Compute as an Asset Class: Inside NVIDIA's $500 Billion Compute Capital Stack

Two proprietary metrics run through every section below: Compute Yield, a cap-rate-style measure of what a compute asset actually earns against what it cost to install, and Compute Duration, the years of economically useful revenue-generating life before obsolescence — not physical failure — erodes an asset's earning power. Both exist to answer one question, repeatedly, from different angles: can revenue durability outrun hardware obsolescence?

On August 10, 2026, NVIDIA signed memoranda of understanding with six of the world's largest financial institutions — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR — to establish independent compute financing platforms intended to mobilize over $500 billion of third-party capital for the AI infrastructure buildout over time. NVIDIA's own release frames the goal explicitly: turn NVIDIA compute and full-stack AI infrastructure into an investable asset class for global capital.

Jensen Huang's language is the cornerstone citation here, not a caveat that needs hedging — NVIDIA wrote the framing into its own headline. Huang described the company's shift from building chips toward helping create "a new class of productive, investable infrastructure: AI factories," and, addressing circular-financing concerns in a separate essay, argued that NVIDIA compute is well-suited to underwriting because it is broadly adopted, fungible across customers, and continuously improved through software — extending its useful life.

On NVIDIA's own exposure, Huang disclosed — case by case, and not yet as a contractual term — that NVIDIA may provide residual-value support for up to roughly 25% of an opportunity. The mechanism matters more than the number: it covers up to a quarter of the shortfall if a chip's resale or redeployment value falls short of expectations at the end of a financing term. It is a partial cap on someone else's loss, not a transfer of the risk to NVIDIA — the credit assessment of customer, use case, cash flows and residual value stays with the capital providers. Huang has described the level as lower than in other compute financing arrangements. Separately, NVIDIA is reported to be in talks to provide a roughly $250 billion financing backstop tied to a 10-gigawatt OpenAI data center project in Ohio — an economically analogous bilateral structure, at a scale that would dwarf anything disclosed inside the six-firm platform. We treat this as reported and unconfirmed until contracts are public; see §09.

We deliberately draw no conclusion from the day's share-price moves, and readers should be sceptical of anyone who does: NVIDIA itself fell more than 3% on the announcement, alongside Amazon, Microsoft and Alphabet. Any number of unrelated drivers — competing in-house silicon, China exposure, model-efficiency concerns — were live in the same session. A structural thesis about financing architecture is not testable on a single day's tape, which is why §10 deliberately excludes share prices from its falsification criteria.

Strip away the press-release framing and the mechanism is a fairly conventional infrastructure-finance stack wrapped around an unconventional asset. Long-duration capital — pension, insurance and sovereign-wealth money — sits at the top, several steps removed from the hardware. It reaches the compute layer through alternative asset managers who structure credit and equity tranches, into a financing vehicle that actually owns the GPUs, and down to the operator that leases them to the AI lab, enterprise or sovereign customer that ultimately generates the cash flow.

NVIDIA's platform is the largest instance of this mechanism, not the first. Three already-live precedents, running from January 2026, show what the same idea looks like at a fraction of the scale — and the first of them is the cleanest expression of the thesis on record.

Read together, the sequence is legible, and it starts long before August. In January, institutional capital takes outright ownership of a GPU fleet and leases it to the operator under a triple net lease — the textbook form of the thesis, with NVIDIA sitting in the equity as an anchor LP rather than merely selling the hardware. In May, a vertically integrated hyperscaler opens its own chip program to third-party capital. In June, a chip vendor secures a $35 billion capital solution against its platform, anchored by two of the same firms that reappear in NVIDIA's August announcement. Only then does the largest GPU supplier attempt to standardize the structure — and it does so with a mobilization target, not a transaction.

NVIDIA did not invent compute-as-collateral. It is attempting to industrialize a structure that Apollo had already underwritten, priced and closed.

