GAVIN BAKER: Unfunded memoranda are a weak focus for the AI bear case when stronger bear cases are available
The Gist
Gavin Baker says skeptics are worried about the wrong thing. If nobody builds for the unfunded promises, those promises cannot sink the suppliers. There are better reasons to be bearish. This is a steelman: each side is reconstructed in its strongest fair form, not endorsed, and no winner is declared.
Conclusion
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.
Premises
- Skeptics often anchor on the roughly $1 trillion of unfunded commitments as a central systemic risk.
- If suppliers do not build against those commitments, their failure would not strand much supplier capacity.
- Other risks, such as funding depth, pricing, utilization and power, are more realistic channels of loss.
Assumptions
- Baker acknowledges that realistic bear cases exist; the claim ranks risks rather than dismissing risk.
- Baker said he did not expect the original post to be controversial; that response is a matter of perception, not evidence.
Analysis
Argument type: Inductive.
Potential Fallacies
- Undischarged conditional (unasserted antecedent) (Premise 2) — P2 says that if suppliers do not build against the commitments, little capacity is stranded. That is nearly true by definition, but the argument never establishes that suppliers actually refrain from building. The conclusion therefore leans on a factual premise that is assumed rather than shown, and whether suppliers pre-build is the central empirical dispute.
- Possible straw man (Premise 1) — 'Skeptics often anchor' on the $1T figure summarizes an unnamed group by their simplest argument. Many bears use the figure to point to circularity, concentration or financing dependence, not to a direct stranded-asset loss. If so, the rebuttal addresses a weaker version of the opposing case.
- Unsupported comparative claim (Premise 3 and conclusion) — Calling the other channels 'more realistic' or 'more defensible' gives no metric, magnitude or probability. The ranking rests mostly on the author's standing and judgment, so it is asserted more than demonstrated.
- Treating coupled risks as independent (Premise 3 and conclusion) — Funding depth, utilization and pricing may be downstream of the commitment pipeline. Unfunded commitments can shape how much gets built, how it is financed, and how counterparties are valued. Ranking them as separate buckets can understate how one feeds the other, and funding depth is the variable that decides whether memoranda ever become funded.
- Narrowed loss channel (non-binding treated as non-consequential) (Premises 2 and 3) — The argument evaluates unfunded commitments only through stranded supplier capacity. It leaves out signaling, valuation, credit-access and sentiment effects, through which announced but unfunded commitments can still cause losses.
Counterarguments
These challenges are not ranked.
- Premise 2 (High impact) — Soft demand signals have historically produced hard overbuilds, as in the 1999-2001 telecom and fiber cycle and in crypto-mining hardware. Memoranda are exactly what induces suppliers to pre-build, pre-order and borrow. Assuming restraint in a boom assumes the behavior booms tend to violate, so the conditional may fail where risk is greatest.
- Conclusion (High impact) — Unfunded commitments are a transmission mechanism, not a separate weaker risk. Lenders, equity investors and suppliers underwrite funded capex against the headline pipeline. Credible doubt about the memoranda can therefore tighten funding and compress valuations, which are among the 'realistic' risks the argument favors.
- Premise 3 (Medium impact) — Funding depth determines whether unfunded commitments ever become funded, so the 'stronger' bear case contains the 'weaker' one rather than competing with it. A ranking by separate channels dissolves once the channels are coupled.
- Premise 1 (Medium impact) — Informed bears often use the $1T figure to illustrate circular financing and counterparty concentration, and they focus on funded exposure such as vendor financing, neocloud leverage and depreciation. If so, the target is a popularized caricature, and the real debate is not between 'unfunded' and 'funded' cases.
- Conclusion (Low impact) — The source is a prominent investor with exposure to AI infrastructure, and redirecting critics toward more manageable risks serves his interests. This does not make the claim false, but it calls for independent verification of the ranking.
Suggested Improvements
- Assert and evidence the antecedent of P2 — State explicitly that suppliers are not building ahead of these commitments, and support it with capex, backlog, inventory, take-or-pay and prepayment data for the major suppliers. The conclusion rests on supplier behavior that is currently assumed. Evidence would turn a conditional into a usable premise.
- Decompose the $1T — Break the figure down by counterparty, instrument, bindingness and milestone structure, and show what share is truly non-binding. P2 only applies to the genuinely non-binding portion. Without this split, the claim could be over- or under-applied.
- Address signaling and reflexive channels — Add a premise explaining why valuation, credit-access and sentiment effects of failed memoranda are small, or concede them as a separate risk. This closes the gap between 'little stranded supplier capacity' and 'little overall risk'.
- Quantify the ranking — Give rough expected-loss or exposure estimates for funding depth, pricing, utilization and power against the unfunded channel, using GPU rental prices, financing spreads, utilization and interconnection data. A comparative claim needs a stated criterion and magnitudes to be persuasive.
- Engage the strongest skeptic version and align terms — Quote specific skeptics and rebut their best form of the argument. Map the P3 channels explicitly onto the conclusion's 'funded exposure and operating economics', and address how the two bear cases interact. This avoids straw-man concerns and removes term drift between premises and conclusion.
Scenario Tests
- A large lab fails to fund its commitments, and the suppliers involved had only attended announcements and signed non-binding frameworks without building. (Supports) — Little capacity is stranded and direct losses are small, which is the case where P2 holds. Equity and sentiment effects may remain.
- Neoclouds and data-center developers borrow heavily and order long-lead equipment, power and land against announced but unfunded deals. Funding then tightens and the commitments lapse. (Challenges) — Capacity is stranded and funded debt is impaired, so unfunded exposure converts into funded loss. This is the scenario P2 assumes away.
- GPU rental prices fall and utilization drops while spending is fully funded and built. (Supports) — Realized losses come from operating economics and funded capacity, consistent with P3 and the telecom precedent.
- Doubt about the commitment pipeline causes a drop in valuations and a widening of credit spreads, even though no physical capacity is stranded. (Challenges) — Unfunded memoranda create loss through market and financing channels, so the risk cannot be dismissed on the stranded-capacity logic alone. It also shows the channels are coupled.
- The claim is read in its modest form, as a ranking of risks that concedes realistic bear cases exist. (Supports) — In this form the conclusion is defensible and needs only relative, not absolute, support, though a ranking criterion is still missing.
Coherence & Relevance
The argument has a clear structure: identify a popular bear case, show that its main loss channel is weak if suppliers hold back, and redirect attention to channels that are more directly tied to money spent and operating returns. Read as a modest ranking, as the stated assumptions intend, it is coherent and plausible. The main gaps are the unasserted antecedent in P2, the lack of a bridge from 'little stranded capacity' to 'little overall risk', and the underlying coupling between the unfunded and funded risks. The argument is plausible but under-evidenced, and it would be considerably stronger with supplier-behavior data and a quantified comparison.
- Skeptics often anchor on the roughly $1 trillion of unfunded commitments as a central systemic risk. (Moderate) — It identifies the target of the critique but does not bear on whether the risk is overstated. It also needs a premise showing that stranded supplier capacity is the main loss channel skeptics have in mind.
- If suppliers do not build against those commitments, their failure would not strand much supplier capacity. (Strong) — This is the core of the 'overstates' claim, but its antecedent is not asserted or evidenced. It addresses stranded assets only, not sentiment, financing or valuation channels, and it treats supplier behavior as static.
- Other risks, such as funding depth, pricing, utilization and power, are more realistic channels of loss. (Strong) — It supports the second half of the conclusion, but the ranking has no criterion or magnitudes. The channels named do not exactly match 'funded exposure and operating economics', and they are not independent of the commitments.