SRIDHAR RAMASWAMY: Compute books work like a commodity merchant's take-or-pay contracts, so small demand or price errors can erase the spread

The Gist

Sridhar Ramaswamy says AI labs are like gas traders who promised to buy supply for years: they owe the money even if sales fall, so a modest sales miss can wipe out most of the profit. This is a steelman: Sridhar Ramaswamy's arguments are reconstructed in their strongest fair form, not endorsed.

Conclusion

Because an AI lab pays for committed compute regardless of demand while selling at uncertain volumes and prices, a modest demand or pricing error can erase nearly all of its spread, just as it can for a merchant holding take-or-pay contracts.

Premises

1. A commodity merchant signs a long take-or-pay contract, such as a 10-year LNG supply deal, and then sells cargoes at market prices.
2. If prices or volumes collapse, the merchant must still pay the supplier, so the merchant carries the price and volume risk.
3. An AI lab's compute book has the same shape: committed capacity costs are paid regardless, while revenue depends on usage and price per token.
4. Illustration: committing $100 to capacity while expecting $150 of revenue yields $50 of contribution; a 30 percent revenue miss cuts revenue to $105 and contribution of $5.
5. High fixed commitments therefore create strong operating leverage in both directions.

Assumptions

• Sridhar Ramaswamy is CEO of Snowflake. Snowflake resells AI model usage (including Claude under a multi-year $200 million partnership with Anthropic announced in December 2025), earns consumption-based revenue, and itself carries multi-year minimum spend commitments with cloud providers (for example a cumulative $6.0 billion minimum spend over a five-year term ending March 31, 2031, under a cloud infrastructure agreement amended in April 2026). He therefore has a business interest both in AI lab health and in how commitment risk is understood. Sources: Snowflake press release of December 3…
• The $100 and $150 figures are an illustration, not reported numbers for any lab.
• Differs: Not all lab compute is take-or-pay: Anthropic also has agreements with no fixed or minimum dollar commitment, where payments depend on capacity delivered, timing and pricing.
• Differs: Anthropic spent $5.6 billion on compute against $11.6 billion of revenue in the second quarter of 2026, about $2.08 of revenue per compute dollar, a wider cushion than the 1.5 times illustration.

Analysis

Argument type: Inductive.

Potential Fallacies

• Weak or overextended analogy (Premise 3 and the conclusion ('just as it can for a merchant')) — Fixed costs against variable revenue are the one shared feature of the lab and the merchant, and the argument uses it to import the merchant's whole risk profile. Merchants resell a fungible commodity at benchmark prices and can hedge and index. Labs sell differentiated products with some pricing power, face secular token-price declines alongside rapid volume growth, and can redeploy idle capacity to training. 'Same shape' holds for the committed portion of the book, and the argument never says…
• Illustration treated as support for magnitude (hasty generalization) (Premise 4 to the conclusion) — A hypothetical with a chosen 1.5x ratio, 100% fixed cost and a 30% miss shows that severe fragility is possible. It cannot show that real labs are that fragile, and the sensitivity depends entirely on the chosen parameters. The disclosed 2.08x ratio roughly halves the damage for the same miss.
• Equivocation on 'small' or 'modest' (Conclusion versus Premise 4) — The conclusion speaks of a modest error, but the worked example needs a 30% revenue miss, which is close to break-even at a 1.5x cushion. Whether 30% is modest depends on how accurately lab revenue can be forecast, and no evidence on that is offered. The word does persuasive work without being defined.
• Neglect of probability (Conclusion and Premise 5) — The argument shows how damaging a miss would be, but not how likely one is. Premise 5 says leverage works in both directions, and lab revenue has recently surprised on the upside, so the relevant tail may be favorable. The argument supports 'high variance' more directly than 'high fragility'.
• Quantifier overreach / suppressed premise (Premise 3, Premise 4 and the conclusion) — The conclusion implies that an AI lab pays for committed compute regardless of demand, as if all of it were committed. The stated assumptions show that part of lab compute carries no minimum. The argument needs an unstated premise that the committed share is large enough to drive the result, and the illustration simply sets it at 100%. The model also treats compute minus revenue as the whole 'spread', ignoring other costs and the discretionary nature of training compute.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — At Anthropic's reported 2.08x revenue per compute dollar, a 30% miss still leaves about $1.46 of revenue per $1 of compute. Erasing the spread would take a miss of roughly 52%, which is not modest. The headline claim holds only for thin-margin cases.
• Premise 3 (High impact) — Not all lab compute is take-or-pay. Some agreements tie payments to delivered capacity, timing and pricing. Unused capacity can often be resold or redirected to training and research. Exposure depends on the committed share, which is unstated, so 'same shape' overstates the fixed-cost rigidity.
• Premises 1-3 (Medium impact) — Labs differ from LNG merchants. They sell differentiated, fast-growing products with some pricing power, and token prices decline structurally rather than mean-revert. Counterparties such as cloud providers also have incentives to renegotiate rather than enforce against a major customer. The analogy therefore may mislead about the nature and size of the risk.
• Premise 5 (Medium impact) — If leverage truly runs in both directions and demand has been growing sharply, forecast errors may skew toward upside. The binding risk could be under-provisioning rather than stranded commitments, so the argument supports high variance more than fragility.
• Premise 4 (Medium impact) — Treating revenue minus compute as the whole 'spread' ignores other costs and the discretionary nature of training compute. It also ignores the time structure of staggered, ramping commitments. It is a gross-margin proxy and not a merchant-style trading spread, so 'erasing the spread' is not the same as distress or insolvency.
• Conclusion (Low impact) — Every high-fixed-cost business, such as airlines, telecoms and cloud-hosted software, shows operating leverage. Labeling it 'commodity merchant take-or-pay' adds a vivid image but little AI-specific insight, so the analogy has limited discriminating power.

