Rental indices, secondary venues and syndicated credit already exist for GPU compute. What lenders don't have is a residual-value history to underwrite against.
On August 10, NVIDIA announced memorandums of understanding with six global investment firms (Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR) that "aim to establish the first compute financing platforms": dedicated pools of capital to finance NVIDIA-based AI factories for frontier labs, enterprises and AI clouds. The release's verb matters: the platforms aim to "mobilize over $500 billion of third-party capital for the buildout of AI infrastructure over time." That is an announced mobilization target, not committed capital. Jensen Huang makes the larger claim in a companion essay on NVIDIA's blog: AI factory compute, in the title's words, "is becoming an investable asset class."
Much of a compute-finance market already exists. Rental price indices publish daily. Used GPUs trade through brokers and, since this summer, on dedicated electronic venues. CoreWeave has taken contract-backed GPU debt through public syndication repeatedly. What the market does not have is the piece long-duration lenders lean on hardest: a standardized, independently observable record of what the hardware is worth across product generations and demand cycles. Aircraft finance needed decades of appraisal conventions and realized sales to build its residual-value curves. The August 10 announcement asks institutional capital to move at $500 billion scale before the equivalent history exists.
The release is specific about intent and silent on structure. The MOUs are "subject to execution of the final agreements." The August 10 release discloses no per-partner commitments, fund sizes, legal structures, closing dates or first transactions, and does not define what counts toward the $500 billion. The most concrete statement of purpose came from Goldman Sachs: "excited for the new opportunity to create a market for credit backed by NVIDIA compute."
The announcement landed into a conversation already underway, an extension of the private-capital turn SCN traced in July. Two days earlier on the All-In podcast, host David Sacks sketched xAI's compute growing "from two to eight" gigawatts (an illustration; Elon Musk's guidance was roughly 2 GW by year-end), and guest Brad Gerstner, founder of Altimeter Capital and a large, publicly bullish NVIDIA shareholder, put the cost at "$50 billion per gigawatt to build minimum." His list: "you either have to go into the market and borrow the money, or you have to do a dilutive equity raise... or you get NVIDIA to backstop it," plus off-balance-sheet SPVs.
Huang's essay engages the skepticism around vendor-supported demand under a section he titled "Is this circular financing?" His answer: the demand is real, coming from "frontier AI labs, AI-native startups, enterprises, cloud providers and countries building AI services," and the capital providers "independently underwrite each project — including the customer, demand, utilization, cash flow and residual value."
That last noun is where this story's question lives.
"AI compute" names four different assets: GPU boards, configured servers and racks, rented capacity under contract, and factory-scale projects with land, power and leases attached. Each is different collateral with different lenders. A creditor on a 20-year powered-shell lease to an investment-grade tenant is underwriting the tenant; a lender to a GPU-server SPV is underwriting the hardware.
Cutting across all four is a distinction the depreciation debate keeps blurring: service life (does the hardware run), economic life (does deployed compute keep earning), and collateral life (what a creditor recovers from the pledged asset and contracts if things go wrong). A financing platform's risk sits mostly in the third category, and the third is the one without a history.
Huang's asset-class case rests on four characteristics: AI factory compute "produces revenue, serves a broad market, improves in performance over time, and can be redeployed." The essay backs the revenue leg with rental pricing: one-year H100 contracts rising from $1.70 per GPU-hour in October 2025 to about $2.35 in March 2026, cross-provider on-demand medians from $2.00 to $2.70 by June, B200s at $5.30 to $7.05. The essay names no sources for its figures, though the one-year H100 arc matches SemiAnalysis's index almost exactly.
The "improves over time" leg runs through software. CUDA updates raise the output of installed hardware, which NVIDIA argues extends economic life, the same dynamic SCN examined earlier this week in the CUDA moat's collision with AMD's composability push. The A100, introduced in 2020, "remains in active commercial use" six years later, per the essay, which describes its economic life as extending "toward a decade." That is NVIDIA's argument to make; lenders will want measures they can obtain independently.
The price-discovery layer is real and young. SemiAnalysis has maintained rental indices for clients since 2023; its one-year H100 index went public in April 2026. Silicon Data has published daily rental indices for A100, H100, B200 and other SKUs since 2025, pitched at financial institutions doing "asset valuation and depreciation modeling." SemiAnalysis wrote after the announcement that lenders entering this market "will need a good GPU rental price index, a way to forecast GPU rental prices into the future, and a model for estimating and tracking GPU residual value over time". Those are products it sells. The data itself replicates across trackers; the plumbing is simply new, and the governance questions lenders ask of any young benchmark remain open: which transactions enter the index, whether quotes are executable or indicative, how conflicts are managed, and how the series behaves in a downturn none of them has yet recorded.
