On SemiAnalysis Episode 33, host Jordan Nanos asked a question that AI infrastructure finance has not yet answered: how do you underwrite a terminal value on hardware that has never been retired?
In SemiAnalysis Episode 33, titled “ClusterMAX 3.0 Is Here! Neoclouds Ranked,” Jordan Nanos — joined on the call by Sam Harshe and Prathmesh Bhatt — described a GPU market in which supply constraints are reshaping the basic economics of infrastructure finance. His central observation: there are four-year contracts being signed today for H100 capacity. That single data point reframes what would otherwise be a straightforward asset-finance question into something harder to answer cleanly.
Nanos also made a specific claim about OpenAI’s A100 fleet, delivered as his characterisation rather than independently verified: that OpenAI has never given up an A100 GPU that they’ve gotten — units that arrived around 2020, making them now over six years old. He then turned the question toward the financial side of the table: How can you underwrite the terminal value at 5 years or 6 years on some of these chips when these contracts expire?
For context on the scale involved, Nanos put Nvidia’s off-balance-sheet backstop universe at over $588 billion at the end of FY27. He also cited a conditional SemiAnalysis forecast — framed explicitly as “the forecasts if this continues” — that the figure could grow to over $2 trillion by the end of 2031. That conditional forecast is his, not ours, and it is repeated here only to establish what the structural question below is actually pointing at.
The key insight: Every mature asset class prices long contracts against a residual value anchored in observed retirements. The GPU fleet is young enough that no such anchor exists — which means the number the entire arrangement depends on is not uncertain, it is unobserved. Those are structurally different problems.
The scale is stated. The residual value it rests on has never been watched all the way down.
The Structural Read
In aircraft finance, truck leasing, commercial real estate, and maritime shipping, underwriters price long-duration contracts by anchoring the terminal value in historical retirement data — thousands of observed outcomes, written down by people who watched the asset from delivery to scrapyard. The residual value is uncertain, but it has a distribution. You can argue about where in that distribution you land. The argument is productive because it is constrained by evidence.
For the current GPU fleet, that archive does not exist. The fleet is young. On Nanos’s account, the oldest units — the A100s — are still in active service rather than retired. The input that would anchor a residual value estimate has not yet been generated by the market. This is the position Nanos’s question names directly, and it is worth being precise about what it means: the problem is not that the residual value is hard to estimate. It is that the evidence base for estimating it does not yet exist. An uncertain number has a distribution you can argue about. An unobserved number does not yet have a distribution at all — only a set of assumptions standing in for one.
Map of AI — Infrastructure Layer
The Unobserved Input Problem
The Map of AI framework distinguishes between companies that build infrastructure, companies that aggregate it, and companies that deploy it. The residual-value gap sits at the infrastructure layer and propagates upward: every contract that layers above it — neocloud leases, enterprise commitments, compute-backed financing — rests on a terminal assumption that the infrastructure layer has not yet empirically validated. The gap is not a flaw in any individual contract; it is a structural property of a market that has not yet aged enough to produce the evidence it needs.
Nanos’s A100 retention detail deserves the opposite reading from the obvious one. If the oldest hardware in the fleet has never been released, the absence of retirements is itself data. It suggests the binding constraint in this market is not performance-per-chip but available capacity — so older hardware keeps earning because the realistic alternative for a buyer is not newer hardware, it is nothing. That is evidence that cuts toward longer useful lives and higher residual values, not lower ones. The sceptical reading of AI infrastructure typically assumes rapid obsolescence; this evidence runs in the opposite direction and should be stated plainly.
The limit of that reading is equally plain: an asset that has not yet been retired tells you it has not yet been retired. It does not tell you when it will be. The data point extends the period of uncertainty; it does not resolve the question. Both of those things are true at the same time, which is exactly what makes the structural position Nanos describes genuinely interesting rather than merely alarming.
Jordan Nanos — SemiAnalysis Ep. 033 (not independently verified)
“How can you underwrite the terminal value at 5 years or 6 years on some of these chips when these contracts expire?”
