CARMEN LI: H100 and B200 term rental rates have shifted up across the curve over the past year, and the implied forward rates do not price a decline over 36 months

The Gist

Carmen Li says renting H100 and B200 GPUs costs more now than a year ago, for short and long contracts alike, and prices for future rentals no longer point steeply down. This is a steelman: Carmen Li's arguments are reconstructed in their strongest fair form, not endorsed.

Conclusion

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.

Premises

1. A year ago, longer H100 rental terms were cheaper than short ones.
2. During 2026, H100 term rates rose at every tenor; one-year rates rose from about $1.70 to about $2.65 an hour between October 2025 and March 2026.
3. B200 term rates are also higher at every tenor than a year ago.
4. The H100 and B200 curves have since flattened, so the implied forward rates show no steep expected decline over 36 months.
5. A forward curve reflects current market pricing, not a guaranteed forecast.

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 curves cited are built by Silicon Data from term rental rates; other trackers report different rate levels for the same GPUs.
• Differs: Silicon Data's own published curves were in backwardation, with longer terms cheaper than spot, in July and September 2026; the talk describes the curve as contango earlier in the year.
• Differs: Ornn data showed B200 rentals fell about 30 percent in three weeks after a May 30, 2026 peak, before rising again.

Analysis

Argument type: Inductive.

Potential Fallacies

• Reification of a model output as market expectation (P4 and the conclusion) — Implied forwards derived from term rental quotes in a thin, largely bilateral market are treated as 'current market pricing' that 'expects' outcomes. Such curves also embed term premia, credit and liquidity effects, depreciation and the vendor's interpolation choices, so flatness does not reliably reveal beliefs. P5's caveat only partly offsets this.
• Non sequitur from level shift to future expectations (P2, P3 to Conclusion) — Rates being higher than a year ago (P2, P3) describes the past and present. It says nothing about the 36-month path, since an elevated level can coexist with expected mean reversion. Only the curve's shape bears on expectations, and that is the least supported claim.
• Stale evidence / selective time window (P2, P4 versus the conclusion's 'current' claim) — The quantified rise covers Oct 2025 to Mar 2026, while the conclusion speaks in the present tense about a talk dated Oct 2026. Later evidence (backwardation in July and September, a ~30% B200 drop after May 30) is in tension with the picture of a stable, flat curve. The word 'flattened' is not given a date.
• Partial circularity in P4 (P4 to Conclusion) — P4 asserts both that the curves have flattened and that this means no steep decline is implied. The conclusion largely restates that, so the central inferential step is assumed rather than demonstrated with figures.
• Hasty generalization across tenors and products (P2, P3, Conclusion) — The only quantified datum is the H100 one-year rate, yet the claims extend to 'every tenor', to a 36-month horizon, and to B200, which is a newer product with different lifecycle risk.
• Headline stronger than hedge (Title versus P5) — The title ('do not price a decline over 36 months') reads as reassurance about durable prices, while the body concedes forwards are not forecasts. A casual reader is likely to take away the stronger message.

Counterarguments

These challenges are not ranked.

• Premise 4 (High impact) — Silicon Data's own curves were in backwardation (longer terms cheaper than spot) in July and September 2026. That shape is consistent with expected declines, and it conflicts with a 'flattened, no decline priced' reading unless the flattening is explicitly dated and shown to postdate the backwardation.
• Premise 4 (High impact) — Implied forwards from thin, bilateral term-rental quotes are model outputs, not traded prices. A flat 24–36 month segment may reflect sparse observations, interpolation, smoothing, or risk and liquidity premia rather than a consensus that rates will hold.
• Conclusion (High impact) — H100 is an older-generation chip. Blackwell and successor ramps, new capacity and hardware aging should push older-chip rental rates down over 36 months. A flat curve for a depreciating asset more plausibly signals illiquidity or a temporary scarcity premium than durable expectations.
• Premise 2 (Medium impact) — The Oct 2025 to Mar 2026 comparison may start from a trough, and the later ~30% three-week B200 drop after the May 30 peak (A4) shows rates are volatile and mean-reverting. A year-over-year rise therefore does not indicate a durable new level.
• Premises 2, 3 and 4 (Medium impact) — The curves, indices and forwards are largely from one vendor, whose CEO is the speaker. Other trackers differ on levels (A2), and corroboration confirms direction but not curve shape. H100 and B200 are also driven by common demand and supply factors, so they are not independent confirmations.
• Conclusion (Medium impact) — Because forwards are 'not forecasts' (P5), and any curve shape can be called 'what the market expects', the thesis risks being unfalsifiable. It needs a dated claim and a track record of how past forwards compared with realized rates.

