GAVIN BAKER: Anthropic's $413 billion of non-cancellable commitments is manageable once spread over its duration and set against its revenue

The Gist

Gavin Baker says Anthropic's scary $413 billion is mostly about $31 billion a year of cloud bills, which its revenue can cover, plus Broadcom chips that lenders are already financing. His weekend point was about loose memos, not these contracts. This is a steelman: Gavin Baker's thread is reconstructed in its strongest fair form, not endorsed.

Conclusion

The headline $413 billion of Anthropic non-cancellable contracts is large but reasonable relative to its duration and Anthropic's scale, because the cloud portion is about $31 billion a year that operations can fund and the Broadcom portion is being externally financed; the same test should be applied to OpenAI once its filing is public.

Premises

1. His weekend comments about unserious commitments referred to more than $1 trillion of 2025 memoranda with Asian supply chain and semiconductor companies, not to signed contracts.
2. Contracts matter and are taken seriously.
3. Per Anthropic's reported draft prospectus, its cloud commitments run about $11 billion a year with Amazon for a decade, about $15 billion a year with Google for seven years, and about $5 billion a year with Microsoft for about six years.
4. That is roughly $31 billion a year, priced below current spot and forward compute prices.
5. Against a reported revenue run rate he puts near $70 billion and high inference margins, that cloud spend can be paid from operations.
6. The roughly $161 billion of Broadcom-related commitments will be financed by Broadcom and private credit, which has begun.

Assumptions

• The prospectus was confidential and reported, not publicly filed.
• Anthropic's last reported run rate was about $65 billion at the end of July 2026; third-party estimates put it higher by October.
• Anthropic raised more than $95 billion of outside capital in 2026 and reported losses, and critics argue take-or-pay contracts function as debt equivalents for creditors.
• Some 2025 memoranda he refers to, such as OpenAI's Samsung and SK Hynix letters of intent, remained non-binding in October 2026.

Analysis

Argument type: Inductive.

