through the Open Compute Project.[31] AMD’s chair and chief executive Lisa Su described the custom part as the company’s first custom GPU effort of this style and indicated that production shipments of both the standard MI450 series and the Meta custom variant would begin in the second half of 2026.[30] On October 27, 2026, the companies expanded the partnership further: AMD announced that Helios racks were installed and running at Meta, that the collaboration was informing the MI500 series and next-generation rack-scale systems, and that AMD would invest $1 billion in Meta’s equity offering.[29]
“With Helios now running at Meta, we are taking that partnership to a new scale”
— Lisa Su, Chair and Chief Executive Officer, AMD [78]
The AMD–Meta collaboration illustrates a structural shift with direct consequences for collateral value: hardware design is increasingly following specialized customer workloads rather than the reverse. A custom MI450 optimized for Meta’s ranking, recommendation, and inference workloads is, by construction, less fungible than a standard part, and fungibility is the foundation of collateral value. The same logic applies with greater force to the fully custom accelerators: Google’s TPUs, which Anthropic is leasing at a scale of $36 billion; Amazon’s Trainium; Microsoft’s Maia; and Meta’s own MTIA. Each of these is designed for a single operator’s workloads, runs a software stack that operator controls, and has, in most cases, no secondary market at all. The Broadcom residual value backstop in the Anthropic transaction exists precisely because a TPU, unlike an H100, cannot be sold on a spot marketplace; the vendor’s guarantee substitutes for the market that does not exist.
For NVIDIA-based collateral, the competitive dynamic cuts in two directions. On one hand, every gigawatt of hyperscaler demand that migrates to AMD or to custom silicon is a gigawatt of demand that will not bid for used NVIDIA systems in the secondary market, which lowers recovery values. On the other hand, the proliferation of architectures fragments the software ecosystem and raises the switching costs of moving workloads off CUDA, which supports the installed base’s value. The net effect, in my judgment, is that NVIDIA systems will retain a liquidity premium in secondary markets for the balance of the decade, but that the premium will be narrower than it was in the Hopper era, and that lenders should not assume the Hopper rebound of 2026 will repeat for Vera Rubin in 2030.
2.4 Measuring collateral half-life
I now define the metric formally. Let V(0) denote the initial net recoverable market value of a specific GPU system at the financing date, where “net recoverable” means the price at which a lender or lessor could realistically sell or re-lease the system to a third party, less the costs of removal, transport, reinstallation, and remarketing. Let V(t) denote the same quantity at time t after the financing date. The collateral half-life, t½, is the first point at which the measured value has declined by 50 percent:
V(t½) = V(0) / 2
The metric is deliberately simple, and its simplicity conceals a number of choices that a serious analyst must make explicit. The first is the definition of V(0). It is not the invoice price of the equipment. A lender who records V(0) as the financed cost has already overstated the collateral, because the financed cost includes installation, integration, and margin that no secondary buyer will pay for. In the Hopper experience, the gap between new-unit list price and the price at which a two-year-old unit changed hands was 25 to 40 percent even before the Blackwell repricing, and the gap between the price of a loose GPU and the recoverable value of a complete, installed rack is larger still.[84] A realistic V(0) is therefore some fraction of financed cost, and the choice of that fraction is the first and most consequential underwriting decision.
The second choice is the shape of the decay function. The physical analogy of exponential decay, in which the asset loses a constant fraction of its remaining value each period, is a convenient starting point, and it yields the familiar relationship between half-life and the annual decay rate: a half-life of two years implies a decay rate of roughly 29 percent per year; a half-life of three years implies roughly 21 percent; a half-life of four years implies roughly 16 percent. But GPU values do not decay smoothly. They decay in steps, at the arrival of each new architectural generation, with plateaus between steps during which inference demand can stabilize or even lift values, as the H100 experience of 2026 demonstrated. A more realistic model treats the decay as a staircase whose risers are the generational transitions and whose treads are the intervening demand cycles, and it recognizes that the height of each riser depends on the magnitude of the performance-per-watt improvement the new generation delivers. NVIDIA’s claim of a tenfold reduction in inference token cost for Vera Rubin relative to Blackwell, if realized, implies a riser for Blackwell systems in 2027 that is considerably steeper than the riser Hopper experienced in 2025.
The third choice is the set of factors that determine where a particular system falls within the range. The eventual manuscript of a full study would develop scenarios around at least the following six variables. Removal and redeployment costs, which rise sharply for liquid-cooled rack-scale systems such as the NVL72 and its successors, whose value is in the integrated rack rather than in the individual boards. Software compatibility, which determines whether a secondary buyer can run its workloads on the system without a costly migration, and which favors CUDA-based systems over custom silicon. Electricity efficiency, which determines whether an older system can compete on cost per token when power is priced at the margin, and which is the variable through which Layer 1 of the Five-Layer AI Economy reaches directly into collateral value. Liquidity, which depends on the depth of the secondary market at the moment of sale and which collapses precisely when many holders try to sell at once. Equipment generation, in the sense of how many architectural steps separate the system from the current frontier at the moment of valuation. And demand conditions, which encompass everything from the growth of inference workloads to the arrival of robotics to the state of the macroeconomy.
The table below sketches three illustrative collateral half-life scenarios for a hypothetical Vera Rubin NVL72 system financed at the start of 2027. The figures are analytical constructions intended to show how the variables interact, not forecasts of SpaceX equipment values or of any specific contract’s recovery.
| Scenario | Assumed V(0) as % of financed cost | Annual value decay | Implied collateral half-life | Principal outstanding at t½ (20-quarter straight-line) | Recovery coverage at t½ |
| Adaptive expansion | 75% | ~16% | ~4.0 years | 20% of financed cost | ~1.9× outstanding principal |
| Orderly repricing | 70% | ~21% | ~3.0 years | 40% of financed cost | ~0.9× outstanding principal |
| Correlated impairment | 65% | ~29% | ~2.0 years | 60% of financed cost | ~0.5× outstanding principal |
Table 1. Illustrative collateral half-life scenarios for a hypothetical rack-scale GPU system under a five-year, twenty-payment straight-line amortization. Hypothetical values for analytical illustration only; not forecasts of SpaceX equipment prices, contract protections, or financing outcomes.
The table makes visible the relationship that Section 3 develops at length: whether a GPU lease is adequately collateralized depends not on the half-life alone, nor on the amortization schedule alone, but on the race between them. A half-life of four years comfortably outruns a five-year straight-line amortization; a half-life of three years roughly matches it at the midpoint; a half-life of two years loses to it decisively. The entire credit question reduces to which of these three regimes the actual market will occupy between 2027 and 2032.
2.5 The second economic life of GPUs
Everything in the collateral half-life framework depends on the proposition that a GPU displaced from the frontier can find a second economic life that sustains its value, and that proposition deserves to be examined on its merits rather than assumed in either direction.
The case for a robust second life rests on the shape of AI demand. Training the largest frontier models is a small number of enormous workloads that migrate to the newest hardware immediately; inference, fine-tuning, and the long tail of enterprise applications are an enormous number of modest workloads that are indifferent to whether they run on the newest hardware so long as the cost per unit of useful output is competitive. As AI adoption broadens, the second category grows faster than the first, and the installed base of prior-generation accelerators is the natural home for it. The 2026 Hopper experience is the strongest evidence for this case: a chip two generations behind the frontier saw its rental rates rise 30 percent in the back half of the year because the inference workloads that could run on it multiplied faster than the Blackwell and Rubin supply that could displace it.[52] Model fine-tuning, batch inference for non-latency-sensitive applications, enterprise retrieval and summarization, scientific computing, simulation, and increasingly the training and operation of robotics and autonomous systems all represent workload categories that can absorb prior-generation compute at prices that cover its operating cost with a margin.
The case against a robust second life rests on three counterarguments. The first is energy. As the performance-per-watt gap between generations widens, an older system’s cost per token at the margin rises relative to a newer one’s, and in a power-constrained market where every megawatt of energized capacity has an opportunity cost, an operator may rationally retire a functioning older system to free its power allocation for a newer one, even if the older system could still cover its direct operating cost. The second is software. As models are increasingly designed and quantized for the numerical formats and memory architectures of the newest hardware, running them efficiently on older hardware requires engineering effort that small operators may not have. The third is the lumpiness of supply. The second-life market works when prior-generation systems arrive gradually; it breaks when they arrive all at once, as they would if several large operators refreshed simultaneously or if a demand shock forced several to liquidate.
The honest conclusion is that the second economic life of GPUs is real, is substantial, and is conditional. It is real because the Hopper evidence shows it operating. It is substantial because the inference and agentic workloads that drive it are the fastest-growing segment of AI demand. And it is conditional on the three variables the counterarguments identify: the availability of energized capacity, the portability of the software stack, and the orderliness of the supply of displaced systems. A lender who underwrites a GPU lease on the assumption of a robust second life is making a bet on all three, and the collateral half-life metric is a way of making that bet explicit and testable rather than implicit and forgotten.
One further observation belongs here, because it reframes the entire question. The secondary market does not merely reveal the collateral value of a GPU; it reveals what the primary price actually was. A trade publication covering the used-GPU market put the point with a bluntness that financial analysts would do well to remember.
“The sticker price was never the real price. The secondary market was.”
— Hashrate Index, used-GPU market analysis [82]
If that is correct, then the appropriate V(0) for any GPU financing is not the invoice but the price at which a comparable system of the prior generation traded in the secondary market at a comparable point in its life, adjusted for the performance differential. That is a far more conservative starting point than most financings appear to use, and it is the starting point this paper recommends.
Section 3: Collateral Underwriting, Lease Design, and Risk Allocation
3.1 Underwriting beyond GPU purchase prices
If the first section established that the GPU has become a financeable asset and the second established that its value decays on a timeline that may be shorter than its financing, the third section turns to the question that occupies the credit committees of the institutions that have committed $40 billion to SpaceX and tens of billions more to its peers: how does one underwrite such an exposure responsibly? The answer, which the better-structured transactions already reflect, is that one does not underwrite the GPU at all, or not primarily. One underwrites the cash flow the GPU generates, the counterparty that pays it, the facility that houses it, and the contractual protections that govern what happens when any of those fails. The GPU purchase price is the beginning of the analysis, not the end.
Utilization is the first and most fundamental variable. A GPU system generates revenue only when it is running workloads for which someone is paying, and the difference between a fleet operating at 90 percent utilization and one operating at 60 percent is, in a high-fixed-cost business, the difference between comfortable debt service and distress. The utilization that matters for underwriting is energized utilization, meaning the fraction of installed capacity that is both powered and sold, since a GPU sitting in a rack awaiting a grid interconnection contributes nothing to debt service while consuming interest. SpaceX’s disclosure that its nameplate compute reached 1.7 gigawatts at September 30, 2026, up from 1.4 gigawatts at the end of June and 0.6 gigawatts a year earlier, is the kind of operational metric that an underwriter needs, but it must be read alongside the question of how much of that nameplate is energized and contracted.[10]
Contracted revenue is the second variable, and it is the one on which the hyperscaler defense of long depreciation schedules principally rests. When a lessee can demonstrate that the capacity being financed is already sold under multi-year contracts to creditworthy customers, the lender’s exposure to collateral value is correspondingly reduced, because the lease will be serviced from contracted cash flow regardless of what the equipment would fetch in a sale. SpaceX’s compute-leasing business illustrates the strength and the limits of this argument simultaneously. The company has disclosed arrangements under which Anthropic pays approximately $1.25 billion per month through May 2029 for capacity at the Colossus 1 datacenter, a $13.3 billion annualized compute lease signed in September 2026 with an undisclosed customer, and total annualized compute lease revenue of approximately $41.1 billion across its customer base.[12] Those are very large numbers, and they support the thesis that the Vera Rubin capacity will be sold before it arrives. But they also reveal that SpaceX’s AI segment is substantially a compute landlord whose revenue depends on a small number of large, sophisticated tenants, each of whom is simultaneously a competitor, a customer, and a builder of its own capacity, and whose leases will be re-bid against newer hardware when they expire.
