CARMEN LI: Market valuations of used A100, H100 and B200 GPUs remain well above five-year straight-line depreciation

The Gist

Carmen Li says used GPUs are worth more than accounting rules assume. Older A100s still beat the book value, H100 prices went up, and B200s sell above list. This is a steelman: Carmen Li's arguments are reconstructed in their strongest fair form, not endorsed.

Conclusion

Used A100, H100 and B200 GPUs are currently valued well above what five-year straight-line depreciation implies, because their rental income remains strong.

Premises

1. A GPU's value can be estimated as the discounted cash flow from renting it, using forward rental rates, a declining utilization rate and operating costs.
2. On that basis, six-year-old A100s have lost value but remain above five-year straight-line book value.
3. H100 values have risen as rental rates rose, and used H100 servers sell for well over half their original price.
4. B200s have traded above their list price in the secondary market.
5. Lenders use such appraisals when financing GPU servers.

Assumptions

• Carmen Li is CEO of Compute Exchange, the venue she describes, and of Silicon Data, which publishes the GPU rental indices, forward curves and residual value estimates she cites and which sponsors the index that the planned CME compute futures would settle on. She has a disclosed commercial interest in the exchange model and in the use of these benchmarks.
• Differs: The A100, H100 and B200 valuations come mainly from Silicon Data's model; independent resale data partly agrees for H100 and differs for A100 and B200.
• Differs: Most large clouds depreciate servers over five to six years, and CoreWeave uses six.
• Differs: Independent forecasts expect H100 resale values to fall steeply.

Analysis

Argument type: Inductive.

Potential Fallacies

• Circular validation (model as evidence) (P1, P2 and the conclusion) — The valuations come from a rental-income DCF, and the conclusion says they are high because rental income is strong. A model's output is being used to support the model's own premise, with no independent test of its predictive accuracy.
• Equivocation on 'value' (P1/P2 versus P3/P4 and the conclusion) — 'Valuation' slides between a modeled DCF estimate (P1, P2) and observed secondary-market prices (P3, P4). These are different quantities, and the conclusion treats them as one.
• Inapt or trivially satisfied benchmark (P2, P3, P4 and the conclusion) — Straight-line book value is a cost-allocation convention, not a market forecast. For six-year-old A100s it is near zero, so almost any positive price exceeds it. For B200s, trading above list price is not a depreciation comparison at all. The H100 'well over half' claim is not tied to the units' age.
• Hasty generalization across products (P2-P4 to the conclusion) — The three chips are at different lifecycle stages and supported by different kinds of evidence (model-only for A100, partly corroborated for H100, a list-price comparison for B200). Strength for H100 does not transfer to the other two, and the conclusion pools all three.
• Extrapolation from a transient state (P3, P4 and 'because rental income remains strong') — Current rental strength and scarcity premiums are treated as durable. New GPU generations, supply responses and the independent forecasts of steep H100 declines can erode this. A value that exceeds the straight-line curve now can still fall below it later.
• Appeal to institutional practice (P5) — That lenders use appraisals shows they are accepted, not that they are accurate. Lenders may rely on the same vendor, apply large haircuts, or follow boom-time optimism, so P5 gives little independent support.
• Unsupported causal claim (Conclusion ('because their rental income remains strong')) — Rental income is an input to the DCF, so 'because' is largely definitional. It is not shown that rental strength, rather than scarcity, export controls or thin-market effects, drives observed resale prices, especially for B200.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — Technological obsolescence dominates GPU value. Each generation (Blackwell, Rubin) pushes older chips' rental rates toward the marginal cost of newer compute. Today's strong rents are a transient scarcity rent, so a DCF that extrapolates them overstates value, and independent forecasts of steep H100 declines point to convergence toward or below depreciation schedules.
• Premise 1 (High impact) — The DCF is highly sensitive to the discount rate, the utilization decay and the forward rental path, and the forward curve comes from the same firm that produces the valuation. A higher discount rate, faster decay or a 30-50% rental decline could push value below straight-line book, so the method can justify a wide range of prices.
• Premise 2 (High impact) — A six-year-old asset has near-zero five-year book value, so being above it is a trivial bar. Independent resale data also differs for A100, which suggests the model overstates realized prices.
• Premise 4 (Medium impact) — B200 above-list trades reflect delivery-time arbitrage and rationed primary supply on near-new units. Like 2021 crypto-era GPU premiums, such scarcity premiums can vanish quickly and are not evidence of durable residual value.
• Premise 5 (Medium impact) — Lender acceptance is not validation. Collateral appraisals have been widely used and inflated before busts, and a benchmark used for lending and futures settlement can become self-referential, with thin trades moving the marks that determine loan values.
• Conclusion (Medium impact) — Straight-line depreciation is an accounting policy, not a prediction of resale value. Exceeding it does not show that depreciation policies are wrong, and large clouds' five-to-six-year schedules may reflect conservative accounting choices for different purposes.

