GAVIN BAKER: AI suppliers price risk on funded contracts, not headline pledges, so roughly $1 trillion of unfunded commitments is not the key AI infrastructure risk

The Gist

Gavin Baker says chip and cloud suppliers build for deals that are signed and paid for, and they ignore loose pledges. So the trillion dollars of unfunded promises is not the thing to worry about. This is a steelman: each side is reconstructed in its strongest fair form, not endorsed, and no winner is declared.

Conclusion

The AI infrastructure supply chain plans and prices risk around funded, contracted commitments and aggregate token demand, so unfunded headline pledges are neither reliable demand nor the central risk, and lower model-layer margins need not reduce infrastructure demand.

Premises

1. Lower revenue or margin at the model layer does not imply lower spending at the infrastructure layer; a market shifting toward open-weight models and cheaper tokens is more likely to raise infrastructure demand than to lower it, and a two-lab duopoly would be the least favorable outcome for infrastructure suppliers.
2. 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.
3. Pointing to funded deals does not rebut the claim that unfunded commitments are discounted, because funding, not announcement size or legal form alone, is what the supply chain treats as decisive.
4. Commitment totals should count only contracted or funded obligations; non-binding memoranda inflate headline figures and do not represent promises to pay.
5. A bear case built on unfunded memoranda overstates the risk those memoranda carry, and the more defensible bear cases rest on funded exposure and operating economics.

Assumptions

• The claim concerns supplier capacity planning and infrastructure demand, not whether OpenAI or Anthropic can meet their own obligations.
• Baker manages Atreides Management, which invests in technology and AI-related companies; no specific holding was disclosed in these posts.
• A large share of Anthropic's commitments is non-cancellable, which narrows how much of the $1 trillion fits the unfunded memorandum category.

Analysis

Argument type: Inductive.