Customer concentration is therefore the third variable, and it is the one that links the SpaceX exposure most directly to the correlated-default scenario of Section 3.5. The institutional investors who are lending to SpaceX against Vera Rubin GPUs are, in many cases, the same institutions that are lending to Anthropic against Google TPUs, to CoreWeave against Blackwell systems, and to the datacenter developers that house all three. If a single large model developer were to reduce its compute commitments, the effect would propagate through several financings at once.
Power costs are the fourth variable and are the channel through which Layer 1 of the Five-Layer AI Economy enters the credit analysis directly. Electricity is the largest operating cost of a GPU fleet after depreciation and interest, and its price determines the margin on which the second economic life of the equipment depends. A lender should model debt-service coverage under a range of power price assumptions, including the scenario in which regional constraints force the operator to procure behind-the-meter generation at a premium.
System maintenance, the fifth variable, is less glamorous but no less real. Liquid-cooled rack-scale systems have failure modes, parts-replacement needs, and firmware dependencies that loose GPUs did not, and the covenants governing who bears maintenance cost and who bears the risk of a system falling out of warranty bear directly on recovery value.
Cash flow available for debt service, finally, is the synthesis of all of the above. It is the quantity that an underwriter should project, quarter by quarter, across the life of each lease schedule, under base and stressed assumptions, and it is the quantity against which the quarterly rent of one-twentieth of financed cost plus interest should be measured. Debt-service coverage, not loan-to-value, is the primary metric for a GPU lease to a going concern; loan-to-value becomes primary only when the going concern stops going.
3.2 Amortization versus depreciation
The relationship between the amortization of the financial obligation and the depreciation of the collateral is, in my view, the single most important and least discussed feature of GPU lease design, and it deserves to be made concrete. Consider a $100 equipment tranche financed on the SpaceX–Apollo template: twenty quarterly principal payments of $5 each, so that the principal balance declines by $20 per year on a straight line and reaches zero at the end of year five. Now suppose that two years into the lease, after eight quarterly payments, the lessee defaults and the lessor must look to the equipment. The principal outstanding is $60. What is the equipment worth?
| Item | Amount |
| Original financed cost of the tranche | $100 |
| Quarterly principal payments made (8 of 20) | $40 |
| Principal remaining after eight quarters | $60 |
| Hypothetical stressed net recovery value of the equipment | $35 |
| Illustrative equipment-only shortfall | $25 |
Table 2. Illustrative $100 equipment tranche after eight quarterly principal payments. Hypothetical values for analytical illustration only. They are not forecasts of SpaceX equipment prices, contract protections, or financing losses.
The hypothetical recovery of $35 corresponds, in the collateral half-life framework, to a system whose V(0) was roughly 70 percent of financed cost and whose value then decayed by about 30 percent per year, which is the “correlated impairment” regime of Table 1, in which the half-life is close to two years. Under that regime, the equipment alone covers slightly less than 60 percent of the outstanding principal at the two-year mark, and the lessor’s loss, absent other protections, is about a quarter of the original financing. The arithmetic improves rapidly as the lease ages, because principal continues to amortize at $20 per year while the equipment, having already absorbed its steepest decline, decays more slowly in absolute dollars: at the three-year mark, under the same assumptions, principal outstanding is $40 and the equipment might be worth $24, a shortfall of $16; at the four-year mark, principal outstanding is $20 and equipment might be worth $17, a shortfall of $3. The exposure is front-loaded, which has two implications. The first is that the critical window for a GPU lease is the first thirty months, during which the amortization has not yet caught up with the decay. The second is that any structural feature that accelerates amortization in the early period, or that provides supplemental collateral during it, is worth disproportionately more than the same feature applied later.
This is why the straight-line, one-twentieth-per-quarter amortization in the SpaceX–Apollo structure, while conventional, is not obviously the right shape for an asset whose decay is front-loaded. A declining-balance or front-loaded amortization, in which the lessee pays, say, 7 percent of financed cost per quarter in the first two years and 3 percent per quarter thereafter, would track the equipment’s value curve more closely and would reduce the lessor’s peak exposure substantially, at the cost of higher early rent for the lessee. Whether the parties considered such a structure is not disclosed. What is clear is that the comparison between the amortization curve and the depreciation curve, rather than either curve considered independently, is the quantity that determines the adequacy of the collateral, and that a lender who looks only at the ending loan-to-value, which is zero by construction in a fully amortizing lease, has not looked at the risk.
3.3 Collateral protections and contractual safeguards
Because the equipment alone may not cover the outstanding obligation in the early years of a lease, the documentation of a GPU financing typically layers additional protections on top of the lessor’s ownership of the equipment, and the quality of those protections is the second half of the credit analysis. The SpaceX 8-K does not disclose the contractual safeguards of the master finance lease, and the following discussion therefore describes the protections that are customary in the asset class and that an analyst should look for when the documentation becomes available, rather than asserting that any particular protection is present.
Maintenance covenants require the lessee to keep the equipment in the condition, and under the warranty and support arrangements, necessary to preserve its value and its remarketability. For rack-scale liquid-cooled systems these covenants are more demanding than for loose servers, and their breach is an early warning of distress. Insurance requirements protect against physical loss, and in some transactions against specified business interruption. Reserves, whether in the form of cash collateral, a debt-service reserve account, or a prepayment of several quarters’ rent, provide a liquidity cushion that allows the lessor to continue to be paid while it exercises remedies. The fact that rent under the SpaceX lease is payable quarterly in advance is itself a modest form of reserve, since the lessor always holds one quarter’s payment ahead of the service period.[1]
Recourse and guarantees determine whether the lessor’s claim runs only to the equipment and the lessee’s lease payments or extends to the lessee’s general assets. A finance lease to an operating company of SpaceX’s size is ordinarily full recourse to that company, which means that in a default the lessor is an unsecured creditor for any deficiency after realizing on the equipment; the recovery on that unsecured claim depends on the lessee’s broader solvency and on the lessor’s position relative to the company’s bondholders. Where the lessee is a special purpose vehicle or a project subsidiary, as in the Valor structure, a parent guarantee may be the principal source of credit support, and its scope, whether it is a guarantee of payment or merely of performance, and whether it is capped, is a central term.
Lien priorities govern the lessor’s position relative to other creditors with claims on the same or related assets. In a datacenter where financed GPUs sit in racks financed by one lender, inside a building financed by another, drawing power under a contract pledged to a third, the question of who may remove what, and when, is not trivial, and the inter-creditor arrangements that resolve it are among the most heavily negotiated documents in the structure.
Equipment substitution rights, finally, are the protection most specific to the obsolescence problem. A lease that permits the lessee to substitute newer equipment of equal or greater value for the financed equipment, with the lessor’s lien attaching to the replacement, allows the collateral to be refreshed over the life of the lease and partially insulates the lessor from the decay of a single generation. Conversely, a lease that permits the lessee to return equipment early, or that gives the lessee a purchase option at a fixed price, shifts residual value risk toward the lessor. The direction in which these options run is a direct allocation of collateral half-life risk between the parties, and it is the term I would most want to read in the SpaceX documentation.
Vendor support deserves separate mention because it is the innovation that distinguishes the 2026 vintage of GPU financing from its predecessors. In the Anthropic transaction, Broadcom’s residual value support agreement on the A1 and A2 notes converted a credit exposure to a startup, secured by custom chips with no secondary market, into something that institutional investors could treat as near-investment-grade, by placing the residual risk with a party that had both the balance sheet to bear it and a commercial interest in the chips retaining value.[7] The structural logic is compelling: the vendor is the party best positioned to remarket returned equipment, best informed about its future relevance, and most motivated to see the financing succeed so that the next one can be done. Whether NVIDIA has provided, or will provide, comparable support in the SpaceX transaction is not disclosed. Given NVIDIA’s deep involvement in the capital structure of its customers’ demand, its anchor role in the Valor vehicle, and its indicated $3 billion participation in Meta’s equity offering in the same week, it would not be surprising if some form of vendor engagement with the collateral existed, but the record is silent, and an analyst should treat the question as open.
3.4 The installed-system valuation problem
A persistent error in the public discussion of GPU collateral is to reason from the resale price of an individual chip to the recovery value of a financed system. The two are related, but the relationship is neither simple nor stable, and it runs in both directions.
The reason is that what is financed in a transaction like SpaceX–Apollo is not a pallet of GPUs but a set of GPU pods, which is to say complete, integrated computing systems. A Vera Rubin NVL72 rack comprises 72 Rubin GPUs, 36 Vera CPUs, HBM4 memory, NVLink switches delivering 3.6 terabytes per second per GPU, Spectrum-X or Quantum networking, power distribution, liquid cooling manifolds, and the cable-free tray design that NVIDIA says makes assembly and servicing eighteen times faster than Blackwell systems.[26] The GPUs are the most expensive components, but they are not separable from the rest without destroying much of the system’s value, and the rest has its own, quite different, depreciation characteristics. Networking switches, power distribution units, and cooling infrastructure have longer useful lives than accelerators and are less exposed to generational repricing; one analysis has proposed component-based depreciation in which GPU modules are depreciated over three and a half to four and a half years with a 30 to 35 percent salvage assumption while chassis, networking fabric, and cooling are depreciated over six to eight years.[37] For collateral purposes, the implication is that a rack’s recoverable value is a weighted blend of a fast-decaying component and a slow-decaying one, and that the blend’s half-life is longer than the GPU’s alone.
But there is a countervailing consideration that often dominates. A rack-scale system has value to a secondary buyer only if that buyer has somewhere to put it, and “somewhere” means an energized datacenter with the power density, the liquid cooling capacity, and the networking topology the rack requires. In a market where grid interconnection queues in PJM average 40 months against an 8-to-11-month target, where substation transformer lead times exceed 160 weeks, and where ERCOT’s large-load queue reached approximately 410 gigawatts in April 2026 with datacenters accounting for 87 percent of it, the energized slot is scarcer than the rack.[61] A lender repossessing racks from a datacenter whose power contract has been terminated holds equipment that, however technically sound, cannot generate revenue until a new home is found for it, and the pool of buyers with a vacant energized slot of the right specification at the right moment is narrow. The individual chip resale quotations that circulate in secondary markets, which reflect the price at which a loose GPU can be shipped to a buyer who already has a slot, systematically overstate the recovery economics of a complete system that must be removed, transported, and re-homed.
The practical conclusion for underwriting is that the installed-system valuation problem argues for two adjustments. The first is a haircut to chip-level resale quotations that reflects the costs and delays of system-level remarketing; the second, and more important, is a recognition that the most valuable protection a lender can obtain is not a claim on the equipment alone but a claim on, or at least a stable relationship with, the energized facility in which it sits. A lease whose remedies include the right to continue operating the equipment in place, under the existing power and colocation arrangements, for a period sufficient to remarket it, is worth far more than one whose only remedy is removal.
3.5 Refinancing and correlated default risks
The underwriting framework of this section has so far treated each GPU financing as an isolated exposure. The final subsection considers what happens when the exposures are not isolated, and it is the consideration that connects the SpaceX–Apollo transaction to the financial stability questions that Section 4 takes up.
The AI infrastructure financing market has a structure that concentrates risk along several dimensions simultaneously. It concentrates by hardware generation, because the great majority of the equipment financed in any given year is the current NVIDIA architecture, so that a repricing of that architecture affects every lender at once. It concentrates by counterparty, because a small number of model developers, SpaceX, Anthropic, OpenAI, Meta, Google, and Microsoft, account for the overwhelming majority of compute demand, so that a change in any one of their plans propagates through many financings. It concentrates by financier, because Apollo, Blackstone, Blue Owl, KKR, BlackRock, and a handful of others lead a large share of the transactions and hold meaningful retained positions in each. It concentrates by maturity, because the five-year leases and term loans written in 2025 and 2026 will all require refinancing or repayment in 2030 and 2031, within a window of a few quarters. And it concentrates by geography, because the facilities that house the equipment cluster in Texas, Virginia, Arizona, Tennessee, and a handful of other states whose grids are under correlated stress.