Suggested Improvements

• Benchmark specification — State the cost basis, salvage value, unit age and schedule length (including the six-year schedules in actual use) for each chip, and quantify 'well above' with a stated margin. The straight-line comparison cannot be evaluated without these. It is trivial for A100s and not comparable for new B200s.
• Independent validation — Show Silicon Data's valuations beside transaction-level data (brokers, auctions, lender sales) and publish a backtest of past residual-value estimates against realized prices. This would address the circularity, and the divergence for A100 and B200, which are the main evidential weaknesses.
• Sensitivity and scenario analysis — Present valuations under varied discount rates, utilization decay paths and rental declines tied to next-generation supply, and show where they cross below book value. It shows how robust the conclusion is to the forecast inputs, which are the model's most uncertain element.
• Engagement with contrary evidence — Address the independent forecasts of steep H100 declines and explain why current prices or rental rates should be trusted over them. Current prices cannot refute a forecast about future prices, so the argument needs a direct reply.
• Scope and claim calibration — Restrict the conclusion to what is supported: H100 above straight-line book value by a stated margin, with model-based and less certain findings for A100 and B200. Treat B200 above-list trades as a scarcity signal and not as depreciation evidence. It matches the claim to the quality of evidence for each chip and removes the unsupported causal and generalizing language.
• Lender evidence — Replace the bare fact of lender use with data on lender haircuts, loan-to-value ratios and recoveries on defaulted GPU-backed loans. Realized recoveries and conservative haircuts would test the appraisals, whereas use alone does not.

Scenario Tests

• A new architecture (e.g., Rubin) ships at scale and cuts H100 and A100 rental rates by 30-50%. (Challenges) — The DCF values would fall sharply, possibly below straight-line book. Strong current rent does not protect value from a supply and performance reset.
• Independent transaction records (brokers, auctions) for A100 and B200 are compared with Silicon Data's marks. (Challenges) — Independent data already differs for these chips. If the gaps are material, the claim holds only for H100, which weakens the generalization.
• H100 resale prices stay well above 40-50% of original price over the next 12-24 months despite the forecasts of steep declines. (Supports) — This would support the H100 portion of the claim and the usefulness of rental-based valuation, although it would not by itself confirm the A100 or B200 claims.
• Primary B200 supply normalizes and the above-list secondary premium disappears. (Challenges) — It would show that P4 reflected temporary scarcity, not durable income-based value.
• The same valuation is tested against a six-year straight-line schedule, as CoreWeave uses. (Neutral) — H100 and B200 margins narrow somewhat. The A100 comparison remains trivially satisfied at both horizons. The outcome shows how sensitive the 'well above' claim is to the chosen benchmark.
• A downturn in AI demand lowers utilization and triggers a wave of GPU-backed loan covenant breaches. (Challenges) — Appraisals tied to current rents would reprice together across lenders, exposing the reflexive and procyclical nature of mark-to-model collateral.

Coherence & Relevance

The argument has a clear structure (a valuation method, three chip-specific observations and a supporting institutional practice), but it is inductive and its inferences are loose. The key gaps are a bridge from model value to market price, comparable benchmarks for each chip's age, evidence that rental strength is the actual driver, and a response to contrary forecasts. The H100 claim is reasonably coherent and corroborated, while the A100 and B200 claims fit poorly with 'well above' and with the stated rental-income mechanism. The conclusion as stated overreaches the premises.

• A GPU's value can be estimated as the discounted cash flow from renting it, using forward rental rates, a declining utilization rate and operating costs. (Strong) — It supplies the method that links rental income to value, but gives no inputs, discount rate or validation. It also does not bridge model-based value to observed market prices.
• On that basis, six-year-old A100s have lost value but remain above five-year straight-line book value. (Moderate) — It addresses only the A100 and restates part of the conclusion. After six years, the benchmark is near zero, so the comparison is weakly informative about 'well above'.
• H100 values have risen as rental rates rose, and used H100 servers sell for well over half their original price. (Strong) — It directly supports the H100 claim and the rental-value link, but does not state unit age to compute book value, and the co-movement with rental rates is not separated from common drivers such as supply scarcity.
• B200s have traded above their list price in the secondary market. (Weak) — Trading above list price is not a comparison with depreciated book value and is better explained by launch-period scarcity than by rental income, so it does not connect to the DCF logic or the five-year benchmark.
• Lenders use such appraisals when financing GPU servers. (Weak) — It concerns the use of appraisals, not their level or accuracy, and does not bear on whether values exceed book or why.

View this argument on LogicFirst.ai