Suggested Improvements

• Calibrate the illustration to reported data — Run the sensitivity at both 1.5x and the reported ~2.08x, and show the break-even miss for each (about 33% versus about 52%). Present a table of 10, 20, 30 and 50% misses. This ties 'modest' and 'nearly all' to real parameters and shows the reader how sensitive the result is to the starting cushion.
• Define the claim conditionally — State that contribution falls by roughly miss × ratio / (ratio − 1), where the ratio is revenue to committed cost and fixed costs are fully committed. Then say for which ratios and miss sizes the spread is largely erased. This makes the conclusion precise, falsifiable and valid as stated, instead of resting on one chosen example.
• Specify the committed share and tenor — Separate committed take-or-pay compute from flexible or usage-linked compute, and training from inference, and state contract tenors. Operating leverage scales with the fixed share, so this is the parameter that determines how far the analogy holds.
• Address disanalogies and mitigants — Discuss resale, reallocation to training, renegotiation, hedging and indexing, and pricing power. Say where labs differ from merchants and where they are more exposed, such as falling token prices and hardware obsolescence. Naming the differences keeps the analogy honest and makes the argument more persuasive to informed readers.
• Add probability and symmetry — Give some indication of the likely distribution of demand and price forecast errors, and state the upside case as prominently as the downside. Magnitude without likelihood overstates the risk, and the author's own premise says leverage cuts both ways.
• Integrate caveats into the main chain — Move the points that the compute book is only partly take-or-pay and that the real cushion is wider into the premises or the conclusion's wording. Disclose the author's commercial interest alongside the claim. This aligns the headline with what the author already acknowledges and lets readers weigh the claim with full context.

Scenario Tests

• A lab has a 1.5x revenue-to-committed-compute ratio, with nearly all compute under take-or-pay contracts, and revenue falls 30% short of forecast. (Supports) — Contribution falls about 90%, from $50 to $5, as in the illustration. In thin-cushion, fully committed cases the claim holds, but those are the conditions the argument has to establish for real labs.
• The same 30% miss is applied to Anthropic's reported Q2 2026 figures ($11.6B revenue, $5.6B compute). (Challenges) — Revenue falls to about $8.1B, leaving roughly $2.5B of margin over compute, a cut of about 58% but not erasure. Break-even needs a miss of about 52%. Leverage is still large, but 'modest error erases nearly all' does not describe this case.
• A substantial share of a lab's compute is flexible, usage-linked or resalable, and idle capacity can be redirected to training. (Challenges) — Downside is capped well below the full commitment, so the 'must still pay regardless' framing overstates the exposure. The analogy then describes only part of the book.
• Demand growth exceeds forecasts and revenue surprises on the upside while commitments are fixed. (Neutral) — The leverage works in the lab's favor, consistent with Premise 5. The mechanism is validated, but the fragility narrative is not, which shows that the argument establishes variance more than downside risk.
• Demand falls sharply, token prices drop and capacity cannot be resold, as in the telecom and fiber overbuild or LNG take-or-pay stress episodes. (Supports) — Fixed commitments combined with large demand or price shortfalls have caused real distress historically, although renegotiation and resale often moderated it. The mechanism is real, but the outcome depends on how large the shock is and on the mitigants available.

Coherence & Relevance

The argument is internally coherent at the level of mechanism: the take-or-pay structure, the lab analogy and the operating-leverage corollary fit together, and the arithmetic is right. The weak point is the final step from mechanism to magnitude. The words 'modest' and 'nearly all' depend on a thin 1.5x cushion that the stated assumptions (a wider reported cushion and partly flexible compute) undercut. The argument is best read as a sound qualitative point about operating leverage, with an overstated quantitative headline.

• A commodity merchant signs a long take-or-pay contract, such as a 10-year LNG supply deal, and then sells cargoes at market prices. (Strong) — Sets up the source domain well. It does not mention that merchants usually hedge, index prices or resell, which affects how far the analogy transfers.
• If prices or volumes collapse, the merchant must still pay the supplier, so the merchant carries the price and volume risk. (Strong) — Establishes the risk allocation that drives the argument. It says nothing about the probability of a collapse, or about how often and by how much the obligation is renegotiated.
• An AI lab's compute book has the same shape: committed capacity costs are paid regardless, while revenue depends on usage and price per token. (Strong) — This is the crucial bridge, and it is asserted without specifying the share of compute that is committed, contract tenor, or lab-specific price dynamics. The assumptions show the shape is only partial.
• Illustration: committing $100 to capacity while expecting $150 of revenue yields $50 of contribution; a 30 percent revenue miss cuts revenue to $105 and contribution of $5. (Moderate) — It demonstrates the mechanism but its parameters (1.5x, 100% fixed, 30% miss) are chosen and not benchmarked to real data. It does not substantiate 'modest' error or real-lab fragility.
• High fixed commitments therefore create strong operating leverage in both directions. (Moderate) — A correct generalization of the illustration, but it adds no independent support for the conclusion's magnitude, and the conclusion emphasizes only the downside direction.

View this argument on LogicFirst.ai