The scenario that follows from this structure is not difficult to construct. Suppose that in 2028 the arrival of a new architecture, combined with a moderation in the growth of model developers’ compute budgets, produces a decline in GPU rental rates of a magnitude comparable to the Hopper decline of 2025. Several operators who financed Blackwell and early Rubin systems on five-year schedules find that their contracted revenue, as contracts roll off and are re-bid against newer hardware, no longer covers debt service with the margin their lenders require. Some seek to refinance; the refinancing market, which was open in 2026 at spreads that already reflected concern, is tighter, because the lenders who would provide the refinancing are the same lenders who hold the existing exposures and are now marking them down. Some seek to sell equipment; the secondary market, which absorbed the gradual arrival of Hopper systems in 2025 and 2026, is confronted with the simultaneous arrival of systems from several distressed sellers and prices fall faster than any individual seller anticipated. The decline in secondary prices lowers the recovery values that other lenders are carrying, which tightens their covenants, which produces more sellers. This is the correlated impairment regime of Table 1, and it is characterized by the simultaneous decline of compute prices, utilization, and secondary-market liquidity, which is to say by the failure of exactly the three variables on which the second economic life of GPUs depends.
The early warnings of such a dynamic were visible in the credit markets in the autumn of 2026. CoreWeave’s bonds were yielding close to 15 percent in mid-October, a level at which, as one analysis observed, credit markets were pricing the company closer to distressed debt than to a growth company.[16] Credit risk gauges for Oracle, Broadcom, and SpaceX climbed to fresh highs on October 20, as investors braced for a further wave of bond issuance from companies that had collectively raised nearly $500 billion of new debt in 2026 to finance AI infrastructure.[64] S&P had downgraded Oracle to BBB-, one notch above speculative grade, in July, citing elevated debt levels, and Oracle’s five-year credit default swaps traded at multi-year highs.[65] None of this constituted a crisis, and the strongest hyperscalers’ spreads remained at or below those of the broad AA-rated corporate index.[79][80] But the market was beginning to differentiate, and the direction of the differentiation, toward wider spreads for the operators whose capital spending was closest to hyperscaler scale without a hyperscaler’s balance sheet, was exactly what the correlated-default analysis would predict.
The underwriting implication is that diversification across GPU financings is less protective than it appears, because the financings share common factors, and that the appropriate stress test for a GPU lease is not a scenario in which the lessee alone defaults but one in which the lessee defaults at the same time as several others, into a secondary market that is already falling. That is a demanding standard, and it is the standard that the asset class will be held to if the correlated scenario ever materializes.
Section 4: The Institutional Investment Market for AI Collateral
4.1 The rise of specialized infrastructure credit
The transactions examined in the preceding sections are individually large, but their collective significance lies in what they suggest about the formation of a market. A single $40 billion lease is an event; a sequence of structurally similar financings, led by overlapping groups of institutions, secured by overlapping classes of equipment, and distributed to overlapping pools of insurance and pension capital, is the beginning of an asset class. The question this section addresses is whether GPU leasing is developing into a standardized institutional investment market, with the depth, the transparency, and the secondary liquidity that term implies, or whether it remains a collection of bespoke transactions whose apparent similarity conceals material differences in risk.
The evidence for standardization is considerable. The template that Apollo refined from Valor in January to SpaceX in October, and that Apollo and Blackstone applied to Anthropic in between, has recognizable common elements: an asset-owning vehicle, a multi-year amortizing lease to the operator, tranched or syndicated senior capital, and in at least one case vendor residual support. JPMorgan projected that annual datacenter securitization issuance could reach $30 billion to $40 billion in 2026 and beyond, and the asset-backed securities market has begun to absorb GPU-collateralized paper alongside the datacenter real estate paper it has financed for years.[67] Apollo has stated an explicit ambition to create secondary trading markets for its GPU lease exposures, and has forecast that by the end of the decade roughly half of the ten largest borrowers in the U.S. bond market will be hyperscalers.[68] Ares has estimated that private credit could finance about $5.5 trillion of capital across debt and equity in global infrastructure, including AI-related projects, through 2035.[67] These are the projections of participants with an interest in the market’s growth, and they should be discounted accordingly, but they describe a direction of travel that the deal flow of 2026 has confirmed.
The evidence against standardization is equally real, and it is of a kind that the participants have less incentive to emphasize. The transactions differ in ways that matter enormously for credit: in whether the lessee is an investment-grade public company or a venture-backed startup, in whether the collateral is fungible NVIDIA hardware or single-operator custom silicon, in whether a vendor stands behind the residual value, in whether the equipment is already installed in an energized facility or awaiting a grid connection, and in the allocation of substitution and early-return rights. A market in which these differences are priced is a healthy one; a market in which they are obscured by a common label is one in which risk is being mispriced. The history of structured finance offers ample precedent for the latter, and the pace at which GPU financing has scaled, from a $3.5 billion capital solution in January to a $40 billion commitment in October of the same year, leaves little time for the development of the independent valuation infrastructure, the rating methodologies, and the disclosure norms that a mature asset class requires.
4.2 Apollo, BlackRock, KKR, banks, and insurers
Institutional investors can participate in AI collateral through several vehicles, and the vehicle shapes both the risk the investor bears and the information it receives.
Direct lease ownership, in which the investor’s fund or affiliate is the lessor and holds title to the equipment, is the most concentrated form of exposure and offers the most control. It is the position Apollo’s managed entity occupies in the SpaceX transaction. The lessor bears the full residual value risk, subject to whatever contractual protections it has negotiated, and it captures the full return. It also bears the operational burden of remarketing equipment in a default, which requires expertise that financial institutions have had to acquire or contract for.
Direct lending, in which the investor provides a secured term loan to an operator that owns the equipment, is the form CoreWeave’s delayed-draw term loans took and is the form most familiar to private credit funds. The lender’s recovery in a default depends on its security interest in the equipment and in the contracted receivables, and on its position relative to other creditors.
Private credit funds, whether drawdown vehicles or the perpetual, semi-liquid structures that have grown rapidly among wealth channels, pool these exposures across transactions and offer investors diversification, at the cost of the concentration risk that Section 3.5 described. The Federal Reserve’s May 2026 Financial Stability Report noted redemption requests and weaker sentiment in parts of private credit, and reporting in April indicated that Blue Owl, Ares, Apollo, Blackstone, and KKR had all limited redemptions from certain private credit funds as investor concern mounted.[17] Fitch’s private credit default rate reached a record 6.0 percent in May 2026.[23] These stresses were not principally AI-related, but they bear on the capacity of the private credit system to absorb AI collateral losses if they come.
Structured obligations, including the tranched notes of the Anthropic transaction and the asset-backed securities that refinance equipment leases, allow risk to be sliced by seniority and distributed to investors with different appetites. The senior tranches, particularly where vendor-supported, can be placed with insurance company general accounts and other investors constrained to investment-grade exposure; the subordinated tranches are held by the arrangers or sold to funds with higher return targets. Structuring is what allows the market to scale, and it is also what allows the risk to migrate to holders who are furthest from the underlying asset and least equipped to evaluate it.
Insurers deserve particular attention because they are, through affiliates like Athene and through their general accounts, the ultimate holders of a large share of this paper, and because their regulatory and accounting treatment of it shapes demand. A five-year amortizing lease to an investment-grade lessee, held to maturity and secured by equipment, receives favorable capital treatment under risk-based capital frameworks, and the yield pickup over comparable corporate bonds is attractive. The question regulators have begun to ask is whether the frameworks adequately capture the obsolescence risk that distinguishes GPU collateral from the aircraft and railcars those frameworks were designed around.
4.3 The CoreWeave comparison
CoreWeave is the most transparent case study the market offers, because it is a public company whose business consists almost entirely of financing GPU systems and renting their capacity, and whose SEC disclosures lay bare the mechanics that private transactions conceal.[13] Its experience through the second quarter of 2026, the most recent period reported as this paper is written, with third-quarter results scheduled for November 10, 2026, illuminates the SpaceX–Apollo structure by contrast.
At June 30, 2026, CoreWeave reported a revenue backlog of approximately $104.2 billion, of which $103.7 billion was remaining performance obligations, up 246 percent year over year, and the figure excluded more than $25 billion of net new customer commitments added early in the third quarter.[14] Second-quarter revenue more than doubled to $2.58 billion; the company’s full-year 2026 revenue guidance rose to between $12.4 billion and $13.2 billion. The backlog is the strongest evidence that the demand for financed compute is real and contracted. Against it stood approximately $35.1 billion of recourse and non-recourse debt principal, more than $16 billion of lease obligations, 2026 capital expenditure guidance of $35 billion to $39 billion, a second-quarter GAAP net loss of $626 million, quarterly interest expense of $640 million guided toward $860 million to $940 million in the third quarter, and an adjusted operating margin that had slid to 5 percent from 16 percent a year earlier.[15][66] The company had raised more than $10 billion of unsecured debt and convertible bonds plus a $3.1 billion term loan in the quarter, and its capital spending, as S&P Global Market Intelligence observed, was close to hyperscaler scale without a hyperscaler’s balance sheet.[16] By mid-October its bonds yielded close to 15 percent.
The CoreWeave disclosures reveal four mechanics that bear on GPU collateral generally. The first is debt sizing against contracted revenue: CoreWeave’s delayed-draw term loans were sized to the contracted cash flows of specific customer agreements, with advance rates and covenants tied to the backlog’s conversion, which is the project-finance logic that the SpaceX lease applies at a larger scale and to a stronger lessee. The second is the role of financial guarantees and customer credit: a backlog is only as good as the customers who owe it, and CoreWeave’s concentration in a small number of very large counterparties, including Microsoft, OpenAI, Meta, and NVIDIA, means that its credit is, in substantial part, a derivative of theirs. The third is the project-level operating assumption: CoreWeave’s six-year useful life, its utilization targets, and its power cost assumptions are the inputs to the debt-service coverage on which its lenders rely, and any one of them proving optimistic flows directly to coverage. The fourth, and most telling, is the gap between adjusted and GAAP results: CoreWeave’s 59 percent adjusted EBITDA margin coexisted with a $626 million GAAP loss because depreciation and interest, which the adjusted figures exclude, are not accounting noise for a GPU lessor; they are the core costs of the business model.
The contrast with SpaceX is instructive on each count. SpaceX enters the Apollo lease with roughly $90 billion of cash and marketable securities, a BBB+ rating, a diversified revenue base in which connectivity generated about 85 percent of operating income even as AI generated 52 percent of revenue, and a net profit of $4.4 billion in its most recent quarter.[9] It is, in every respect that matters to a lender, a far stronger lessee than CoreWeave is a borrower. But it is also guiding to fourth-quarter capital expenditure of $35 billion to $40 billion, more than double the third quarter’s $19.6 billion, and analysts have projected its total debt rising from roughly $50 billion in 2025 toward $350 billion by 2029 if its capital spending trajectory holds.[11] The lesson of CoreWeave is not that SpaceX will follow its path; it is that the path exists, that it is paved with contracted backlog and five-year financing, and that the distance between a strong lessee and a stressed one can be traversed in a small number of quarters when capital spending exceeds revenue and the financing is secured by equipment whose value is falling.
4.4 Financial-system implications
The question of whether GPU financing poses risks beyond the parties to each transaction moved, in 2026, from the margins of the policy debate to its center, and the official sector’s assessments provide the institutional context in which the SpaceX–Apollo deal should be read, without implying that the deal itself has been classified as a systemic risk by any authority.
The Federal Reserve’s Financial Stability Report of May 8, 2026, recorded that 50 percent of surveyed market participants cited artificial intelligence as a salient risk to the financial system, up from 30 percent in the fall of 2025 and from single digits a year before that.[18] The report observed that AI capital spending was increasingly debt-funded, creating leverage across companies, lenders, and funding markets, and that stress in private credit could spill into broader markets if expectations about AI returns weakened.[17] It did not single out any transaction or asset class for supervisory action, and its principal concerns remained geopolitical risk and asset valuations, but it placed debt-funded AI capital expenditure and private credit vulnerabilities squarely within its framework for the first time at that intensity.