That chronology matters for how much weight the thesis can bear. A single announcement can be dismissed as promotion. Three financing precedents and one announced platform across eight months — with billions of dollars of committed capital, and at least one already operating through a closed GPU-lease structure — describe a market forming. We are deliberately careful with that phrasing: only the January Apollo/Valor transaction is documented as a completed acquisition and lease. The Broadcom platform is described by Apollo as committed capital across a multi-year draw schedule, which is not the same as capital already drawn, and NVIDIA's $500 billion is a target rather than a transaction of any kind. Comparing the two figures directly would be a category error.

Real estate and infrastructure investors underwrite assets against a cap rate: net operating income divided by asset cost. Compute has no established equivalent, so we propose one — built the same way, in the same units. The discipline that makes it useful is keeping it strictly unlevered.

Compute Yield is a pre-financing, asset-level measure. No interest cost, no amortization, no residual value enters it. That is what makes it comparable to a data-center cap rate or an aircraft-leasing yield: it isolates what the asset earns from how it was funded. It creates a common percentage framework for comparing the operating economics of compute against conventional infrastructure — it is explicitly not a substitute for debt yield, equity IRR or credit-spread analysis. A 20% unlevered operating yield and a 20% private-credit yield are not the same economic object: the latter embeds seniority, leverage, a contractual coupon, maturity, expected loss and recovery.

Everything the yield deliberately excludes belongs one layer down:

Keeping the two apart resolves a confusion that runs through most published commentary on these deals, where an operating yield, a lender's coupon and an equity return get quoted as if they were the same number. They answer different questions: Compute Yield asks whether the asset works; Compute Asset IRR asks whether the trade works.

A second distinction does real work in this market. Contracted Compute Yield uses minimum contracted revenue — take-or-pay, reserved capacity, fixed lease obligations. Realized Compute Yield uses revenue actually earned at actual utilization. The gap between them is precisely what makes these assets financeable: under a triple net lease of the kind Apollo used in the xAI transaction (§03), physical utilization can fall without contractual revenue falling proportionally, because the lessee — not the asset owner — carries that exposure. Where revenue is usage-linked rather than contracted, the two converge and the financing looks far more like equipment finance than infrastructure debt.

These are A.L. Capital Advisory analytical constructs, not industry-standard disclosure metrics — no financing platform, hyperscaler or chip vendor currently reports either figure. But the inputs are no longer entirely dark. Per SemiAnalysis's H100 1-Year Rental Price Index — built from monthly surveys of more than 100 market participants and validated against transaction data — one-year H100 contract rates rose roughly 40%, from about $1.70 per GPU-hour in October 2025 to $2.35 by March 2026.

Two things about that number matter more than its level. First, these are negotiated committed-term contract rates, not posted on-demand prices — which is precisely the revenue line that feeds Contracted Compute Yield, and it is not the series most public GPU price trackers publish. Second, the market has split: abundant, cheap spot capacity coexists with tight and rising committed pricing. That divergence is the empirical fingerprint of the contracted-versus-realized distinction above, and it cuts against the simplest oversupply bear case. What remains undisclosed is the other half of the equation: realized utilization and residual value, deal by deal. The framework's value is in what it forces an analyst to ask for.

Applied across the three structures this page covers, Compute Yield reads differently even before a single number is filled in. Under NVIDIA's platform — and, already, under the Apollo/Valor lease — the yield an outside financier underwrites is explicitly a third-party return, priced against a contracted lease and whatever residual-value support is negotiated. Under Oracle's model, there is no separate yield line for an outside investor to underwrite in the same way; the economics run through Oracle's own corporate credit and margins. Under Google's TPU program, the concept barely applies in third-party form at all — Google owns the asset outright and captures, or absorbs, the full economics itself (§08). That three-way contrast — externally underwritten, corporate-balance-sheet-underwritten, and internally owned with no external yield — is the more useful output of the framework than any single number would be.

Compute Duration is the years of economically useful, revenue-generating life a compute asset has before technological obsolescence — not physical hardware failure — materially erodes its competitive earning power or resale value. It sits apart from, and below, four related but distinct numbers: physical useful life (the hardware still runs), accounting depreciation life (4–6 years, the schedule most hyperscalers currently use), financing maturity (the term of the debt raised against the asset), and architecture refresh cadence (NVIDIA ships a new GPU generation roughly every 12–24 months).

