GAVIN BAKER: The AI supply chain does not treat roughly $1 trillion of unfunded OpenAI and Anthropic spending commitments as real demand

The Gist

Gavin Baker says the trillion dollars of AI spending pledges from OpenAI and Anthropic that have no money behind them are not taken seriously by the chip and cloud companies. Those companies build for deals that are paid for. This is a steelman: each side is reconstructed in its strongest fair form, not endorsed, and no winner is declared.

Conclusion

Unfunded headline commitments from the two leading labs are discounted by the suppliers expected to fulfill them, so those headline figures are not a reliable measure of either committed demand or supplier exposure.

Premises

1. OpenAI and Anthropic have announced roughly $1 trillion of spending commitments that are not backed by identified funding.
2. Suppliers plan capacity on commitments they can bank, not on headline pledges.
3. TSMC executives reportedly dismissed Sam Altman as a nontechnical "podcasting bro" after he pitched an outsized multi-year build, which shows how suppliers discount ambitious unfunded proposals.
4. Public-company executives, asked about these commitments, describe them in terms that show they are not underwriting capacity against them.

Assumptions

• "Unfunded" means lacking an identified financing source, not lacking legal form; Baker later states that some unfunded commitments may be non-cancellable.
• The TSMC anecdote dates from 2023 to 2024, concerns a different proposal (a reported $7 trillion fab plan), and rests on one unnamed source in a New York Times report.
• "Literally no one" is read as a strong generalization about supplier practice, not a census of every counterparty.

Analysis

Argument type: Inductive.

Potential Fallacies

• Hasty generalization / unrepresentative anecdote (P3 → P2 and conclusion) — One reported reaction from one supplier, about a different and far larger proposal, is used to illustrate how the whole supply chain treats today's commitments. A vivid single case shows that skepticism existed at one point. It does not show a general practice, and the personal insult says little about how contracts are assessed. TSMC is also an unusually conservative, capacity-constrained supplier and may not represent direct counterparties such as cloud, chip or neocloud firms.
• Equivocation on 'unfunded' / 'real demand' (P1, P2 and conclusion) — 'Unfunded' (no identified financing) slides toward 'not bankable' and then 'not real demand'. These are different properties. A commitment can lack identified funding and still be legally binding, which the author concedes. Suppliers may then rationally discount it as a credit risk while still being exposed to it. So discounting in planning does not show that exposure is absent.
• Inferring behavior from speech (P4) — Executives' public language is shaped by investor relations, legal caution and negotiating leverage, so it is a noisy signal of internal capacity planning. Because the statements are paraphrased rather than quoted, readers must also trust the author's reading of them. Capex, backlog, prepayment and purchase-obligation data would reveal underwriting far better.
• Argument from absence of evidence (P1) — 'Not backed by identified funding' is treated as 'not funded'. Financing may exist privately (vendor financing, future raises, strategic investors) without being publicly visible, and suppliers may know contract terms outsiders do not.
• Scope shift / conflating planning with exposure (Inference to conclusion) — The premises concern how suppliers plan capacity. The conclusion extends this to how reliable the headline figures are as a measure of supplier exposure. Not building against a commitment does not mean having no exposure to it.
• Treating a heterogeneous group as a single actor (P2, P4 and title) — 'The AI supply chain' covers foundries, memory makers, chip designers, hyperscalers, neoclouds and power and data-center firms, with very different contracts and incentives. Some have booked backlog or raised debt against these deals. Conservative upstream behavior is allowed to stand in for the whole chain, which makes the 'literally no one' framing easy to falsify with one counterexample.

Counterarguments

These challenges are not ranked.

• Conclusion (supplier exposure) (High impact) — Discounting in planning is compatible with real exposure. Some unfunded commitments may be non-cancellable (as A1 concedes), and suppliers may hold equity stakes, vendor financing, prepayments or hyperscaler-intermediated orders tied to lab solvency. If the labs failed to pay, suppliers would bear real losses. Headline figures could then be a meaningful, if imperfect, indicator of exposure, and the argument could even be turned into a contagion argument.
• Premise 2 and Premise 4 (High impact) — Revealed behavior matters more than rhetoric. Suppliers and direct counterparties such as Nvidia, AMD, Oracle, CoreWeave, Broadcom and the hyperscalers have reportedly expanded capex, raised debt, booked backlog or signed offtake and equity-linked deals explicitly tied to OpenAI and Anthropic. That suggests partial underwriting of the commitments, whatever executives say publicly.
• Premise 3 (Medium impact) — The TSMC anecdote concerns a different proposal, an earlier period and a fab build rather than a purchase contract. It reflects a reaction to an outsized ask and to the proposer, not a method for discounting contract commitments. Supplier engagement has changed as the labs' revenue and funding grew.
• Premise 1 (Medium impact) — 'Unfunded' at signing is the normal condition of multi-year growth-stage commitments, and funding often follows the demand signal. Announcements can help raise the capital that later funds them, so the discount may apply only to a snapshot. Sovereign capital, debt markets, strategic investors and revenue growth could convert unfunded commitments into funded ones.
• Title / 'literally no one' (Medium impact) — The universal framing is falsified by a single supplier with take-or-pay or prepayment exposure. Even as a strong generalization (A3), it invites an easy counterexample hunt and distracts from the more defensible core claim.
• Conclusion (Low impact) — Unreliable as a precise measure is not the same as uninformative. Headline commitments still signal intent, help secure allocation and priority, and influence financing and valuations. Suppliers may treat them as option-like, probability-weighted demand and not as zero.