Potential Fallacies

• Run rate treated as cash available to pay (equivocation on 'operations') (P5) — Revenue run rate and inference gross margin are not the same as free operating cash flow, which is what actually services a fixed obligation. Training, R&D and payroll sit outside inference margin, and the stated context of reported losses and heavy outside funding shows operations have not yet covered current spend. The comparison is a useful scale check but does not show the spend 'can be paid from operations'.
• Averaging away timing and tail risk (P3–P4 and the 'spread over its duration' framing in the conclusion) — Dividing commitments by their term gives a smooth average, but take-or-pay schedules are usually ramped or back-loaded. The three contracts also have different lengths (10, 7 and about 6 years), so $31 billion a year holds only while they overlap. Solvency is decided in the worst years, not the average year.
• Projecting a snapshot onto a long horizon (P5) — A run rate from a hypergrowth period is used to support obligations lasting 6–10 years. Growth, margins and pricing could all move sharply, but the fixed obligations will not. The argument offers no sensitivity or downside analysis.
• Conflating financing with risk reduction (P6) — Broadcom and private credit financing shows lenders are willing to extend capital now. It does not show the obligation is safe. Lenders are typically underwriting Anthropic's future payments, so the burden is moved rather than removed, and the willingness to lend is pro-cyclical and tied to the same demand outlook. Critics who treat take-or-pay contracts as debt equivalents make exactly this point.
• Non sequitur: price advantage is not ability to pay (P4) — Being priced below spot and forward compute prices bears on whether the contracts are a good deal. It does not bear on whether Anthropic can meet the payments. If compute prices fall, a locked-in rate that is cheap today becomes above-market. The claim is also asserted without a benchmark or disclosed contract terms.
• Credibility conflated with affordability (irrelevant support) (P1–P2) — Showing that the commitments are binding contracts, not memoranda, answers the charge that they are unserious. It does not show they are affordable, and being binding raises the downside if revenue falls short. These premises clarify Baker's earlier remarks but do little inferential work for the conclusion.
• Partial test applied to the whole (P3–P6 → Conclusion) — The cloud-versus-revenue test is applied to one component, and the other roughly 39% (the Broadcom portion) is handled by pointing to who finances it. Neither branch tests the combined obligation against a measure such as free cash flow coverage or leverage.
• Reliance on authority and unverified figures (P4–P5) — Several load-bearing claims (below-spot pricing, high inference margins, a $70 billion run rate above the last reported ~$65 billion) rest on a prominent investor's judgment, who may have exposure to the sector, and on a leaked draft prospectus. These are reasonable starting points but are not independently checkable.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — Non-cancellable take-or-pay commitments behave like fixed debt. For a loss-making company that needed $95 billion+ of outside capital in a single year, the 'manageable' verdict depends on three forecasts holding for 6–10 years: revenue keeps growing, margins stay high despite competition, and external financing stays available. The correct test is leverage and coverage under stress, not annual spend against today's run rate. Telecom and fiber build-outs around 2000 involved binding commitments by credible firms against strong projected demand, and many ended in…
• Premise 5 (High impact) — Run-rate revenue is an annualized snapshot, not recognized revenue or cash. Inference gross margin leaves out training, R&D and personnel. The relevant quantity is free cash flow after all of these, and the argument supplies none. The stated losses and heavy outside funding indicate that operations have not yet covered even current spend.
• Premise 6 (High impact) — Third-party financing relies on Anthropic's contracted payments, so the obligation remains Anthropic's in substance and adds leverage and covenants. Financing is plentiful when the AI story is strong and tightens in a downturn, which is exactly when a non-cancellable obligation bites. Financing by a counterparty such as Broadcom also raises circularity concerns.
• Premise 3 (Medium impact) — The $31 billion a year is an average. Compute contracts are generally ramped as capacity comes online, so peak-year payments could be considerably higher, and the three terms overlap only partly. Year-by-year schedules and any step-down or termination terms are needed before the average can be treated as the real annual burden.
• Premise 4 (Medium impact) — A fixed price below today's spot is a hedge only if scarcity persists. If hardware efficiency gains or new supply push compute prices down, 7–10 year commitments become above-market and leave Anthropic at a cost disadvantage. Obsolescence over contract tenor adds to this.
• Premises 1–2 (Medium impact) — Distinguishing signed contracts from memoranda establishes that the commitments are real, not that they are affordable. Binding obligations are more dangerous to the signer, so credibility and safety are separate properties.
• Conclusion (Medium impact) — Suppliers such as Amazon, Google and Microsoft are also investors in or partners of Anthropic, so some of the capital raised may return as cloud spend. This circularity and the concentration across a few counterparties are not addressed, and the argument also comes from an investor who may hold positions in the sector.
• Conclusion (Low impact) — The 'same test' for OpenAI only compares annualized obligations with run-rate revenue, which flatters any firm with a large run rate. OpenAI's commitments differ in structure (more memoranda and counterparty-financed arrangements), so a test built on this one may not transfer cleanly and could mask funding gaps for both firms.

Suggested Improvements

• Cash-flow basis for affordability — Test commitments against operating cash flow or free cash flow after training, R&D and payroll, and against gross margin on a recognized-revenue basis, rather than run-rate revenue and inference margin. The claim is that operations can fund the spend, which requires a cash measure. This also reconciles the claim with the reported losses and outside capital raised.
• Payment schedule and peak-year exposure — Use year-by-year commitment schedules, including ramps, minimum-spend escalators and termination or step-down terms, and report peak-year coverage next to the average. Solvency is determined by the worst years, and averaging hides back-loaded obligations.
• Stress testing — Add downside scenarios, such as revenue growth halving, margins compressing, and compute prices falling 30–50%, and show the resulting commitment-to-revenue and coverage ratios. A fixed non-cancellable obligation is a leveraged bet on growth, and the conclusion should be shown to hold when that bet goes wrong.
• Financing structure — Disclose recourse, guarantees, covenants, cross-default provisions and repayment dependence on Anthropic's credit for the Broadcom and private-credit financing, and estimate the cost of capital. Without these terms, 'externally financed' cannot be distinguished from 'Anthropic's obligation funded by someone else'.
• Evidence for pricing and margin claims — Cite benchmarks (GPU/TPU/Trainium hour rates and forward curves for matching hardware generations) and quantify inference margin, or drop the below-market claim as a support for manageability. Both are load-bearing assertions that are currently uncheckable, and below-market pricing is irrelevant to affordability unless prices hold.
• Define 'manageable' in advance — State thresholds, such as commitments to forward revenue or coverage ratios, and apply them identically to Anthropic and OpenAI once filings are public. Treat take-or-pay as debt-like in the leverage calculation. This makes the claim falsifiable, makes the 'same test' meaningful, and engages directly with the debt-equivalence critique.
• Verification and disclosure — Re-run the analysis against the public filing and disclose any positions the commentator holds in counterparties or in the AI sector. The figures come from a confidential draft, and disclosure lets readers weigh potential bias.