The International Monetary Fund’s October 2026 Global Financial Stability Report, released on October 13 during the Annual Meetings in Bangkok, went further. It named a potential AI-driven market correction as one of three salient near-term risks to global financial stability, alongside a sharp rise in sovereign bond yields and a major cyberattack, and it included a background analysis titled “Rising Leverage in the AI Ecosystem.”[19] The report documented that large cloud and datacenter operators’ share of global investment-grade corporate bond issuance had risen to 10 percent as of July 2026 from less than 1 percent at the end of 2024, that dwindling free cash flow had pushed hyperscalers to rely increasingly on debt, and that U.S. bank loans to nonbank financial intermediaries and exposures to private credit firms had risen sharply.[21] It warned policymakers against a specific failure of vigilance.
“complacency”
— International Monetary Fund, Global Financial Stability Report, October 2026 [20]
The IMF’s companion World Economic Outlook modeled a scenario in which the unwinding of AI-related investment directly reduces global GDP by as much as 0.4 percent and investment by as much as 1.3 percent by 2028, a reminder that the AI capital expenditure that has supported global growth in 2026 is a macroeconomic variable whose reversal has macroeconomic consequences.[70] The European Central Bank’s May 2026 Financial Stability Review drew an explicit comparison between the AI private credit boom and the pre-2008 subprime buildup, with stress models showing pension fund losses of 5 to 6 percent if AI capital spending disappointed, and the Financial Stability Board’s first dedicated report on private credit vulnerabilities, published May 6, estimated the market at $1.5 trillion to $2 trillion and named leverage in opaque, multi-layered structures as its first pressure point.[69] The official sector’s message, taken together, was not that AI infrastructure finance is unsound but that it is large, leveraged, concentrated, and insufficiently transparent, and that those four properties have historically been the preconditions of financial instability when the underlying asset disappoints.
Academic economists have reached a range of conclusions that bracket the official sector’s caution. Aswath Damodaran of New York University’s Stern School of Business, whose October 8, 2026 analysis traced how the AI capital expenditure boom was flowing through the aggregate financial statements of U.S. public companies, has emphasized the financing channel specifically.
“a big chunk of it is funded with debt, the debt coming from private capital rather than banks”
— Aswath Damodaran, Professor of Finance, NYU Stern School of Business [43]
Damodaran’s slides accompanying the October analysis laid out the asymmetry plainly: if the AI investment delivers a return on capital above the cost of capital, the builders’ earnings rise and their value rises with them; if it does not, the accountants will eventually revisit the investments, and if the capital expenditure was funded disproportionately with debt, there will be defaults and distress on the part of the borrowers and pain for the lenders, banks and private credit alike.[42] Daron Acemoglu of MIT, the 2024 Nobel laureate in economics, has been the most consistently skeptical mainstream voice on the near-term productivity returns to AI investment, and his assessment of the capital being deployed has not softened as the numbers have grown.[46]
“A lot of money is going to get wasted.”
— Daron Acemoglu, Institute Professor, Massachusetts Institute of Technology [45]
Erik Brynjolfsson of Stanford’s Digital Economy Lab represents the optimistic pole, arguing that the investment phase is giving way to measurable productivity gains and that the scale of the buildout is proportionate to the scale of the opportunity.
“We’re creating the potential to have massive productivity gains and a lot more wealth.”
— Erik Brynjolfsson, Director, Stanford Digital Economy Lab [47]
The disagreement among these scholars is not a disagreement about the facts of the financing; it is a disagreement about the returns at Layers 4 and 5 of the Five-Layer AI Economy, and it is precisely because that disagreement is unresolved that the collateral at Layer 2 matters. If Brynjolfsson is right, the leases will be serviced from operating cash flow and the equipment’s resale value will never be tested. If Acemoglu is right, the equipment is what the lenders will have left.
4.5 Toward transparent GPU collateral reporting
The gap between what the market needs to know about GPU collateral and what it is told is wide, and it is widest precisely where the exposures are largest, in the private transactions whose terms are disclosed, if at all, in the summary form of an 8-K. The SpaceX filing discloses the size, the term, the amortization shape, and the timing of the Apollo lease; it does not disclose the fixed rate, the advance rate against equipment cost, the substitution and early-return rights, the maintenance and insurance covenants, the presence or absence of vendor support, the inter-creditor arrangements with SpaceX’s other secured and unsecured creditors, or the lessor’s assumptions about residual value. An institutional investor in the senior tranche presumably knows these things; a bondholder of SpaceX, a shareholder, a regulator assessing systemic exposure, or an analyst attempting to compare the transaction with its peers does not.
I propose that the market, and if the market will not then its supervisors, converge on a consistent minimum disclosure standard for GPU-collateralized financings above a materiality threshold, comprising six elements. First, hardware generation: the architecture and configuration of the financed equipment, stated with enough specificity that its position relative to the technological frontier can be tracked over the life of the financing. Second, installation status: whether the equipment is delivered, installed, energized, and operating, and the expected timeline for each stage if not, since equipment awaiting a grid connection is collateral of a different character from equipment generating revenue. Third, cash-flow coverage: the ratio of contracted and projected cash flow available for debt service to the required lease or loan payments, under base and stressed assumptions, with the key assumptions disclosed. Fourth, concentration: the share of the financed capacity’s revenue attributable to the largest customer and the largest application category. Fifth, recovery valuation: an independent estimate of the equipment’s net recoverable value at inception and at regular intervals thereafter, together with the methodology, so that the collateral half-life can be observed rather than inferred. Sixth, maturity schedule: the amortization profile and any balloon, extension, or refinancing dependency, so that the clustering of maturities across the market can be monitored.
None of these elements is exotic. Each has an analogue in the disclosure regimes for aircraft finance, commercial mortgage-backed securities, or project finance, asset classes that developed standardized reporting after, rather than before, episodes of mispricing. The argument for developing it before, in the case of GPU collateral, is that the asset decays faster than any of those, and that the window in which a disclosure regime could be built in calm conditions may be shorter than the five-year leases now being written.
Section 5: The 2027–2030 Outlook — When Innovation Outruns the Financial Contract
5.1 Three possible financing futures
The analysis of the preceding sections can be organized into three scenarios for the period during which the SpaceX–Apollo lease schedules, and the broader cohort of 2025–2027 GPU financings, will run. The scenarios are not predictions; they are structured descriptions of the ways in which the variables identified in this paper could combine, and they are intended to help an analyst locate the actual future, as it arrives, within a framework rather than to forecast it.
| Scenario | Hardware economics | Collateral half-life regime | Financing implications |
| Adaptive expansion | AI demand absorbs successive GPU generations; inference and agentic workloads grow faster than frontier supply; second-life markets deepen | ~4 years; decay outrun by straight-line amortization | Strong cash flow supports repayment; equipment value rarely tested; spreads tighten; market standardizes |
| Orderly repricing | New accelerators lower rents and resale prices on a predictable cadence; demand grows but more slowly than capacity; margins compress | ~3 years; decay roughly matches amortization at midpoint | Lenders shorten maturities, raise advance-rate haircuts, require larger reserves and vendor support; weaker operators refinance at wider spreads; selective impairments |
| Correlated impairment | Demand weakens or capacity exceeds effective use; power constraints strand capacity; several operators refresh or liquidate simultaneously | ~2 years or less; decay decisively outruns amortization | Default and recovery pressures rise together; secondary market liquidity collapses; losses migrate to private credit and insurance holders; official-sector intervention debated |
Table 3. Three financing futures for AI GPU collateral, 2027–2030. The collateral half-life regimes correspond to the illustrative scenarios of Table 1.
Adaptive expansion is the scenario in which the Hopper experience of 2026 generalizes. Each new architecture expands the total market for computation faster than it reprices the installed base, inference and agentic demand absorb prior-generation capacity at prices that cover operating cost with a margin, the second-life market deepens as more buyers acquire the expertise and the energized slots to use displaced systems, and the leases are serviced from operating cash flow without the equipment’s value ever being tested in a distressed sale. Under this scenario the collateral half-life is four years or longer, the straight-line amortization comfortably outruns the decay, and the asset class matures into something resembling aircraft finance, with standardized disclosure, rating methodologies, and a liquid secondary market in both equipment and paper. It is the scenario that the participants in the SpaceX–Apollo transaction are, implicitly, underwriting.
Orderly repricing is the scenario in which the technology cadence does what it has done but demand grows more slowly than capacity. New accelerators lower rents and resale prices on a predictable schedule, operators who financed prior generations see their margins compress as contracts roll and re-bid, and lenders respond by shortening maturities, raising the haircuts they apply to equipment value, requiring larger reserves, and insisting on vendor support as a condition of senior placement. Weaker operators refinance at wider spreads or are absorbed by stronger ones; selective impairments occur but do not cascade. Under this scenario the collateral half-life is about three years, roughly matching the amortization at the midpoint of the lease, and the outcome for a given financing depends on the quality of its contractual protections and the strength of its lessee. This is, in my assessment, the central case, and it is the scenario in which the design choices discussed in Section 3 matter most.
Correlated impairment is the scenario in which the variables fail together. Demand weakens, whether because model developers’ revenue disappoints, because a macroeconomic shock interrupts the capital flows that fund them, or because efficiency gains in models reduce the compute required per unit of output faster than usage grows. Power constraints strand capacity that was financed on the assumption it would be energized. Several operators refresh or liquidate at once, the secondary market is overwhelmed, and the recovery values that lenders are carrying prove unrealizable. Under this scenario the collateral half-life is two years or less, the decay decisively outruns the amortization, and the losses migrate through the structured tranches to the private credit funds and insurance general accounts that hold them, which is the transmission channel the IMF and the ECB have warned about. The probability of this scenario is not negligible, and its severity is a function of the leverage and concentration that the market has accumulated by the time it arrives.
5.2 Power grids and datacenter commissioning
Of all the variables that determine which scenario materializes, the one that has moved most decisively against the optimists in 2026 is the availability of energized datacenter capacity, and it deserves extended treatment because it is the variable that financial analysts, trained to think about chips and cash flows, are most likely to underweight.
The time between acquiring a GPU and earning revenue from it is, in the current environment, governed not by the chip’s delivery schedule but by the grid’s. A datacenter shell can be built in twelve to eighteen months; the power connection owns the schedule, through permitting and approvals, interconnection studies that in PJM averaged 40 months against an 8-to-11-month target, transformer and switchgear procurement, and commissioning.[61] Substation transformer lead times, which averaged roughly 140 weeks in 2023, exceeded 160 weeks in 2026.[59] The Electric Reliability Council of Texas was tracking a large-load interconnection queue of approximately 410 gigawatts as of April 2026, with datacenters accounting for roughly 87 percent of it, against a state whose entire generating capacity is a small fraction of that figure; ERCOT voted in June 2026 to introduce a batch-based review process for the queue.[58] PJM, the thirteen-state grid that serves Virginia, Pennsylvania, Ohio, Indiana, and Michigan among others, fell 6.6 gigawatts short of its reliability target in its December 2025 capacity auction, and Gartner has projected that 40 percent of AI datacenters will face power constraints by 2027.[60] In Northern Virginia, the largest datacenter market in the world, the wait for a grid connection has reached seven years by one industry estimate; Dominion Energy received interconnection requests for 60 gigawatts of new datacenter load, more power than the entire state consumes, against a realistic delivery capacity of 7 to 10 gigawatts by 2030.[59] Arizona Public Service had roughly 4.5 gigawatts of committed large-load demand in its queue with discussions underway for a further 19 gigawatts. Bloomberg reported that close to half of all planned U.S. datacenter builds in 2026 were projected to be delayed or canceled, and Stanford’s 2026 AI Index recorded $64 billion of datacenter projects blocked or delayed over the prior two years amid rising local resistance.[49]
The regional comparison the outline called for yields a consistent pattern with important local variation. Texas, operating its own grid under ERCOT, interconnects projects in two to three years, faster than PJM’s territory, and may overtake Virginia as the largest datacenter market by 2030; but its queue is the largest in absolute terms and its transmission network is straining, and developers are increasingly going behind the meter with on-site gas generation, as the Chevron–Microsoft arrangements and smaller projects illustrate. Arizona offers land, solar resources, and a utility actively courting load, but its committed queue already exceeds its near-term delivery capacity by a wide margin. Virginia is saturated, is phasing out the datacenter sales-and-use tax exemption that built Loudoun County, and faces the longest interconnection timelines in the country. Indiana, Ohio, Pennsylvania, and Michigan, all within PJM, have attracted a wave of announced projects, including Anthropic’s TPU deployments in Indiana, on the strength of available land and comparatively welcoming policy, but they share PJM’s capacity shortfall and its 40-month interconnection studies, and their advantage over Virginia is one of degree rather than kind. Tennessee, where SpaceX is building Colossus II in Memphis, sits outside PJM in the Tennessee Valley Authority’s territory, and Musk’s statement that he was spending four days a week in Memphis bringing the facility online, because power, cooling, and infrastructure rather than chips were the binding constraint, is the most direct testimony available that the Layer 1 problem reaches the largest and best-capitalized operators.[9]
The federal government’s response has been to treat permitting as the obstacle. The White House’s July 2025 AI Action Plan framed the issue in terms that have since become the administration’s refrain.