The worked example that matters most isn't hypothetical — it has already happened, and it is more specific than the general narrative suggests. In March 2026, Oracle, Crusoe and OpenAI cancelled a planned ~600MW expansion of the flagship Abilene, Texas Stargate campus, which would have taken it from roughly 1.2GW to 2.0GW. The reported reason is precisely a Compute Duration problem: Abilene is built around Blackwell-generation hardware whose power wasn't projected to come online for about a year, by which point OpenAI expected to be deploying Vera Rubin silicon in larger clusters elsewhere. OpenAI's compute lead has said publicly the company chose to place that capacity at other sites.

Three qualifications keep this honest. The broader 4.5GW Oracle–OpenAI commitment remains on track, and Oracle publicly called reports of a deteriorating relationship "false and incorrect," stating that leasing across that commitment was completed. Microsoft subsequently leased 900MW at the same site from Crusoe. And NVIDIA itself paid Crusoe a reported $150 million deposit to secure capacity there — an intervention worth noting, because a supplier paying cash to control where its installed base ends up is an early, informal version of the residual-value management this entire financing structure depends on.

The practical implication for a lender is now measurable rather than assumed. We previously described market financing maturities in general terms; the Saline Township campus gives a disclosed one. That deal placed 19-year money, maturing in 2045, at a 7.5% coupon with six years of interest-only before amortization begins, against a compute asset class whose economic life is argued at anywhere from two to ten years. The interest-only period alone is longer than Burry's entire estimate of the hardware's economic life. That mismatch doesn't automatically make the debt unsound — the borrower is a data-center campus with a long-lived shell and power infrastructure, not a rack of GPUs, and a lease can contract revenue well beyond the point specific hardware becomes uncompetitive. But it does locate the question precisely: how much of the collateral value in these structures is the building and the interconnect, and how much is the silicon? The longer the tenor, the more the answer has to be "the building" — and the less the financing is really about compute at all.

The precedent nobody in this cycle wants to name: vendor financing

There is a reason the phrase "circular financing" attaches itself to every one of these transactions, and it is not merely rhetorical. The equipment industry has run this experiment before, at scale, and the results are in the record.

Between roughly 1997 and 2001, Lucent Technologies and Nortel Networks extended credit to their own customers to fund purchases of their own equipment. The logic was identical to the one being made today: demand was real, the buyers were credible, the equipment was productive, and the vendor understood the collateral better than any bank. The practice flattered reported revenue, because a sale booked against vendor-provided credit is still a sale. When the competitive-carrier buyers proved unable to service that debt, the receivables were written down, in economic terms, some of that demand proved to have been pulled forward through vendor credit, and both vendors were structurally impaired — Lucent was ultimately absorbed into Alcatel, and Nortel filed for bankruptcy protection in 2009. The specifics — commitment levels, receivable write-downs and their accounting treatment — are documented in both companies' SEC filings from the period and in contemporaneous financial-press and academic coverage of telecom vendor financing; see references 25–26.

The differences from today matter, and they cut in NVIDIA's favour. Lucent lent directly to weak counterparties and held the paper itself. NVIDIA's platform structure interposes six independently capitalized underwriters between the supplier and the credit, caps its own participation at a stated share of the shortfall, and is attached to hardware with a genuinely fungible, multi-customer secondary use — which telecom switching gear installed in a specific carrier's network largely did not have. Those are real structural improvements, not cosmetic ones.

But the similarity that survives all of them is the one that mattered last time: the supplier's revenue and the buyer's ability to pay are not independent variables. Every mechanism on this page — the ≤25% support, the anchor-LP equity in the Valor vehicle, the reported Ohio guarantee, the $150M paid to control a competitor's access to a site — routes some portion of demand risk back toward the entity whose revenue depends on that demand being real. Diversifying the intermediaries does not make the underlying exposure independent; it makes it harder to see. The correct conclusion is not that this ends the way telecom did. It is that the burden of proof sits with the structure, and the single most useful question to ask of any deal in this market is the one Lucent's investors asked too late: if the offtaker could raise this money on its own credit, why is the supplier standing behind it?