Scenario Tests

• Revenue continues to grow quickly, inference margins stay high, and compute prices stay elevated or rise. (Supports) — A roughly $31 billion annual cloud cost is a modest share of revenue and the below-market contracts look like a bargain. External financing remains available, so 'manageable' holds. This is the regime the argument implicitly assumes.
• Revenue growth slows sharply because of price competition, open-source commoditization or customer insourcing, while payments stay fixed. (Challenges) — Cloud spend consumes a far larger share of revenue, and the 'paid from operations' claim breaks. Anthropic would have to raise more equity at lower valuations or renegotiate, showing that the average-versus-run-rate test misses tail risk.
• Compute prices fall quickly as hardware efficiency improves and new capacity floods the market. (Challenges) — The locked-in rates become above-market, so the 'below spot' feature turns into a liability, and rivals with more flexible capacity gain a cost advantage.
• Credit markets tighten and private-credit appetite for AI infrastructure fades. (Challenges) — The $161 billion Broadcom-related financing becomes a direct liquidity problem for Anthropic. Because financing is pro-cyclical, it is least available when the obligation matters most.
• The public filing shows year-by-year payments that are roughly level, termination or step-down rights, and operating cash flow that covers cloud commitments. (Supports) — The annualized framing would be vindicated and the cash-flow test passed, making the 'manageable' conclusion considerably firmer.
• The public filing shows back-loaded, steeply ramped payments with limited flexibility. (Challenges) — The $31 billion average understates peak-year burden, and the argument's central normalization would no longer describe the actual obligation.
• OpenAI's public filing is tested using the same annualized-commitment-versus-revenue method. (Neutral) — The method is a reasonable consistency check, but it only informs if the test includes cash flow and stress scenarios. Differences in contract structure, pricing and financing between the two firms limit direct comparability.

Coherence & Relevance

The structure is clear and sensible. The headline is decomposed into a cloud branch (annualized and tested against revenue) and a Broadcom branch (externally financed), and the figures broadly reconcile to the $413 billion total. Coherence weakens at the inferences that carry 'manageable': the cloud branch tests against revenue instead of cash flow and ignores timing and downside risk, and the Broadcom branch reports who funds it, not whether Anthropic's exposure falls. The opening premises on binding versus non-binding commitments contribute context and credibility but do not address affordability, and the stated context of losses, heavy outside capital and debt-equivalence is acknowledged but not answered. The conclusion is plausible and defensible as a first-order screen, but the evidence supports 'serviceable while growth continues' more strongly than 'reasonable' in a robust sense. The closing recommendation for OpenAI is a fair consistency norm that the premises neither support nor undercut.

• His weekend comments about unserious commitments referred to more than $1 trillion of 2025 memoranda with Asian supply chain and semiconductor companies, not to signed contracts. (Weak) — This clarifies his earlier remarks and separates non-binding memoranda from signed contracts, but it does not bear on whether Anthropic's signed commitments are affordable.
• Contracts matter and are taken seriously. (Weak) — This is a framing statement. It establishes that the obligations are real, which raises the stakes, but provides no support for manageability.
• Per Anthropic's reported draft prospectus, its cloud commitments run about $11 billion a year with Amazon for a decade, about $15 billion a year with Google for seven years, and about $5 billion a year with Microsoft for about six years. (Strong) — These supply the cloud half of the decomposition. Missing are the payment schedule, ramp profile and termination terms, plus an account of the remaining few billion dollars that do not tie exactly to the headline.
• That is roughly $31 billion a year, priced below current spot and forward compute prices. (Moderate) — The sum is relevant to the annual burden. The below-market pricing speaks to deal quality, not ability to pay, and is unsupported and price-path dependent.
• Against a reported revenue run rate he puts near $70 billion and high inference margins, that cloud spend can be paid from operations. (Strong) — This is the main inferential step for the cloud branch, but it moves from revenue and gross margin to cash capacity without cash-flow data, training and R&D costs, or durability over the contract term.
• The roughly $161 billion of Broadcom-related commitments will be financed by Broadcom and private credit, which has begun. (Moderate) — This addresses who advances the cash for the Broadcom branch but not whether Anthropic's obligation, recourse or exposure to credit-market conditions is reduced.

View this argument on LogicFirst.ai