“almost impossible to build this infrastructure in the United States with the speed that is required”
— The White House, America’s AI Action Plan, July 2025 [63]
The plan’s second pillar, “Build American AI Infrastructure,” and the accompanying executive order on accelerating federal permitting of datacenter infrastructure, directed agencies to streamline environmental review, expand the FAST-41 process to datacenter and energy projects, and open federal lands for datacenter and power development.[62] These measures address the permitting component of the timeline; they do not manufacture transformers, and they do not shorten the interconnection studies that grid operators must perform to maintain reliability. The physical constraint will ease, but on the timescale of transmission construction and transformer manufacturing, which is years, not on the timescale of GPU delivery, which is quarters.
The implication for collateral analysis is direct and, I think, underappreciated. Energized datacenter capacity is becoming the scarce complement to GPU collateral, in the sense that a GPU without an energized slot has no second economic life and a lender without a claim on the slot has a weaker claim on the GPU. The SpaceX lease contemplates deployment at datacenters selected by the company, and SpaceX’s ability to energize 1.7 gigawatts of nameplate compute, with a target of two gigawatts or more by the end of 2026, is a material credit strength that a lessee without that capability would lack.[10] But it also means that the recoverable value of the Vera Rubin systems, in a default, would depend heavily on whether the lessor could keep them operating in place at Colossus or elsewhere rather than removing them into a market with few vacant slots of the required specification.
5.3 Export controls and secondary markets
A second variable that bears on secondary-market depth, and that financial analysts frequently overlook, is the regime of U.S. technology export controls, which constrains the geography of the buyer pool for financed equipment and the ability of a lessor to relocate it.
The regulatory landscape shifted repeatedly in 2025 and 2026. Between January and May 2025, the Bureau of Industry and Security issued and then rescinded the Framework for Artificial Intelligence Diffusion, which would have imposed a tiered global licensing regime on advanced computing integrated circuits, citing concerns that it would hinder U.S. innovation and complicate diplomatic engagement.[54] On January 15, 2026, BIS issued a final rule revising its licensing policy for certain earlier-generation chips, including NVIDIA’s H200 and AMD’s MI325X, from a presumption of denial to case-by-case review for exports to China and Macau, subject to performance thresholds and conditions, while maintaining the presumption of denial for the most advanced parts and for entities headquartered in countries of concern.[55] Then, on May 31, 2026, following reporting that hundreds of thousands of Blackwell-class GPUs had reached China-linked datacenters in Southeast Asia through foreign subsidiaries, BIS issued guidance clarifying that the location of a company’s ultimate parent, not the registration of its subsidiary, triggers licensing requirements for the leading-edge chips, including Blackwell and Rubin.[57] The guidance did not require already-deployed equipment to cease operating, but it tightened the compliance obligations of datacenter operators and infrastructure-as-a-service providers, who are now explicitly within the enforcement perimeter.
NVIDIA’s chief executive has been publicly critical of the controls’ effectiveness, noting that the company’s share of China’s AI accelerator market fell from roughly 95 percent to zero.
“The goals of the export controls are not being achieved”
— Jensen Huang, Chief Executive Officer, NVIDIA [56]
Whatever one’s view of the policy, its consequences for collateral are concrete. First, it narrows the potential buyer pool for Vera Rubin systems in a default by excluding, as a practical matter, the largest single market for AI compute outside the United States, and it imposes licensing friction on sales to buyers in many other jurisdictions. Second, it constrains the geographic relocation of financed equipment: a lessor that repossesses racks in Tennessee cannot simply ship them to the highest bidder abroad, and must instead find a domestic or allied-country buyer, which, given the energized-slot constraint, is a narrower set. Third, it creates a compliance risk that attaches to the equipment itself and to the operators who house it, since a datacenter found to be serving a prohibited end user faces enforcement that could interrupt the cash flows on which the lease depends. Fourth, and more subtly, it has created a black-market price signal: the Financial Times reported that NVIDIA’s export-restricted chips more than doubled in price on China’s gray market in the first half of 2026, which indicates that global demand for the hardware exceeds what the controlled channels supply, and which is, perversely, evidence that the collateral has value that lawful holders cannot fully realize.[51]
For the collateral half-life framework, export controls function as a liquidity haircut whose size varies with the policy regime, and the policy regime has changed three times in eighteen months. An underwriter should assume that the secondary market for financed equipment is, in the first instance, a domestic and allied market, and should value the collateral accordingly.
5.4 Models, inference, and agentic demand
The scenario that materializes between 2027 and 2030 will be determined, in the end, at Layers 4 and 5 of the Five-Layer AI Economy: by whether frontier models, enterprise AI agents, autonomous machines, software engineering tools, and robotics generate enough recurring revenue to sustain large fleets of accelerators across successive hardware generations. The evidence on that question in November 2026 is mixed, and the mixture is itself informative.
On the demand side, the evidence of sustained growth is strong. SpaceX’s AI segment grew from $2.56 billion of revenue in the second quarter of 2026 to $7.56 billion in the third, and the company has guided to first-quarter 2027 revenue of $33 billion to $37 billion across all segments.[11] CoreWeave’s backlog exceeds $104 billion. Anthropic raised $65 billion at a $965 billion valuation in May 2026 and committed to a $36 billion TPU lease; Meta committed to six gigawatts of AMD accelerators on top of its NVIDIA deployments; the IMF credited AI infrastructure investment with offsetting higher oil prices in its global growth forecast.[70] The rebound in H100 rental rates in the autumn of 2026 is direct evidence that inference demand is absorbing capacity faster than the obsolescence narrative predicted. Stanford’s AI Index recorded that generative AI reached 53 percent adoption in three years, faster than personal computers or the internet, and estimated U.S. consumer surplus from generative AI at $172 billion annually by early 2026.[49] Brynjolfsson has argued that the macroeconomic data now show the transition.
“transitioning out of this investment phase into a harvest phase”
— Erik Brynjolfsson, Stanford Digital Economy Lab, Financial Times op-ed, February 2026 [48]
On the revenue-quality side, the evidence is more equivocal. A substantial share of the demand that fills the backlogs is demand from model developers who are themselves not yet profitable and whose capacity commitments are funded by equity raised at valuations that assume future revenue. The circularity by which chip vendors invest in model developers who buy chips, and model developers lease capacity from infrastructure operators who are also their competitors, means that the same dollar of capital can appear as revenue at several points in the chain without a customer at Layer 5 having paid for anything. Damodaran’s observation that the only point of consensus is that companies have collectively invested a huge amount in AI capital expenditure captures the epistemic state of the market precisely: the spending is observable, the returns are projected, and the projections are what the financing is secured against.[41] The PwC global CEO survey in January 2026 found that 56 percent of chief executives reported no financial benefit from AI to date, even as an EY survey found 96 percent of firms investing in AI reporting some productivity gain, which suggests that the benefits are real at the task level and not yet consolidated at the enterprise level.[86]
The question for collateral is not whether AI will be valuable, on which the evidence is overwhelming, but whether the recurring revenue it generates will arrive on the schedule, and in the places, that the five-year leases assume. Agentic AI, in which models execute multi-step tasks autonomously, is the workload that most directly favors sustained demand for large fleets, because it consumes orders of magnitude more inference per user than conversational AI, and NVIDIA has explicitly positioned Vera Rubin’s disaggregated inference architecture, with its context-processing extension and its Groq-derived low-latency tier, around that workload.[26] Software engineering, where SpaceX’s acquisition of Cursor and the broader adoption of coding agents provide the clearest current evidence of enterprise willingness to pay, is the leading indicator. Robotics and autonomous systems are the longer-dated opportunity, and they are the workload category most likely to absorb prior-generation compute, because their inference requirements are modest per unit and vast in aggregate. If these workloads scale as their proponents expect, the adaptive expansion scenario follows. If they scale more slowly, or if efficiency gains in models reduce the compute required per task faster than the number of tasks grows, the orderly repricing scenario follows, and the equipment financed in 2027 will be re-bid in 2029 against hardware that does the same work for a fraction of the power.
5.5 A proposed Collateral Half-Life Dashboard
I close this section by proposing a compact set of measurements that an institutional investor in GPU-collateralized financings should track, quarterly, for each exposure and for the portfolio in aggregate. The dashboard is intended to make the collateral half-life framework operational, and each of its six measurements corresponds to a variable this paper has identified as determinative.
| Measurement | What it captures | Why it matters for the lease |
| Time until estimated recoverable value falls by half | The collateral half-life itself, estimated from independent valuation of comparable systems | Determines whether equipment decay outruns amortization; the core metric |
| Outstanding principal relative to stressed recovery value | Loan-to-value under a stressed, system-level (not chip-level) recovery assumption | Measures the equipment-only shortfall at each point in the lease; front-loaded exposure |
| Cash-flow debt-service coverage | Contracted and projected cash flow available for debt service divided by required rent | Primary protection while the lessee is a going concern; the metric that makes equipment value secondary |
| Energized utilization and actual compute revenue | Share of installed capacity that is powered and sold, and the revenue it produces | Distinguishes revenue-generating collateral from stranded collateral; the Layer 1 and Layer 3 channel |
| Customer and application concentration | Share of revenue from the largest customer and the largest workload category | Measures exposure to correlated default across financings; the Layer 4 and Layer 5 channel |
| Sensitivity to hardware replacement and electricity prices | Modeled change in coverage and recovery under a new-architecture repricing and under power price shocks | Captures the two exogenous shocks most likely to move the exposure between scenarios |
Table 4. The proposed Collateral Half-Life Dashboard: six quarterly measurements for institutional holders of GPU-collateralized financings.
The dashboard’s value lies less in any single measurement than in their joint movement. A financing whose half-life estimate is shortening while its coverage is stable and its energized utilization is high is one in which the lessee is outrunning the obsolescence, which is the adaptive scenario at the exposure level. A financing whose coverage is compressing while its concentration is rising and its recovery value is falling is one that is drifting toward correlated impairment, and the dashboard’s purpose is to make that drift visible early enough to act on it, through reserve requirements, through substitution, through refinancing, or through sale, before the market as a whole has recognized it.
Section 6: What Have We Learned? Seven Pillars
The preceding sections have moved from the anatomy of a single transaction to the structure of a market and from the structure of a market to the scenarios that will test it. Before concluding, it is worth distilling what the analysis has established into a small number of durable propositions, which I offer as pillars because each supports a different part of the argument and because, taken together, they describe a way of thinking about AI infrastructure finance that I believe will outlast the particular numbers of 2026. The outline with which this paper began proposed five; the analysis has added two.
