CARMEN LI: Rising rental curves mean GPUs are currently holding value well above standard accounting depreciation

The Gist

Carmen Li says GPU rental prices are up and used GPUs are worth more than accounting rules assume, so they are acting like assets that earn their keep. This is a steelman: Carmen Li's arguments are reconstructed in their strongest fair form, not endorsed.

Conclusion

Because H100 and B200 rental curves have shifted up and do not price a steep decline, and used A100, H100 and B200 GPUs are valued well above five-year straight-line depreciation, GPUs are currently behaving more like income-producing assets valued on rental income than like fast-depreciating IT equipment.

Premises

1. H100 and B200 rental curves have shifted up across all terms compared with a year ago and have flattened, so current market pricing does not expect a steep fall in rental rates over the next 36 months, though forward rates are not forecasts.
2. Used A100, H100 and B200 GPUs are currently valued well above what five-year straight-line depreciation implies, because their rental income remains strong.

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.
• Much of the evidence comes from Silicon Data's own indices and model; independent trackers broadly agree on rising H100 rates and H100 resale strength but differ on A100 and B200 resale values.
• The claim is about current values and could reverse with new supply or a new GPU generation.

Analysis

Argument type: Inductive.

Potential Fallacies

• Non sequitur / missing bridge premise (Inference from P1 and P2 to the conclusion) — The step from 'market values exceed five-year straight-line book value and curves are flat' to 'GPUs behave like income-producing assets' depends on an unstated criterion for what income-asset behavior is and what it predicts that a cyclically scarce piece of equipment would not. Market value above book is common for aircraft, trucks and machine tools in tight markets, so it does not by itself indicate a change of category.
• False dichotomy / category mismatch (P2 and the conclusion) — The conclusion contrasts 'income-producing assets valued on rental income' with 'fast-depreciating IT equipment' as if they were exclusive. A depreciating asset can be valued by its income stream, and almost any productive asset is. Accounting depreciation allocates cost over an assumed life; it does not predict resale prices. Beating it is therefore a low bar and does not show a different asset class.
• Partial circularity / non-independent evidence (P2 ('because their rental income remains strong') and P1+P2 jointly) — P2 attributes high valuations to rental income, which is close to the conclusion's claim that GPUs are valued on rental income. Residual-value estimates built from rental models also cannot independently confirm that rental income drives value, and the curves and the residuals come from the same provider. The two premises therefore add less combined support than two independent sources would.
• Reading expectations into forward prices (P1) — P1 says forward rates are not forecasts, then uses the flat curve to say the market does not expect a steep fall. Forward prices embed risk and liquidity premia, hedging demand and, in a young market, model assumptions or few trades. The disclaimer is honest but sits in tension with the inference drawn.
• Unsupported causal inference (alternative explanations not excluded) (P2) — High resale values are attributed to strong rental income without ruling out scarcity premia, power and data-center constraints, speculative demand or modeled valuations. Rental income and resale value are also two views of the same demand shock, so their co-movement does not show that one causes the other.
• Extrapolation from a short, favorable window / hasty generalization (Conclusion and P1) — A year of rising rates in a supply-constrained market is used to characterize the asset class. Earlier hardware cycles (mining GPUs, fiber, telecom equipment) show strong boom-time pricing followed by sharp declines when supply or a new generation arrived. The 'currently' hedge limits the literal claim but not the reclassification it invites.
• Uniform grouping of unevenly supported cases (P2 and the conclusion) — A100, H100 and B200 are bundled into one claim, but independent corroboration exists mainly for H100. A100 and B200 resale values are contested, and B200 has very little secondary-market history, so a five-year comparison is thinly grounded for it.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — Current prices are what a temporary scarcity (power, packaging, data-center capacity) combined with an AI demand surge would produce. Boom-time rents in technology hardware have repeatedly preceded sharp write-downs (crypto-mining GPUs, fiber, telecom gear). The relevant question is how values evolve once new generations and capacity arrive, which a snapshot cannot answer.
• Premise 2 (High impact) — Straight-line depreciation is an accounting allocation, not a market-value forecast. Market value above book is common for many equipment classes in tight markets and does not show that the asset is income-like rather than depreciating. The more relevant comparison is with realistic economic depreciation after obsolescence and power-cost adjustment.
• Premise 2 (High impact) — Residual values are derived from rental-income models supplied by the same provider as the curves, so they cannot independently confirm that rental income drives value. Independent transaction-level resale data (brokers, auctions, secondary-market bids) is needed, and independent trackers already diverge on A100 and B200.
• Premise 1 (Medium impact) — A forward curve in a nascent, thinly traded market with few participants reflects index construction, hedging flows and risk premia, so flatness tells us little about expected future rents. The author concedes forward rates are not forecasts.
• Conclusion (Medium impact) — Gross rental rates ignore power, cooling, hosting, financing and counterparty costs. Older GPUs are less efficient per unit of compute, so net yields can erode faster than headline rates, undercutting valuation on rental income.
• Conclusion (Medium impact) — The argument is made by the operator of the exchange and the benchmark provider, and the framing supports demand for their products. The interest is disclosed, and independent corroboration exists for H100, so this bears on weight rather than validity, but it raises the bar for independent verification.