Potential Fallacies

• Non sequitur (from 'discounted' to 'not the central risk') (Inference from P2, P4 and P5 to the conclusion) — Showing that suppliers discount unfunded pledges shows those pledges are a poor demand signal. It does not show that the risk is peripheral. Discounting may be partial, and funded exposure may depend on the same pledges and the same capital-market sentiment. An added premise linking risk to what suppliers actually price is needed.
• Equivocation among 'unfunded', 'non-binding', 'uncontracted' and 'headline' (P2, P3, P4 and A3) — These are different dimensions. A commitment can be binding but unfunded, or funded but cancellable. A3 concedes that the categories do not line up neatly, yet the argument treats them as one group, so the strong claim rides on the weaker evidence.
• Circular support / question-begging (P3, P4 and P5) — P3 defines funding as 'decisive' for the supply chain, which restates the thesis and pre-empts counterexamples. P4 and P5 largely restate P2. They count as roughly one independent claim, not three. P4 also defines commitment totals so that the headline overstatement is true by construction.
• Rational-actor idealization (is/ought slide) (P2 and P4) — That suppliers should discount unfunded pledges does not establish that they do. Competitive pressure, fear of missing out, vendor financing and long-lead-time capex can lead them to build ahead of firm demand. The argument also moves between what suppliers do and what analysts should count without marking the shift.
• Decoupling of linked layers (P1) — Model-layer margins may not reduce volume demand for compute. But the labs are the marginal buyers, and their margins affect their ability to raise capital and pay. The argument conflates usage volume with solvent, funded demand.
• Narrowed or partial straw man of the bear case (P5) — The bear case is characterized as resting on unfunded memoranda. Stronger versions concern circular financing, counterparty concentration, depreciation and funding dependence, and those are not engaged. Conceding that 'more defensible' bears exist narrows the debate to the author's terms.
• Appeal to authority / unsupported empirical assertion (P2, P3 and the title framing) — Claims about how suppliers price risk rest on the speaker's investor standing, with no cited capex guidance, backlog data, contract terms or financing terms. The speaker's interest in the AI sector is disclosed but not examined.
• Treating a recent trend as an elasticity law (P1) — Rising token volume alongside falling prices is consistent with high elasticity, but it is a historical pattern and not guaranteed. If elasticity falls below 1 or efficiency gains outpace usage, cheaper tokens reduce aggregate spend. The duopoly-as-worst-case claim is also asserted without support.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — The central risk may be the durability of funding itself. Build-out is financed by a narrow, sentiment-dependent loop involving lab fundraising, supplier equity stakes and vendor financing. 'Funded' contracts are only as sound as the counterparty's continued access to capital, so the risk moves one layer down instead of disappearing. Telecom and fiber vendor-financing busts are the precedent.
• Premise 1 (High impact) — Model-layer margin compression can reduce the labs' ability to fund and pay for compute even if usage volume rises. Infrastructure could see rising volume alongside deteriorating utilization, returns and counterparty credit. If elasticity is below 1, or efficiency gains outpace usage growth, cheaper tokens reduce aggregate infrastructure spend.
• Premise 2 (High impact) — Suppliers may not discount headline pledges in practice. Capex raises, long-lead orders for memory and power, vendor financing, equity stakes and debt raised against announcements would show suppliers building on the headlines. Headline pledges also help labs raise the capital that turns pledges into funded contracts, so they affect real exposure even if formally discounted.
• Premise 3 (Medium impact) — The funded/unfunded line is not fixed. Funding status is endogenous to sentiment and fundraising, so today's unfunded pledge may become funded and today's funded contract may be stressed. Treating funding as decisive for the whole supply chain also ignores that foundries, memory makers, neoclouds and power developers differ in credit exposure and bargaining power.
• Premise 4 (Medium impact) — Non-binding memoranda carry signaling and option value and may precede binding orders. Counting them at zero understates expected demand, just as counting them at face value overstates it. Contract terms form a continuum of cancellability and renegotiation likelihood.
• Premise 5 and Assumption A3 (High impact) — If a large share of Anthropic's commitments is non-cancellable, that share is funded, binding exposure and sits at the center of the bear case. This shrinks the 'unfunded memoranda' category without reducing the risk, and it shifts the question to whether the counterparty can pay, which A1 sets aside.
• Conclusion (Low impact) — The source is an investor in AI-related companies, and the bullish conclusion matches that interest. This does not refute the argument, but the key behavioral claims are unsupported by primary data, so they need independent corroboration.

Suggested Improvements

• Quantify the composition of the $1 trillion — Break the headline total into binding and funded, binding but contingently funded, and non-binding components by lab, using contract and filing disclosures. Show how A3's non-cancellable share changes the unfunded portion. The claim that the $1 trillion is not the key risk cannot be assessed without knowing how much of it is actually unfunded and non-binding.
• Provide supplier-side evidence — Cite backlog definitions, capex guidance, prepayments, take-or-pay terms and financing structures that show how suppliers treat funded versus unfunded deals, separating large diversified firms from leveraged neoclouds and developers. P2 is the factual keystone and currently rests on testimony.
• Close the gap between 'discounted' and 'not central' — Add an explicit premise tying risk to the exposure suppliers actually price, and address the funded-exposure risks (counterparty concentration, funding durability, depreciation) head-on. This repairs the main non sequitur and prevents the thesis from winning by redefinition.
• Address the financing channel in P1 — Explain how lower model-layer margins would affect lab fundraising and ability to pay. Separate claims about volume demand from claims about funded, solvent demand, and support elasticity with price and volume data. This removes the strongest objection to the decoupling claim and ties the margin argument to the funded/unfunded framework.
• Engage the strongest bear cases — Take on circular financing, counterparty concentration and depreciation-schedule arguments directly, instead of defining the bear case as reliance on memoranda. Rebutting the weakest version leaves the real debate untouched and invites a straw-man charge.
• Treat funding and conversion probabilistically — Replace the binary funded/unfunded split with probability-weighted exposure that allows partial discounting and conversion of memoranda into contracts, and state what evidence would show that suppliers do not discount. This is more realistic and makes the claim falsifiable.