Pillar 1 — GPUs Are Becoming Capital-Market Infrastructure
Computing systems are no longer merely technological equipment purchased by the companies that use them. Their ownership, their financing rights, and their expected revenue streams have become instruments that connect semiconductor manufacturing to global investment markets in a manner that would have been unrecognizable in 2022. When an insurance company’s general account holds a senior note secured by GPU pods in Memphis, when a public pension fund’s private credit allocation includes a tranche of a TPU lease to a model developer, when a sovereign wealth fund anchors the equity of a compute-infrastructure vehicle, and when the chip vendor itself participates in the capital structure of its customers’ demand, the GPU has become capital-market infrastructure in the same sense that the aircraft, the railcar, and the commercial building became capital-market infrastructure before it. The consequence is that the GPU’s fortunes are now bound up with the fortunes of the capital markets, and vice versa: a disappointment in AI returns is no longer a problem for technology companies alone, and a dislocation in private credit is no longer a problem that leaves the technology sector untouched. The $40 billion SpaceX–Apollo commitment, exceeding the sum of the lessee’s existing debt and lease liabilities in a single transaction, is the clearest illustration to date of how large that binding has become.[72]
Pillar 2 — Technological Life Is Not Collateral Life
A functioning GPU can remain productive for years after it has lost its premium market position, and the fleets of 2020-vintage A100s still earning rental revenue in 2026 prove it. But a lender’s recovery depends not on whether the equipment functions, nor even on whether it is commercially useful to its current operator, but on what a third party would pay for it, net of the costs of removal, transport, and re-homing, at the moment the lender needs to sell, in a market whose depth depends on how many other lenders are selling at the same time. Financial institutions must therefore evaluate four lives separately, technical, economic, accounting, and collateral recovery, and must resist the temptation to let the longest of them stand in for the shortest. The industry’s six-year accounting lives and the pessimists’ two-year replacement cycles are both answers to questions other than the one the lender is asking, and the collateral half-life metric proposed in this paper is an attempt to ask the right question directly. The Hopper price history, in which secondary values fell roughly 70 to 75 percent in eighteen months and then partially recovered, is the only complete observation of that metric the industry possesses, and it should be studied with the attention that aircraft financiers give to the residual value curves of a retiring airframe.[83]
Pillar 3 — Amortization Is a Defense Against Obsolescence
The relationship between the outstanding financial claim and the declining equipment value is more important than either curve considered independently, and it is the relationship that lease design can shape. A five-year straight-line amortization that returns the full financed cost through twenty equal quarterly installments, as the SpaceX–Apollo structure does, is a substantial defense against obsolescence, because by the time the equipment has lost most of its value the lessor has recovered most of its principal. But the defense is weakest in the first thirty months, when the equipment’s decay is steepest and the amortization has barely begun, and it is precisely in that window that the equipment-only shortfall of Table 2 arises. The analytical conclusion is that the shape of the amortization curve should be designed with reference to the shape of the collateral decay curve, that front-loaded amortization or supplemental early collateral is worth disproportionately more than the same protection applied late, and that a lender who evaluates a fully amortizing lease by its ending loan-to-value, which is zero by construction, has not evaluated the risk. Properly structured amortization reduces a lender’s reliance on terminal resale value; it does not eliminate the lender’s exposure to the interim path.
Pillar 4 — Institutional Diversification Does Not Eliminate Shared Risk
Multiple lenders financing multiple operators in multiple transactions can still share exposure to the same chip generations, the same hyperscaler and model-developer counterparties, the same power systems, the same refinancing window, and the same secondary market. The diversification that a private credit fund achieves by holding tranches of a dozen GPU financings is diversification across idiosyncratic risk, not across the common factors, and the common factors are what will move in the correlated impairment scenario. The early differentiation in credit spreads during the autumn of 2026, with CoreWeave’s bonds near 15 percent, Oracle at BBB-, and the strongest hyperscalers still trading inside the AA index, shows a market beginning to price idiosyncratic risk; it does not show a market that has priced the common factors, because the common factors have not yet been tested.[16] The stress test appropriate to this asset class is a joint default into a falling secondary market, and the capital held against GPU collateral should reflect that test rather than the single-name analysis that the structures’ apparent diversity invites. Damodaran’s warning about the private side of the boom applies with particular force to the holders furthest from the underlying asset.
“you’re going to get eaten alive if you go into that space”
— Aswath Damodaran, Professor of Finance, NYU Stern School of Business [44]
Pillar 5 — The Five-Layer AI Economy Ultimately Determines Collateral Value
Energy, chips, datacenters, models, and applications jointly determine whether computing equipment produces enough economic value to support its financing, and the recoverable value of a Layer 2 asset is a derived property of the health of all five layers rather than an intrinsic property of the silicon. A lender who underwrites the chip without underwriting the economy around it has underwritten a fraction of the risk. The Vera Rubin systems that SpaceX will lease in 2027 will be worth, in 2030, whatever the applications that run on the models that run on them are worth to the customers who pay for them, discounted by the cost of the power that feeds them and conditioned on the availability of the datacenters that house them. The disagreement among serious economists about the returns at Layers 4 and 5, which Section 4 documented, is therefore not a sideshow to the credit analysis; it is the credit analysis, and the appropriate response to an unresolved disagreement is not to pick a side but to price the uncertainty.
Pillar 6 — Energized Capacity Is the Hidden Covenant
The first of the two pillars the analysis has added concerns Layer 1 and Layer 3, and it is the lesson of 2026 that financial analysts were least prepared for. The binding constraint on deploying AI compute has shifted from the chip to the grid, and the shift transforms the collateral analysis in two ways. It means that financed equipment awaiting a grid connection is collateral of a fundamentally different character from financed equipment generating revenue, since it consumes interest while producing nothing, and that the installation and energization status of the equipment should be a disclosed and monitored term of every financing. And it means that the second economic life on which recovery values depend is conditional on the existence of vacant energized slots of the right specification at the moment of remarketing, which, in a market with 410 gigawatts queued in Texas alone and seven-year waits in Northern Virginia, cannot be assumed.[58] The most valuable protection a GPU lender can obtain may therefore be not a stronger claim on the equipment but a stable claim on the facility and the power contract that make the equipment useful, and the lessee’s demonstrated ability to energize capacity, as SpaceX has demonstrated at Colossus, is a credit strength of the first order. Energized capacity is, in effect, a covenant that no documentation yet names.
Pillar 7 — Vendor Support and Circular Capital Are Both Cushion and Contagion
The second added pillar concerns the role of the chip vendors in the financing of their own demand. Broadcom’s residual value backstop on the Anthropic TPU notes, NVIDIA’s anchor position in the Valor compute vehicle, and the indicated participation of NVIDIA, AMD, and Broadcom in Meta’s October 2026 equity offering all represent a structural innovation in which the party best positioned to remarket returned equipment, best informed about its future relevance, and most motivated to see the financing succeed assumes a portion of the collateral half-life risk.[78] As a cushion, this is powerful: it converts exposures that institutional investors could not otherwise hold into near-investment-grade paper, and it aligns the vendor’s product roadmap with the financed equipment’s continued value. As a channel of contagion, it is equally powerful: it means that a disappointment in AI demand would be transmitted simultaneously to the vendors’ revenue, to their residual value obligations, and to the equity stakes they hold in their customers, and that the vendors’ own creditworthiness, which underpins the senior tranches, would be weakening at the moment it was called upon. The investor Michael Burry’s characterization of the accounting that accompanies this circularity is contested, and the hyperscalers have substantive responses to it, but the concern it expresses is one that the official sector has now echoed in more measured terms.[36]
“one of the more common frauds of the modern era”
— Michael Burry, Scion Asset Management, on extended useful-life assumptions [81]
The analytical point does not depend on accepting Burry’s framing. It is that vendor support and circular capital flows reduce the measured risk of each individual financing while increasing the correlation among them, and that an investor evaluating the senior tranche of a vendor-supported GPU financing should ask not only whether the vendor can perform on its guarantee today but whether it could perform in the scenario in which the guarantee would be called.
Conclusion: The Contract That Outlives the Frontier
The October 26, 2026, SpaceX–Apollo announcement represents a consequential development in the financial architecture of artificial intelligence, and its significance extends well beyond the approximately $40 billion in anticipated principal payments or the deployment of NVIDIA’s Vera Rubin platform that it will finance. It reveals a new institutional challenge, which this paper has tried to name and to measure: the financing of rapidly evolving computational equipment through obligations whose contractual lives may extend well beyond the equipment’s period of technological leadership. A lease schedule written in 2027 for five to five and one-quarter years will still be amortizing in 2032, by which time the hardware it finances will be, if NVIDIA’s cadence holds, four or five architectural generations behind the frontier. The contract will outlive the frontier. The question is whether it will outlive the equipment’s value.
The relationship between technological improvement and financial risk is not predetermined, and this paper has resisted the temptation to resolve it by assertion in either direction. More capable accelerators reduce the relative attractiveness of earlier systems while simultaneously expanding the total market for computation; the same innovation that reprices the installed base enlarges the demand that the installed base can serve. Older GPUs may continue serving economically useful workloads, as the Hopper rebound of 2026 demonstrated, and properly structured amortization can reduce a lender’s reliance on terminal resale value, as the SpaceX–Apollo structure’s twenty-payment straight line does. The financial challenge is therefore not to predict whether technology will advance, which it will, but to understand when technological progress increases total demand, when it reduces equipment pricing power, and when those changes begin to threaten repayment. The collateral half-life framework, the four lives it distinguishes, the three scenarios it organizes, and the dashboard it proposes are offered as instruments for that understanding, not as answers to it.
This is why Collateral Half-Life fits the paper so well. The title captures an essential mismatch between two forms of time. Collateral represents the durability of financial obligations and recovery rights: the twenty quarterly payments, the lessor’s title, the lender’s lien, the covenant that survives the quarter in which it was written. Half-life represents the changing value of technological assets as innovation advances: the staircase of generational repricing, the plateaus of inference demand between the steps, the decay that is neither smooth nor predictable but that is, over any span of years, relentless. Their intersection establishes a distinctive framework for examining AI equipment financing, private credit, financial stability, and hyperscaler investment, and it is a framework that the official sector, in its 2026 assessments, has begun to adopt in substance if not in name.
By 2030, the competitive advantage of an AI infrastructure investor may depend as much on understanding semiconductor roadmaps, model architectures, inference pricing, and electricity constraints as on evaluating traditional interest rates and borrower credit ratings. The institutions that committed $40 billion to SpaceX in October 2026 have, whether they framed it this way or not, taken a position on all of those questions, and the position will be marked to market, in the only market that matters, over the five years that follow.
The Five-Layer AI Economy reveals why. Layer 1 must supply the energy, on a grid that is queuing 410 gigawatts in Texas and waiting seven years in Virginia. Layer 2 must deliver the accelerators, on an annual cadence that reprices each generation as the next arrives. Layer 3 must provide functioning datacenters, in a market where the shell is built in a year and the power takes three. Layer 4 must transform compute into useful intelligence, at a cost per token that falls by an order of magnitude with each hardware generation. Layer 5 must convert that intelligence into applications customers will pay to use, in enterprises where the task-level gains are visible and the enterprise-level returns are still being counted. The financial contract may be written against a machine, but its value ultimately depends on an entire economic system, and the machine is the only part of that system the lender can repossess.
The central lesson of Collateral Half-Life is that a GPU does not become a sustainable financing asset simply because it is powerful today. It becomes one when its expected earnings, its recoverable value, and its contractual obligations remain sufficiently aligned throughout its financial life, and when the institutions that hold the obligations have measured, disclosed, and priced the possibility that they will not. The SpaceX–Apollo transaction is the largest test of that proposition yet written. Its outcome will not be known until the last lease schedule amortizes to zero sometime in 2032, and by then the frontier it was written against will be a memory. That is the nature of the contract that outlives the frontier, and it is the reason the half-life of the collateral, rather than the brilliance of the chip, is the number that the next five years will be about.