The secondary market is older than it looks and less standardized than lenders need. Brokered resale through ITAD vendors and enterprise resellers has run for years; dedicated electronic venues arrived this summer: Compute Exchange opened a used-H100 and A100 market in July, and Stoa Markets launched an RFQ marketplace in August, claiming $300 million in first-month RFQs (its own figure). Still missing is a standardized mark: ask what a used A100 is worth and the answer depends on where you ask. Jarvis Labs puts used 80GB SXM boards at $4,000 to $9,000, while Hashrate Index's survey of visible commercial listings runs from about $7,800 for refurbished 40GB units to about $18,900 for an 80GB PCIe card (both accessed August 12). The ranges cover different SKUs, conditions, and channels; they barely overlap. Hashrate expects a further 10 to 15 percent decline through 2026 as buyers upgrade to Blackwell.
The depreciation fight sits on top of that dispersion. Michael Burry's running critique is a book-value argument: hyperscalers depreciate servers and networking equipment over five to six years while annual product cadence implies a shorter earning life, understating depreciation by his estimate of roughly $176 billion across 2026-2028. NVIDIA's is an economic-life argument: A100s still earn, with cloud pricing stable at roughly $1.49 to $3.43 per hour across providers (Jarvis Labs, accessed August 12). Both can be true at once. Neither tells a lender what a five-year-old fleet fetches at foreclosure, and no aircraft-style public record of realized disposals across generations and demand cycles exists to settle it.
Which brings us to the essay's most consequential sentence: "In some cases, NVIDIA may provide a residual-value support mechanism for up to 25% of an opportunity, assessed carefully on a project-by-project basis." Huang describes the support as "limited, residual-value based and designed to complement — not replace — independent underwriting."
One reading, and it is an interpretation rather than anything NVIDIA has said: the company is offering its balance sheet as a bridge over the gap where a residual-value record should be, until the market can price the risk itself. How the support would work is undisclosed: no trigger conditions, no seniority, no word on whether it is a guarantee or how NVIDIA would account for it.
There is also a scope question. The support is capped per opportunity within the platforms, and NVIDIA frames its risk share there as limited. Its documented support for customer financing extends beyond them: an order form obligating NVIDIA to purchase CoreWeave's residual unsold cloud capacity through April 2032, with an initial value of $6.3 billion; reported equity of up to $2 billion in a $20 billion SPV that buys NVIDIA GPUs and leases them to xAI; and reported talks over guarantees of up to $250 billion for an OpenAI data-center project in Ohio (the latter two unconfirmed by the companies). Neither the August 10 release nor the essay discloses an aggregate exposure figure; the platform cap and the company-wide perimeter are different things.
The most direct evidence on collateral life comes from CoreWeave. On the same day as NVIDIA's announcement, it closed its $2.6 billion DDTL 5.5 facility, contract-backed like its predecessors but with one difference: the loan runs about five years while the underlying customer contracts average about three. CoreWeave says the structure shows lenders "signaling confidence in the long-term value of NVIDIA GPUs" and "a willingness to underwrite renewal risk." Renewal risk is residual-value risk by another name, taken in increments.
Bonds for a CoreWeave-leased Applied Digital data center priced at 7 percent in June against 10 percent for the same project last November, per Bloomberg data, and developers have raised more than $8 billion in high-yield debt for CoreWeave-leased projects. That credit secures contracted cash flows from data-center projects and customer agreements rather than bare GPU boards; the layer closest to the hardware is still waiting for its curve.
Contracted cash flow is the part lenders already know how to underwrite. The same week as the announcement, Riot Platforms signed a 20-year, 191 MW lease with an unnamed frontier lab: roughly $9.1 billion in initial contract revenue, up to $16.1 billion with extensions, with the first 96 IT MW expected in December 2027 on the company's existing approved interconnection. Decade-scale offtake is the duration logic behind the long-dated power contracts SCN has called the AI power moat and Anthropic's multi-gigawatt TPU reservation, coming online from 2027. With much of the near-term HBM supply committed under long-term agreements and power contracted years out, financing has joined them as a gating constraint on the largest supercomputing buildout in history.
Whether $500 billion mobilizes, and at the "attractive rates" the release promises, depends less on demand than on how fast the residual-value record fills in. Indices need a downturn behind them, and venues need volume; a residual curve needs realized disposals, and the industry has barely begun disposing. Until then, the bridge on offer is NVIDIA's balance sheet: in some cases, up to 25 percent of an opportunity at a time. The MOUs set the target; the final agreements, when they come, will show the structure.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