Scenario Tests

• A major lab's fundraising fails or capital markets tighten sharply after supplier capacity has been built against its funded contracts. (Challenges) — The funded/unfunded line dissolves because funded commitments depend on future raises. Supplier exposure becomes real through receivables and renegotiation, which shows the central risk sits in funding durability and not in the headline memoranda.
• Token prices fall steeply as open-weight models spread, and total token volume and compute purchases rise by more than the price decline, funded by diverse buyers. (Supports) — This matches the P1 mechanism. Margin compression at the model layer coexists with rising infrastructure demand, and a diffuse buyer base is better for suppliers than a duopoly.
• Efficiency gains (distillation, sparsity, edge inference) outpace usage growth, so total compute demand per unit of value falls as tokens get cheaper. (Challenges) — Elasticity below 1 reverses the P1 prediction, and lower model-layer margins would then reduce infrastructure demand.
• Supplier filings show that capex, long-lead orders and debt issuance track contracted backlog and prepayments, and that unfunded announcements produce no change in guidance. (Supports) — This directly corroborates P2 and P3 for large suppliers, leaving the question of funded-exposure risk.
• Suppliers expand capacity, take on debt or invest equity in labs immediately after non-binding announcements, as in vendor-financed telecom buildouts. (Challenges) — Headline pledges would then be shaping real supplier exposure despite formal discounting, and the claim that they are not a measure of supplier exposure would fail.
• Disclosure shows that most of the $1 trillion is non-cancellable but depends on the labs' continued fundraising. (Neutral) — The 'unfunded memoranda' framing shrinks, which fits A3, but the counterparty-durability risk grows. The thesis holds only in a narrow supplier-planning sense.

Coherence & Relevance

The argument is a cluster of related claims, not a single chain. P2, P4 and P5 converge on the narrow point that non-binding headline pledges are poor measures of committed demand, and P1 independently supports the claim that lower model-layer margins need not lower infrastructure demand. The narrow claim is coherent and reasonably well supported. The leap to 'not the central risk' is not secured: it needs a premise about where risk sits, and the argument itself concedes that funded exposure and operating economics are the more defensible battleground. Funding is endogenous to model-layer economics, so P1 and the funded/unfunded framework are less separable than the argument treats them. Stated Assumptions A1 and A3 narrow the scope and the unfunded category, but they do not close this gap.

• Lower revenue or margin at the model layer does not imply lower spending at the infrastructure layer; a market shifting toward open-weight models and cheaper tokens is more likely to raise infrastructure demand than to lower it, and a two-lab duopoly would be the least favorable outcome for infrastructure suppliers. (Moderate) — Supports only the last conjunct of the conclusion. It is independent of the funded/unfunded line of argument, and it does not address how margins feed back into funding capacity.
• 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. (Strong) — Directly supports the claim that unfunded pledges are unreliable demand. It does not establish the extent of discounting, and it does not by itself show that those pledges are not the central risk.
• Pointing to funded deals does not rebut the claim that unfunded commitments are discounted, because funding, not announcement size or legal form alone, is what the supply chain treats as decisive. (Moderate) — Defends the thesis against a rejoinder and is partly circular. It does not supply independent evidence and assumes funding is a stable category.
• Commitment totals should count only contracted or funded obligations; non-binding memoranda inflate headline figures and do not represent promises to pay. (Moderate) — A measurement convention that makes the headline overstatement true by definition. It mixes a normative accounting rule with a descriptive claim and does not address the economic effects of non-binding instruments.
• A bear case built on unfunded memoranda overstates the risk those memoranda carry, and the more defensible bear cases rest on funded exposure and operating economics. (Moderate) — Mostly restates P2 and P4. It concedes where the real risk may lie without sizing it, and it characterizes the opposing case narrowly.

View this argument on LogicFirst.ai