Footnotes and Endnotes:
[1] U.S. Securities and Exchange Commission; Space Exploration Technologies Corp.. Form 8-K, Current Report, October 26, 2026 (binding commitment letter for master finance lease with Apollo-managed lessor). https://www.sec.gov/Archives/edgar/data/1181412/000110465926118614/tm2628247d1_8k.htm
[2] Morningstar (Dow Jones). “SpaceX Strikes $40B GPU Leasing Deal with Private Credit Players,” October 2026. https://www.morningstar.com/stocks/spacex-strikes-40b-gpu-leasing-deal-with-private-credit-players
[3] Investing.com, SEC Filings Desk. “SpaceX signs $40 billion GPU equipment lease agreement with Apollo,” October 26, 2026. https://www.investing.com/news/sec-filings/spacex-signs-40-billion-gpu-equipment-lease-agreement-with-apollo-93CH-4972481
[4] R&D World. “SpaceX commits to $40B NVIDIA GPU lease for terrestrial AI buildout” (including Fitch Ratings BBB+ commentary), October 2026. https://www.rdworldonline.com/spacex-commits-to-40b-nvidia-gpu-lease-for-terrestrial-ai-buildout/
[5] Apollo Global Management, Inc. (Investor Relations). “Apollo Backs $5.4 Billion Valor and xAI Data Center Compute Infrastructure Transaction with $3.5 Billion Capital Solution,” press release, January 7, 2026 (including statement of Antonio Gracias). https://ir.apollo.com/news-events/press-releases/detail/599/apollo-backs-5-4-billion-valor-and-xai-data-center-compute-infrastructure-transaction-with-3-5-billion-capital-solution
[6] Private Equity Wire. “Apollo and Blackstone plan $36bn private credit package to fund Anthropic’s Google TPU deal,” May 29, 2026. https://www.privateequitywire.co.uk/apollo-and-blackstone-plan-36bn-private-credit-package-to-fund-anthropics-google-tpu-deal/
[7] Quartz (citing Bloomberg and Reuters). “Apollo, Blackstone $36 billion debt deal for Anthropic’s Google chips,” May 29, 2026. https://qz.com/apollo-blackstone-36-billion-debt-deal-anthropic-google-chips-052926
[8] Yahoo Finance / Barchart. “SpaceX Reportedly Faces Higher Borrowing Costs on Its $40 Billion GPU Deal,” October 31, 2026. https://finance.yahoo.com/markets/stocks/articles/spacex-reportedly-faces-higher-borrowing-173003427.html
[9] Magica (reporting SpaceX Q3 2026 results and earnings call, October 20, 2026). “SpaceX Q4 spending forecast, AI compute, cash flow” (balance sheet of $89.9B cash and securities; $38.844B debt and finance-lease liabilities; Musk on power constraints). https://magica.com/news/spacex-q4-spending-forecast-ai-compute-cash-flow
[10] FourWeekMBA (The Business Engineer), from SpaceX Form 8-K Exhibit 99.1, October 20, 2026. “SpaceX Q3 Revenue Jumps 161% to $14.6B on AI and Starlink”. https://fourweekmba.com/ai-spacex-q3-revenue-jumps-161-to-14-6b-on-ai-and-starlink/
[11] Metir AI Research (citing Fortune). “SpaceX Q3 2026: The AI Segment Is Now Its Largest Business at $7.56B, and Capex Hit $19.6B,” October 21, 2026. https://www.metirai.com/blog/spacex-q3-2026-ai-unit-7-56-billion-earnings
[12] Engadget. “SpaceX’s first public earnings statement shows the financials of an AI company in 2026” (Anthropic $1.25B per month Colossus 1 arrangement). https://engadget.com/2230455/spacex-s-first-public-earnings-statement-shows-the-financials-of-an-ai-company-in-2026
[13] U.S. Securities and Exchange Commission, EDGAR. CoreWeave, Inc., Annual Report on Form 10-K for fiscal year 2025 and subsequent filings (six-year estimated useful life for technology equipment). https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&company=coreweave&type=10-K
[14] EBC Financial Group. “Why Did CoreWeave Stock Jump 16% Despite a $626M GAAP Loss?” (Q2 2026 results: $104.2B backlog, $35.1B debt), August 2026. https://www.ebc.com/forex/coreweave-stock-q2-2026-earnings-backlog-capex
[15] 24/7 Wall St.. “CoreWeave’s Debt Service Is the True Neocloud Litmus Test,” October 28, 2026. https://247wallst.com/investing/2026/10/28/coreweaves-debt-service-is-the-true-neocloud-litmus-test/
[16] InvestingLive (citing S&P Global Market Intelligence). “CoreWeave bonds trade like distressed debt as AI spending outpaces balance sheets,” October 2026. https://investinglive.com/stocks/coreweave-bonds-trade-like-distressed-debt-as-ai-spending-outpaces-balance-sheets/
[17] Board of Governors of the Federal Reserve System. Financial Stability Report, May 8, 2026. https://www.federalreserve.gov/publications/financial-stability-report.htm
[18] PYMNTS. “Fed Survey Finds Increased Worries About AI’s Impact on Financial Stability,” May 11, 2026. https://www.pymnts.com/cpi-posts/fed-survey-finds-increased-worries-about-ais-impact-on-financial-stability/
[19] Securities.io (reporting IMF Monetary and Capital Markets Department briefing). “IMF Global Financial Stability Report Flags AI, Yield, and Cyber Risks,” October 2026 (Global Financial Stability Report, October 2026, “Resilience, Risk, and Transformation”). https://www.securities.io/imf-global-financial-stability-report-flags-ai-yield-and-cyber-risks/
[20] Reuters (David Lawder and Andrea Shalal), via BusinessDay. “IMF warns of risks from AI market correction and war,” October 13, 2026. https://www.businessday.co.za/world/americas/2026-10-13-imf-warns-of-risks-from-ai-market-correction-and-war/
[21] EconoTimes. “Nasdaq 100 Drop Puts IMF Warning on AI Financing in Focus” (hyperscalers’ 10% share of global investment-grade issuance), October 2026. https://econotimes.com/Nasdaq-100-Drop-Puts-IMF-Warning-on-AI-Financing-in-Focus-1757692
[22] Universal Asset Owners. “UAO Daily Brief,” May 15, 2026 (Financial Stability Board, Report on Vulnerabilities in Private Credit, May 6, 2026; hyperscaler capex and Goldman Sachs capital estimates; pension allocations). https://www.universalassetowners.com/uao-daily-brief-2026-05-15/
[23] Fitch Ratings, via Times of Oman. “US credit outlook tied to AI investment confidence while consumer sectors face headwinds: Fitch” (Private Credit Default Rate record 6.0%, May 2026). https://timesofoman.com/article/174897-us-credit-outlook-tied-to-ai-investment-confidence-while-consumer-sectors-face-headwinds-fitch
[24] Tom’s Hardware. “Nvidia CEO confirms Vera Rubin NVL72 is now in production — Jensen Huang uses CES keynote to announce the milestone,” January 2026. https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-ceo-confirms-vera-rubin-nvl72-is-now-in-production-jensen-huang-uses-ces-keynote-to-announce-the-milestone
[25] The Decoder. “CES 2026: Nvidia promises five times the AI performance and ten times cheaper inference with Vera Rubin,” January 2026. https://the-decoder.com/ces-2026-nvidia-promises-five-times-the-ai-performance-and-ten-times-cheaper-inference-with-vera-rubin/
[26] Hashrate Index (Ian Philpot). “NVIDIA Vera Rubin NVL72 Specs Breakdown,” GTC 2026 coverage. https://hashrateindex.com/blog/nvidia-vera-rubin-nvl72-specs-breakdown/
[27] Tiger Brokers Deep News. “NVIDIA Unveils Rubin Chip Platform with Fivefold Compute Power Surge, Targets Trillion-Dollar Market” (Rubin specifications; Rubin Ultra 2H 2027), March 17, 2026. https://www.itiger.com/news/1106450919
[28] Advanced Micro Devices, Inc. (Investor Relations). “AMD and Meta Advance Strategic Partnership to Power Meta’s Superintelligence Infrastructure,” press release, October 27, 2026. https://ir.amd.com/news-events/press-releases/detail/1302/amd-and-meta-advance-strategic-partnership-to-power-meta-superintelligence-infrastructure
[29] HPCwire / AIwire. “AMD Details Meta AI Infrastructure Collaboration, Plans $1B Investment,” October 27, 2026. https://www.hpcwire.com/aiwire/2026/10/27/amd-details-meta-ai-infrastructure-collaboration-plans-1b-investment/
[30] Advanced Micro Devices, Inc.. Webcast transcript, AMD–Meta announcement, February 24, 2026 (remarks of Dr. Lisa Su). https://d1io3yog0oux5.cloudfront.net/_cead7baaa2cb7f60c17b342e85e15883/amd/db/817/9229/webcast_transcript/AMD_260224+Meta+Announcement.pdf
[31] DatacenterDynamics. “Meta and AMD sign multi-year agreement for 6GW GPU deployment,” February 2026. https://www.datacenterdynamics.com/en/news/meta-and-amd-sign-mutli-year-agreement-for-6gw-gpu-deployment/
[32] Harvard Business School. “Meta: Accounting for AI Data Center Depreciation,” Case 126-034, June 2026. https://store.hbr.org/product/meta-accounting-for-ai-data-center-depreciation/126034
[33] Bloomberg Tax. “Meta’s Accounting Move on AI Servers to Boost Profit This Year” (useful life extended to 5.5 years, January 29, 2025). https://news.bloombergtax.com/financial-accounting/metas-accounting-move-on-ai-servers-to-boost-profit-this-year
[34] Altss Research. “Amazon and Meta’s server-life estimates moved earnings in opposite directions,” updated October 18, 2026. https://altss.com/blog/amazon-meta-server-depreciation-useful-lives-earnings
[35] The National Law Review. “Deep Quarry: Useful Lives of GPUs — Key Considerations,” 2026. https://natlawreview.com/article/deep-quarry-useful-lives-gpus-key-considerations
[36] ValueTheMarkets. “The GPU Depreciation Debate Behind This Week’s AI Dip” (Michael Burry $176 billion estimate), September 16, 2026. https://www.valuethemarkets.com/analysis/gpu-depreciation-debate-behind-this-weeks-ai-dip
[37] Stanley Laman Group, “Signals and Noise”. “Why GPU Useful Life Is the Most Misunderstood Variable in AI Economics” (Satya Nadella remarks; component-based depreciation proposal). https://www.stanleylaman.com/signals-and-noise/gpus-how-long-do-they-really-last
[38] Data Gravity. “How long does a GPU last?” (Microsoft CFO Amy Hood remarks; JPMorgan Asset Management useful-life sensitivity), 2026. https://www.datagravity.dev/p/how-long-does-a-gpu-last
[39] theCUBE Research (Dave Vellante). “Breaking Analysis: Resetting GPU Depreciation — Why AI Factories Bend but Don’t Break Useful-Life Assumptions,” 2025. https://thecuberesearch.com/298-breaking-analysis-resetting-gpu-depreciation-why-ai-factories-bend-but-dont-break-useful-life-assumptions/
[40] WhiteFiber (citing Princeton Center for Information Technology Policy analysis). “Understanding GPU Lifecycle: Frankly, It’s Complicated,” 2026. https://www.whitefiber.com/blog/understanding-gpu-lifecycle
[41] Aswath Damodaran, NYU Stern School of Business, Musings on Markets. “Stock Prices, Earnings and Cashflows: The AI Effect plays through!” October 8, 2026. https://aswathdamodaran.blogspot.com/2026/10/stock-prices-earnings-and-cashflows-ai.html
[42] Aswath Damodaran, NYU Stern School of Business. “The Good, the Bad and the Ugly” — presentation slides accompanying the October 2026 analysis of markets and AI. https://pages.stern.nyu.edu/~adamodar/pdfiles/blog/MktandAI2026.pdf