The most useful reading of the table above isn't which risks get transferred — it's which ones don't, cleanly. Technology obsolescence, residual value, counterparty concentration and refinancing risk all show partial or no transfer, meaning contracts route them to a different party rather than resolving them. That's not a criticism of the structure; it's how project finance has always worked for depreciating physical assets. The distinguishing feature of compute is that the untransferable residual sits on top of an unusually short and contested Compute Duration (§05) — which is why the identity of whoever ends up holding technology-obsolescence risk, deal by deal, is the single most important disclosure this market does not yet reliably make.

Oracle is the closest live example of AI compute financed at scale without the diversified, multi-financier structure NVIDIA is now proposing. Its credit profile is under measurable pressure amid the combination of AI-buildout leverage, large external funding requirements and unusually high customer concentration — a combination, not a single cause.

Google is the natural counterpoint to a page built around "everyone needs external capital to finance AI compute" — for a meaningful share of its own workloads, it doesn't, because it designs and owns its own accelerator, the TPU.

OpenAI and Anthropic cannot be rated as securities — both are private — but their sourcing behavior is one of the more informative data points in this entire framework, because the two companies have taken structurally opposite approaches to the same underlying risk.

The open analytical question this raises: does diversified compute sourcing reduce technological and counterparty duration risk, or does it just relocate the same risk across more counterparties? Anthropic's multi-vendor posture plausibly reduces single-supplier renegotiation leverage and hedges architecture-cadence risk (§05) — no one vendor's roadmap change can strand its entire compute base. It does not eliminate Compute Duration risk generically; it prevents that risk from concentrating in one relationship the way it has in OpenAI's Oracle exposure. Microsoft, Amazon and Google appear in this section only as one-line factual context — principal backer, chip supplier — never as positions.

A framework worth publishing has to say, in advance, what would falsify it. The list below isn't a bear case dressed up as balance — it's the concrete, observable set of things that would tell us the compute-as-an-asset-class thesis is failing, alongside the things that would tell us it's working as designed.

Note what's absent from both columns: NVIDIA's own share price, and any single quarter's headline capex number. Neither is diagnostic for this specific thesis — compute-as-an-asset-class is a claim about financing structure and duration economics, not about demand for AI compute in general, which is a separate (and separately covered) question.

Structural roles, not positions. Where a name carries a rating elsewhere on this site, we link out rather than re-run the call.

The mechanism-level read: this financing wave doesn't change who wins on physical AI infrastructure buildout — that's still the AI Infrastructure thesis. It changes who takes the balance-sheet risk for that buildout, and at what price. That's a capital-structure question, and capital-structure questions are best answered by tracking how much of the disclosed $500B actually converts into signed, priced vehicles — not by re-running an equity thesis this site already publishes elsewhere.

A report built around the phrase "compute as an asset class" owes the reader the test. "Asset class" is not a marketing term, and most things called asset classes on the way up do not survive scrutiny as one. There is no single universally accepted definition, so below we set out the A.L. Capital Five-Criterion Asset-Class Test — a practical institutional test synthesized from the characteristics allocators consistently look for before treating something as a distinct allocation rather than an expression of an existing one.

This is why the terminology on this page is deliberately careful. Compute is a financeable asset today — demonstrably, with closed transactions and drawn capital. Whether it becomes an asset class depends almost entirely on the third and fifth rows above, and both require something no announcement can supply: a functioning resale market, and a completed cycle.

Almost no private investor will ever hold a compute lease directly. The question that matters for a real portfolio is narrower and more useful: if compute financing is becoming a genuine allocation, where does the exposure already sit — usually unintentionally — and what is it correlated with?

None of this is a recommendation to buy or avoid anything. It is the observation that the most common error with a genuinely new financing structure is not mispricing it — it is holding it four times over, in four different wrappers, while believing you hold it once.

Four Intelligence pages now touch the AI compute buildout from four different angles. We show the map explicitly so a reader knows exactly which question each page answers, and doesn't mistake overlap in subject matter for disagreement in conclusions.