Three Implications
IMPLICATION 1 — FOR INFRASTRUCTURE UNDERWRITERS
Four-year contracts on H100s require a terminal value assumption at year four or five. In every comparable asset class, that assumption is calibrated against observed retirements. The GPU market has not yet produced a retirement record, which means underwriters are, by definition, working with assumptions rather than distributions. That is not a scandal — it is the ordinary condition of any market early enough that the evidence has not yet accumulated — but it is a genuine structural feature that pricing models need to account for explicitly rather than implicitly.
IMPLICATION 2 — FOR THE OBSOLESCENCE THESIS
The standard bear case on GPU infrastructure assumes that successive hardware generations rapidly devalue older units. Nanos’s A100 observation — his characterisation, not a confirmed OpenAI fact — is direct evidence against that assumption in the current supply environment. When capacity is the binding constraint rather than performance, older hardware retains value not because it is optimal but because it is available. Anyone building a model that assumes linear performance-driven depreciation should at minimum stress-test that model against a capacity-constrained scenario.
IMPLICATION 3 — FOR THE SCALE OF THE BACKSTOP
The $588 billion off-balance-sheet backstop figure Nanos cited for end of FY27 — offered here with no allegation of any kind about its character or compliance — and his conditional forecast of over $2 trillion by end of 2031 (if current trends continue, per his framing) illustrate the aggregate stakes that rest on the missing residual-value anchor. A structurally uncertain input at the asset level does not stay contained at the asset level when the contract stack above it is measured in the hundreds of billions. The first retirement data, when it eventually arrives, will be informative at a scale well beyond the individual transaction.
Business Engineer Framework
The Map of AI — Infrastructure Layer
The Map of AI traces how value and risk distribute across nine layers of the AI stack — from raw compute and chip supply, through cloud aggregation and neocloud capacity, up to application deployment. Understanding where the residual-value gap sits (at the infrastructure layer) and how it propagates upward (through every contract and commitment that uses that compute) is the analytical move the framework is built to support. If you are trying to read the AI infrastructure market structurally rather than anecdotally, this is the lens.
The question Jordan Nanos put to the finance side of the AI infrastructure market on SemiAnalysis Episode 33 is not a prediction of collapse and it is not investment advice — it is a structural observation about the quality of the inputs that long-duration contracts depend on. The number those contracts need most, the residual value of the hardware at contract expiry, does not yet have an empirical distribution because the fleet has not yet aged out enough to produce one. That is the condition the market is in right now: precisely-priced obligations running on schedule against an asset whose end-of-life value is still, in the strictest sense, unobserved. The first retirements, when they come, will be among the most consequential data points the AI infrastructure industry has ever generated.
Source: SemiAnalysis Episode 33 — “ClusterMAX 3.0 Is Here! Neoclouds Ranked,” Jordan Nanos, Sam Harshe, Prathmesh Bhatt (YouTube). All claims originate with Jordan Nanos on that episode and are relayed here, not independently verified. The $2 trillion figure is a conditional SemiAnalysis forecast, not a fact and not this publication’s projection. The A100 retention claim is Nanos’s characterisation of OpenAI’s fleet. The term “off-balance-sheet” carries no allegation of impropriety, concealment, or non-compliance. Nothing here is investment advice. No outcome is predicted.
Single source. Everything above is what Jordan Nanos said on SemiAnalysis episode 33, relayed here and not independently verified. The figure of over $2 trillion by the end of 2031 is a conditional SemiAnalysis forecast — his framing was “if this continues” — not a fact, and not this publication’s projection. The statement that OpenAI has never given up an A100 is his characterisation, not a confirmed fact about any company’s fleet, and his expectation that prices keep rising is his own and is not adopted here. “Off-balance sheet” is his descriptive term and carries no allegation: nothing above suggests concealment, impropriety, irregularity or non-compliance by Nvidia or anyone else, and Nvidia has not commented here. What the backstops consist of, and who the counterparties are, is not established and is not described above. This is analysis of financing structure. It is not investment advice, it expresses no view on any security, and it predicts nothing about GPU prices, demand, residual values or any company. Nvidia’s financials, GPU prices, rental rates, residual values, depreciation schedules, fleet counts and contract parties are not established and do not appear above.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