[43] OfficeChai (reporting remarks of Aswath Damodaran). “History Suggests There Would Be an AI Bust, and It Would Be More Painful Than the Dot-Com Bust: Aswath Damodaran,” 2026. https://officechai.com/ai/history-suggests-there-would-be-an-ai-bust-and-it-would-be-more-painful-than-the-dot-com-bust-ashwath-damodaran/
[44] Finviz / Benzinga (reporting remarks of Aswath Damodaran). “Top valuation expert says AI market needs trillions in revenue to justify valuations after cashing out of its biggest chipmaker,” 2026. https://finviz.com/news/287849/top-valuation-expert-says-ai-market-needs-trillions-in-revenue-to-justify-valuations-after-cashing-out-of-its-biggest-chipmaker
[45] AOL / Bloomberg (interview with Daron Acemoglu). “‘A lot of money is going to get wasted’: MIT economist sounds the alarm on AI hype”. https://www.aol.com/lot-money-going-wasted-mit-165943061.html
[46] MIT Shaping the Future of Work Initiative. “BBC: Could the tech bubble burst?” — Daron Acemoglu interview, BBC World Business Report, February 5, 2026. https://shapingwork.mit.edu/?p=7316
[47] Federal Reserve Bank of Richmond, Econ Focus. “The Investment Boom” (Q1–Q2 2026 feature; remarks of Erik Brynjolfsson, Stanford University). https://www.richmondfed.org/publications/research/econ_focus/2026/q1-q2_feature2
[48] The Decoder (reporting Erik Brynjolfsson, Financial Times op-ed, February 15, 2026). “Stanford’s Brynjolfsson sees AI boosting US productivity, but he also co-founded an AI consulting firm”. https://the-decoder.com/stanfords-brynjolfsson-sees-ai-boosting-us-productivity-but-he-also-co-founded-an-ai-consulting-firm/
[49] Stanford Institute for Human-Centered Artificial Intelligence (HAI). AI Index Report 2026 and “Inside the AI Index: 12 Takeaways from the 2026 Report,” April 13, 2026. https://hai.stanford.edu/news/infographic-age-artificial-intelligence
[50] Spheron Network. “NVIDIA H100 News 2026: Price Changes, H200 and Successor Updates, and Cloud Availability” (secondary-market H100 pricing). https://www.spheron.network/blog/nvidia-h100-news-2026/
[51] IntuitionLabs (Adrien Laurent). “Data Center GPU Prices: Costs, Configs and Price History,” 2026 (including Reuters/FT report on China gray-market pricing). https://intuitionlabs.ai/articles/data-center-gpu-prices
[52] Crypto Briefing. “Nvidia’s H100 GPU gains about 30% in value as AI compute demand holds firm,” October 2026. https://cryptobriefing.com/nvidia-h100-gpu-value-rises-30-percent/
[53] Latent Space (Eugene Cheah). “$2 H100s: How the GPU Rental Bubble Burst,” 2024. https://www.latent.space/cp/150084975
[54] Morrison Foerster. “Managing Export Control Risks in the AI Chip Ecosystem,” February 9, 2026. https://www.mofo.com/resources/insights/260209-managing-export-control-risks-in-the-ai-chip-ecosystem
[55] JD Supra. “Trump Administration Takes Several Actions on AI Chip Exports” (BIS final rule, January 15, 2026). https://www.jdsupra.com/legalnews/trump-administration-takes-several-7320644/
[56] Spheron Network (citing Tom’s Hardware and Digitimes; remarks of Jensen Huang). “GPU Export Controls 2026: What It Means for Cloud Pricing”. https://www.spheron.network/blog/gpu-export-controls-2026-cloud-pricing/
[57] LetsDataScience (citing Reuters, May 31, 2026, and CNBC). “U.S. Tightens Controls on Nvidia AI Chip Exports” (BIS guidance on China-headquartered affiliates). https://letsdatascience.com/news/us-tightens-controls-on-nvidia-ai-chip-exports-97a95018
[58] Inflect. “Data Center Power Shortage 2026: Why Grid Capacity Is Now the Bigger Constraint Than GPUs” (ERCOT ~410 GW large-load queue, April 2026). https://inflect.com/blog/data-center-power-shortage-2026-why-grid-capacity-is-now-the-bigger-constraint-than-gpus
[59] Spencer Ogden (citing JLL Global Data Center Outlook 2026). “Why Are Data Centers Waiting Years for a Grid Connection?” (transformer lead times; Arizona Public Service queue; ERCOT June 2026 batch review). https://www.spencer-ogden.com/insights/why-are-data-centers-waiting-years-for-a-grid-connection
[60] Introl. “PJM Grid Crisis: 6GW Shortfall Threatens 40% of AI Data Centers by 2027” (PJM capacity auction; Gartner projection; transformer lead times). https://introl.com/blog/pjm-grid-crisis-6gw-shortfall-data-center-power-2027
[61] SoftwareSeni. “Why US AI Data Centre Delays and Cancellations Persist Despite Record Investment,” September 10, 2026 (PJM interconnection studies averaging 40 months). https://www.softwareseni.com/why-so-much-us-ai-data-centre-capacity-is-delayed-or-cancelled/
[62] The White House, Executive Office of the President. “Winning the Race: America’s AI Action Plan,” July 23, 2025. https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-Plan.pdf
[63] Arnold & Porter, Environmental Edge. “Data Centers and Permitting Reform: Is Artificial Intelligence Up to the Challenge?” July 2025 (quoting America’s AI Action Plan). https://www.arnoldporter.com/en/perspectives/blogs/environmental-edge/2025/07/data-centers-and-permitting-reform
[64] Crypto Briefing. “Credit risk for tech giants hits fresh highs as AI debt piles up,” October 20, 2026. https://cryptobriefing.com/tech-credit-risk-hits-highs-ai-debt/
[65] KuCoin News / Crypto Briefing. “Meta and Oracle Face Bond Market Scrutiny Amid AI Spending Surge” (S&P downgrade of Oracle to BBB-, July 2026). https://www.kucoin.com/news/flash/meta-and-oracle-face-bond-market-scrutiny-amid-ai-spending-surge
[66] Luminix Strategic Research (Jon Sinclair). “AI Debt: Hyperscaler Borrowing and Credit Risk 2026,” October 2026 (hyperscaler investment-grade issuance; S&P $7 trillion capex estimate; CoreWeave and Nebius leverage). https://www.useluminix.com/reports/industry-analysis/when-will-the-ai-bubble-burst/source/2
[67] ROIC.ai. “xAI nears deal with Apollo for chips” (JPMorgan data center securitization projection; Apollo secondary-market ambitions), February 2026. https://www.roic.ai/news/xai-nears-3-billion-deal-with-apollo-for-chips-02-09-2026
[68] City A.M. (citing Apollo Global Management forecasts). “Big Tech’s AI borrowing boom tests US bond market,” 2026. https://www.cityam.com/?p=2395318
[69] Startup Fortune. “Christine Lagarde puts AI’s debt-fueled buildout on the financial stability watchlist” (ECB Financial Stability Review, May 2026; FSB private credit report), June 17, 2026. https://startupfortune.com/?p=13968
[70] FourWeekMBA (The Business Engineer). “IMF Holds 2026 Growth at 3.1% as AI Investment Props It Up” (IMF World Economic Outlook, October 2026, Box 1.2; GFSR “Rising Leverage in the AI Ecosystem”). https://fourweekmba.com/ai-imf-holds-2026-growth-at-3-1-as-ai-investment-props-it-up/
[71] Tech Times. “SpaceX Signs Record $40B Apollo GPU Financing for Nvidia Vera Rubin Chips, Topping Anthropic’s Deal,” October 27, 2026. https://www.techtimes.com/articles/329522/20261027/spacex-signs-record-40b-apollo-gpu-financing-nvidia-vera-rubin-chips-topping-anthropics-deal.htm
[72] Iolite. “SpaceX Inks $40B GPU Lease Deal with Apollo to Deploy Next-Gen Nvidia Chips” (principal relative to existing liabilities; Financial Times October 6 report), October 2026. https://iolite.net/en/news/spacex-apollo-40-billion-gpu-lease-nvidia
[73] Barchart (Apollo Global Management press release distribution). “Apollo Backs $5.4 Billion Valor and xAI Data Center Compute Infrastructure Transaction” (statement of Christopher Lahoud, Partner, Apollo), January 7, 2026. https://www.barchart.com/story/news/36924940/apollo-backs-5-4-billion-valor-and-xai-data-center-compute-infrastructure-transaction-with-3-5-billion-capital-solution
[74] Larridin Data. “Apollo Global Management AI Adoption Tracker” (Athene AUM; Apollo infrastructure deployment), updated April 30, 2026. https://data.larridin.com/data/ai-tracker/company/apollo-global-management/
[75] Debt Serious (Substack). “Round #46: Apollo + Blackstone: GPU-Backed Loans” (tranche structure of the Anthropic TPU financing; Broadcom guaranty mechanics), 2026. https://debtserious.substack.com/p/round-46-apollo-blackstone-gpu-backed
[76] Luminix Strategic Research. “GPU Depreciation and Useful Life at Hyperscalers,” updated October 2026 (five durations collapsed into “useful life”; 2026 server cohort depreciation sensitivity). https://www.useluminix.com/reports/industry-analysis/when-will-the-ai-bubble-burst/source/1
[77] Yahoo Finance. “SpaceX Secures Record $40 Billion GPU Lease Deal With Apollo” (Apollo statement on Vera Rubin, Grok, enterprise and cloud customers), October 2026. https://finance.yahoo.com/technology/ai/articles/spacex-secures-record-40-billion-190201846.html
[78] FourWeekMBA (The Business Engineer). “AMD Will Invest $1B in Meta’s Stock as Helios Runs at Meta” (AMD press release of October 27, 2026; NVIDIA press release of October 26, 2026; Meta offering launch release; remarks of Dr. Lisa Su). https://fourweekmba.com/ai-amd-will-invest-1b-in-metas-stock-as-helios-runs-at-meta/
[79] Axios. “The growing jitters over hyperscaler debt,” July 27, 2026. https://axios.com/2026/07/27/debt-data-center-oracle-goole
[80] Advisor Perspectives. “Is There Really Carnage in Hyperscaler Credit?” August 3, 2026. https://www.advisorperspectives.com/articles/2026/08/03/there-carnage-hyperscaler-credit
[81] Tarry Singh, Dispatches. “The capex is real. The NPV is fiction.” June 2, 2026 (quoting Michael Burry on extended useful-life assumptions). https://tarrysingh.com/blog/the-capex-is-real-the-npv-is-fiction
[82] Hashrate Index. “Used GPU Market Pricing: Depreciation and the Secondary AI Hardware Market,” 2026. https://hashrateindex.com/blog/used-gpu-market-pricing-deprecation-secondary-ai/
[83] Lapaas Voice. “The H100 Price Crash: February 2026 Data” (secondary-market decline from mid-2024 peak), February 23, 2026. https://voice.lapaas.com/?p=17179
[84] Luminix Strategic Research. “Understanding Dylan Patel of SemiAnalysis: Deep Dive on AI Compute Scaling Bottlenecks” (H100 resale “mid-life cliff”; Blackwell capex and performance ratios), March 2026 update. https://www.useluminix.com/reports/industry-analysis/understanding-dylan-patel-of-semianalyis-deep-dive-on-ai-compute-scaling-bottlenecks/source/0
[85] Pulse. “Nvidia Moves Vera Rubin AI Platform to Volume Production, Boosting Big Data Compute” (Computex 2026; 350+ partners in 30 countries), June 1, 2026. https://www.pulse.bot/big-data/news/nvidia-moves-vera-rubin-ai-platform-to-volume-production-boosting-big-data-compute-fac2de4b-421a-489d-80e8-075827598238/
[86] AI2.work. “The AI Productivity Paradox: Record Spending, Sluggish Growth” (BLS productivity data; PwC 29th Global CEO Survey; EY US AI Pulse Survey), May 2026. https://ai2.work/blog/the-ai-productivity-paradox-record-spending-sluggish-growth