Suggested Improvements

• Date-stamped curve evidence — Present tenor-by-tenor H100 and B200 rates at specific dates (a year ago, March 2026, July, September and the latest), and show exactly when the curves flattened relative to the backwardation episodes. This would resolve the apparent conflict between 'flattened' and the backwardation in A3, and would anchor 'current' to a real date.
• Curve construction and liquidity — Disclose how the 24–36 month points are derived: transaction counts, quotes versus trades, smoothing and extrapolation, and uncertainty bands on implied forwards. It would show whether the long end carries market information or is mainly model output.
• Independent corroboration of shape — Compare curve shape (not only levels) across Silicon Data, Ornn and other trackers, ideally with a third-party audit of the methodology. Corroboration so far covers direction of rates, not the flatness claim at the core of the conclusion.
• Calibration against outcomes — Backtest past implied forwards against realized rental rates, and estimate the term premium embedded in them. It would show whether the curve has any predictive value, which is what the 'market does not expect a fall' language implicitly asserts.
• Supply-side and generational analysis — Model how Blackwell/Rubin ramps, new capacity and depreciation could affect H100 and B200 rates over 36 months, and report scenarios with and without a supply surge. The main thing that could invalidate a flat reading is a supply-driven repricing, which the curve cannot anticipate.
• Framing and wording — Align the headline with the hedge, for example 'term rates are higher than a year ago and, on this curve, show no steep decline priced', and explicitly address volatility such as the 30% B200 drop. It prevents readers, especially non-specialists and lenders, from treating the curve as a reassurance about future prices.

Scenario Tests

• Independent trackers confirm that H100 and B200 term rates are higher at all tenors than a year earlier. (Supports) — The descriptive half of the conclusion (a level shift upward) is well supported. This says nothing yet about expectations or curve shape.
• The latest dated Silicon Data curves remain in backwardation, as in July and September 2026. (Challenges) — Longer terms priced below spot implies some expected decline, contradicting the 'flattened, no decline priced' reading. The conclusion would hold only if a later re-flattening were documented.
• A new GPU generation or capacity wave pushes H100 and B200 rental rates sharply lower within 12–18 months, as the B200 drop after May 30 hints. (Challenges) — A flat 36-month curve would have failed as a signal, which supports reading forwards as weak, premium-laden indicators. Because the argument concedes forwards are not forecasts, it would be hard to falsify, though anyone relying on the curve for long-term commitments or collateral would be hurt.
• The 24–36 month points are shown to rest on a handful of contracts or on vendor extrapolation. (Challenges) — The 'market pricing' claim reduces to a vendor model, so the flatness has little evidential value about expectations.
• Backtests show that Silicon Data's historical implied forwards tracked subsequent realized rates well, and other trackers' curves are similarly flat. (Supports) — The interpretive claim would gain real credibility, and the conflict-of-interest concern would be partly mitigated by independent verification.

Coherence & Relevance

The argument is coherent in structure but unevenly supported. The claim that rates are higher than a year ago follows reasonably from P2 and P3. The step to 'flattened' and then to 'the market does not expect a steep fall' relies almost entirely on an undated, vendor-built curve, and it sits uneasily with the same vendor's mid-2026 backwardation and the sharp B200 drop. The P5 hedge keeps the conclusion from being outright invalid, but the main claim is only weakly supported until the curve is dated, quantified and independently verified.

• A year ago, longer H100 rental terms were cheaper than short ones. (Moderate) — It supplies the baseline curve shape for 'flattened', but the date and source are unspecified. It conflicts with the later contango description (A3), and the link between a past discount and present expectations is unstated.
• During 2026, H100 term rates rose at every tenor; one-year rates rose from about $1.70 to about $2.65 an hour between October 2025 and March 2026. (Moderate) — It directly supports 'shifted up' but not 'flattened' or 'no decline priced'. The window and single tenor do not match the 'past year' and 'across all terms' scope, and the data stops months before the talk.
• B200 term rates are also higher at every tenor than a year ago. (Moderate) — It extends the level-shift claim to B200 but gives no magnitudes or tenors, and it does not address curve shape.
• The H100 and B200 curves have since flattened, so the implied forward rates show no steep expected decline over 36 months. (Strong) — It is the only premise bearing directly on the expectations claim, but it is asserted with no date, figures, tenor data or method. It collapses the inference from shape to expectation into the premise and conflicts with the documented backwardation in July and September 2026.
• A forward curve reflects current market pricing, not a guaranteed forecast. (Moderate) — It correctly limits the conclusion but does not say how much weight the curve deserves, and it is not reflected in the stronger headline wording.

View this argument on LogicFirst.ai