Suggested Improvements

• Bridge premise and falsifiable criterion — State explicitly what behaving like an income-producing asset predicts (for example, resale prices tracking rental yield, with implied discount rates comparable to other income assets) and what a depreciating-equipment model would predict instead. This would turn a loose analogy into a testable claim and separate a durable regime from a cyclical scarcity premium.
• Independent, transaction-based evidence — Support resale claims with observed transactions from several sources (brokers, auctions, secondary-market sales) and report volumes, rather than relying mainly on modeled residuals from one provider. This removes the circularity between the rental models and the valuations, and resolves the A100 and B200 disagreement.
• Benchmark choice — Compare market values with realistic economic depreciation scenarios, and with the schedules operators actually use (3 to 6 years), not only a single five-year straight line. Beating a conservative accounting convention is a low bar. A stronger test uses the alternatives that reflect obsolescence.
• Net economics — Report net yield after power, hosting, maintenance and financing costs, and implied discount rates on resale prices. Asset value depends on net cash flows over remaining life, and older, less efficient chips may look worse on that basis.
• Forward-curve reliability — Disclose the liquidity, open interest and methodology behind the curves, and backtest earlier forward curves against realized rates through the V100 and A100 transitions. This would show whether flat curves carry real information about future rents.
• Differentiate by GPU generation — Treat A100, H100 and B200 separately and scope the conclusion to the strongest-supported case (H100) unless independent data supports the others. The support is uneven, and a uniform claim overstates what the evidence shows.
• Down-cycle and obsolescence scenarios — Add explicit stress cases for a new generation shipping at scale, a supply glut, and an AI demand pullback, with the resulting rent and value paths. The reclassification only matters if it holds when conditions turn, and users of the framing (lenders, buyers) need to know its failure conditions.

Scenario Tests

• A next-generation architecture ships at scale with much better performance per watt and per dollar, and H100 rental rates and resale prices fall quickly toward or below depreciation schedules. (Challenges) — The 'currently' hedge would protect the literal claim, but the practical implication, valuing GPUs like income-producing assets, would fail when it mattered most. It would show the strength was a transient scarcity premium.
• Independent transaction-level data show H100 resale prices tracking rental yields over time at implied discount rates comparable to other income-producing assets. (Supports) — This would corroborate a rental-income valuation mechanism for H100 and make the classification more than an analogy.
• Independent trackers continue to show A100 and B200 resale values well below the modeled estimates. (Challenges) — P2 would hold only for H100, and the three-chip conclusion would need narrowing. It would also raise doubts about the model's residual values.
• Rental rates stay elevated and forward curves stay flat through the arrival of a new generation and additional capacity over the next two to three years. (Supports) — This would indicate durable income-based pricing and extended useful lives, and would support revisiting depreciation assumptions.
• A supply-constrained market with high prices in older hardware, as in the 2017-2018 crypto-mining GPU boom, followed by a collapse of 50 to 70 percent after supply caught up. (Challenges) — Rising rents and above-book resale prices in a shortage are consistent with a cyclical boom, so these observations alone cannot distinguish an income-asset regime from a temporary squeeze.
• Lenders extend long-amortization, high-advance-rate GPU-backed loans on benchmark-derived residual values, then rental rates fall. (Challenges) — Treating a current valuation snapshot as a durable asset property could amplify losses through collateral shortfalls. Independent valuation and stress haircuts would be needed.

Coherence & Relevance

The argument is coherent and appropriately hedged as an inductive case. It scopes itself to the present, concedes reversibility and discloses the author's interest and the limits of independent corroboration. The two premises converge on a plausible picture of strong current GPU pricing, particularly for H100. The main gaps are structural: no stated criterion for what 'behaving like an income-producing asset' means, a framing that treats income valuation and depreciation as mutually exclusive, partial circularity between the rental models and the valuations, and a benchmark (accounting depreciation) that is not a market-value prediction. The premises make the narrow claim of current H100 strength credible but do not establish the broader reclassification.

• H100 and B200 rental curves have shifted up across all terms compared with a year ago and have flattened, so current market pricing does not expect a steep fall in rental rates over the next 36 months, though forward rates are not forecasts. (Moderate) — A rising, flat curve is consistent with durable income, but also with a temporary shortage, thin liquidity or model smoothing. It concerns rents, not the asset's valuation or depreciation dynamics, and it does not say what separates an income-asset regime from a hot rental market. The explicit hedge that forward rates are not forecasts limits…
• Used A100, H100 and B200 GPUs are currently valued well above what five-year straight-line depreciation implies, because their rental income remains strong. (Moderate) — Exceeding an accounting schedule is relevant to the claim that GPUs retain value, but it does not show a change of category, since book value and market value routinely differ. The causal link to rental income is asserted, not tested, and evidence for A100 and B200 is weaker than for H100.

View this argument on LogicFirst.ai