Suggested Improvements

• Decompose the $1T — Break the total down by lab, counterparty, duration and contract type (firm or take-or-pay, cancellable, framework or MOU, funded or unfunded), and estimate the share that is non-cancellable. The strength of the conclusion depends on how much of the total is truly discretionary versus binding. Without this split the argument cannot distinguish discounted headlines from real exposure.
• Replace rhetoric with revealed behavior — Support P2 and P4 with supplier capex guidance, backlog or remaining performance obligations, prepayments, purchase obligations and specific dated, quoted earnings-call statements, compared against the commitments attributable to each supplier. Actions are better evidence of underwriting than tone. This would also address the strongest counterexamples, such as direct counterparties that have built capacity against these deals.
• Separate demand from exposure — Make two explicit claims: (a) headline totals overstate bankable demand, and (b) supplier exposure is limited. Support (b) with analysis of equity ties, vendor financing, non-cancellable terms and hyperscaler backstops. Planning behavior and balance-sheet exposure are different things. Separating them removes the equivocation and the risk of false comfort.
• Segment the supply chain — Distinguish upstream, capacity-constrained suppliers (foundry, memory, packaging) from direct counterparties (cloud, neocloud, chip vendors with financing ties), and say which the claim covers. The generalization is plausible for upstream manufacturers and weak for direct counterparties. Scoping it makes 'literally no one' defensible.
• Replace or demote the TSMC anecdote — Treat the TSMC story as illustrative color, or replace it with evidence about current supplier treatment of signed contracts, such as required prepayments or credit support. A dated, single-source reaction to a different proposal cannot carry a claim about present practice, and a personal insult invites a reputational reading instead of a contractual one.
• Make the claim dynamic and falsifiable — State predictions, such as supplier capex not rising in proportion to headline totals, and track whether commitments are funded, renegotiated or cancelled as they come due. 'Unfunded' is a snapshot. A time-bound, testable claim is more credible than a static one.

Scenario Tests

• Major direct counterparties (e.g., Oracle, CoreWeave, Nvidia) disclose capex increases, debt raises or backlog explicitly tied to OpenAI and Anthropic contracts. (Challenges) — Suppliers would be underwriting at least part of the commitments. P2 and P4 would fail for those firms, and the claim would be limited to upstream suppliers. Supplier exposure would be real despite any hedged public language.
• Supplier capex guidance and backlog disclosures stay well below the headline totals, and suppliers decline to build ahead without prepayments or credit support. (Supports) — This is the behavioral evidence the argument needs. It would show headline figures overstate bankable demand and that suppliers are discounting in practice, not only in speech.
• A large share of the $1T turns out to be contractually non-cancellable, with suppliers holding equity or vendor-financing positions in the labs. (Challenges) — Exposure would be significant even if suppliers discount the probability of full realization. The conclusion would hold narrowly for 'committed demand' but fail for 'supplier exposure', and headline figures would be more informative than claimed.
• The labs raise large new rounds, sovereign capital or debt that fund most of the commitments within a year. (Challenges) — 'Unfunded' proves a temporary snapshot and the discount shrinks. Headline totals look more like leading indicators of eventual demand, though the earlier supplier skepticism may still have been rational at the time.
• A funding freeze hits the labs, and commitments are renegotiated, scaled back or cancelled with limited losses to suppliers. (Supports) — This would vindicate the view that headline totals overstate both demand and exposure, and show that suppliers had priced in the risk.
• A funding freeze hits the labs, and non-cancellable obligations surface as significant losses for suppliers. (Challenges) — The skeptical view of funding would be right, but the reassurance about limited supplier exposure would be refuted. Discounting in public would have been insufficient protection.

Coherence & Relevance

The core idea is coherent and probably directionally right: suppliers plan on bankable orders, headline totals blend firm and aspirational items, and so headline totals overstate real demand. Coherence weakens where the argument moves from 'suppliers discount' to 'headline figures do not measure supplier exposure', since the author's own concession that some commitments are non-cancellable shows exposure can exist regardless of supplier planning. The evidence is also uneven. The one named-supplier anecdote is stale and off-point, and the more relevant claims about executives are paraphrased, not shown. The argument is best read as a plausible inductive case for a modest claim (headline totals are an imperfect proxy), not as proof of the strong claim its title suggests.

• OpenAI and Anthropic have announced roughly $1 trillion of spending commitments that are not backed by identified funding. (Strong) — Establishes the object of the claim but not suppliers' response. The aggregate is not decomposed by contract type or lab. 'Unfunded' means lacking an identified source, which is not the same as non-binding or non-bankable.
• Suppliers plan capacity on commitments they can bank, not on headline pledges. (Strong) — This is the key bridging principle. It needs the further premise that unfunded commitments are non-bankable and that supplier planning tracks real demand and exposure. 'Bankable' is undefined, and some unfunded commitments may be non-cancellable.
• TSMC executives reportedly dismissed Sam Altman as a nontechnical "podcasting bro" after he pitched an outsized multi-year build, which shows how suppliers discount ambitious unfunded proposals. (Weak) — Differs in time, proposal and kind (fab build versus compute contracts) from the commitments at issue, rests on one unnamed source, and speaks to attitude toward a person and an extreme ask, not contract bankability. One upstream supplier cannot represent the whole chain.
• Public-company executives, asked about these commitments, describe them in terms that show they are not underwriting capacity against them. (Strong) — Directly relevant if substantiated, but no speakers, dates or quotes are given. It infers planning from public tone and is partly circular, since the interpretation 'not underwriting' is close to the conclusion itself. It also says nothing about exposure through equity, financing or binding terms.

View this argument on LogicFirst.ai