GAVIN BAKER: Shortages of power, wafers and capital are slowing the AI buildout, and that reduces the larger risk of an actual overbuild

The Gist

Gavin Baker says it is good that the world is short of power, chips and money for AI, because that stops companies from building too much. The bigger near-term worry is a major cyberattack. This is a steelman: Gavin Baker's thread is reconstructed in its strongest fair form, not endorsed.

Conclusion

Constraints on power, wafers and capital make the AI buildout slower and smoother, which lowers the risk of overbuilding and is better for investors and the public over the long term; a severe cyber event is the more plausible near-term risk.

Premises

1. The worst outcome for AI infrastructure would be an actual overbuild.
2. The industry is structurally short of power, wafers and, increasingly, capital.
3. Those constraints slow and smooth the buildout.
4. With zero interest rates, much larger foundry expansion and no equipment shortages, an overbuild and bubble would be likely.
5. A severe cyber event is the most plausible near-term bear case, though AI cyber defense is improving.

Assumptions

• Capital scarcity also raises the cost of financing for weaker borrowers, as 2026 credit spreads showed.
• The claim that cyber may become defense dominant is a cited expert view, not an established result.

Analysis

Argument type: Inductive.

Potential Fallacies

• Equating delay with risk reduction (P3 → Conclusion) — Slowing a buildout lowers the pace of deployment, but overbuild depends on how capacity compares with realized demand. If constraints only defer capacity, or if capacity arrives together once bottlenecks ease, the same endpoint can be reached later and perhaps more abruptly. The argument needs a premise showing that constraints cap eventual capacity or that they bind persistently.
• Non sequitur (welfare leap) (Conclusion) — The premises concern overbuild probability. The conclusion that constraints are better for investors and the public needs further premises on opportunity costs of delayed AI deployment, electricity and compute prices, concentration of advantage in incumbents, and distress among weaker borrowers (which A1 acknowledges). None of these is weighed.
• Unsupported counterfactual / conflating relative and absolute risk (P4) — P4 shows that overbuild would be likely in a hypothetical unconstrained world. That supports a claim that risk is lower now, not that it is low. The counterfactual also holds demand fixed and cannot be tested.
• Unquantified comparative ranking / severity-probability mismatch (P1 and P5) — Overbuild is called the 'worst' outcome (severity) and cyber the 'most plausible' (likelihood), with no common metric such as probability times impact. Calling cyber 'most plausible' also requires comparison with credit stress, demand or ROI disappointment, and geopolitical shocks, and none is offered.
• Silver-lining framing (P2–P3 and Conclusion) — Shortages are normally read as costs. Recasting them as protective can be correct, but the argument shows the benefits while leaving out the costs, which makes it hard to falsify: if no bust occurs, constraints get the credit, and if one occurs, it is attributed to something exogenous.
• Appeal to authority (mild) (P1, P5 and A2) — Much of the support rests on the standing of a prominent investor and on a cited expert view about cyber defense. The author is open that the latter is not established, but neither claim is backed by data. The author is also a market participant whose incentives may favor a reassuring narrative.

Counterarguments

These challenges are not ranked.

• Premise 3 (High impact) — Constraints delay and concentrate risk instead of reducing it. Shortages inflate prices and margins and encourage leveraged commitments, double-ordering and circular financing. When bottlenecks clear, often simultaneously, capacity may arrive together into softer demand. Semiconductor and fiber cycles show shortage phases followed by gluts.
• Conclusion (High impact) — Capital scarcity (A1) already stresses weaker borrowers such as neoclouds and data-center developers. Defaults and forced asset sales could produce a bust without any aggregate overbuild, so the same constraint that 'smooths' the buildout can be the trigger for a credit-driven correction.
• Premise 1 (Medium impact) — Overbuild may not be the worst outcome. Underbuild that delays productivity gains, concentration of compute among a few incumbents, grid strain, or a leveraged financial crisis could be worse. Surplus capacity can also leave cheap compute with real social value.
• Premise 3 (High impact) — Overbuild risk is driven by demand relative to capacity. If AI monetization disappoints or efficiency gains cut compute needs, even constrained supply can exceed profitable demand, and the slow pace offers no protection.
• Conclusion (Medium impact) — The claim of public benefit is unsupported. Scarcity raises compute and electricity costs, delays beneficial applications, and shifts value to suppliers and well-capitalized incumbents. Investor and public interests are asserted to be aligned, not shown to be.
• Premise 5 (Medium impact) — Demand or ROI shortfall, credit stress, geopolitical disruption around Taiwan, regulatory backlash and power politics may each be at least as plausible a near-term bear case as a severe cyber event. Offense may also keep an advantage over defense, which would raise cyber risk without bearing on the main thesis.

Suggested Improvements

• Define and operationalize 'overbuild' and 'smooth' — Specify overbuild as sustained utilization or compute rental prices below cost of capital over a stated horizon, and smoothness as lower volatility in capex growth, and give asset-specific views (GPUs vs. power and shells). This makes the thesis testable and recognizes that fast-depreciating accelerators carry different overbuild costs than long-lived infrastructure.
• Add demand-side evidence — Include data on AI revenue growth, contracted backlog, utilization, and the share of spending funded by vendors or debt. Overbuild is a supply-demand gap, and shortages alone do not show that demand is durable.
• Address the timing and persistence of constraints — Explain how long power, wafer and capital constraints are expected to bind, whether they ease at different times, and why added capacity would not arrive in a synchronized surge. It would answer the strongest objection, that smoothing is temporary and risk is deferred.
• Weigh the costs of constraints — Integrate A1 into the analysis by assessing credit stress among weaker borrowers, concentration, higher prices, and the opportunity cost of slower deployment, and then state a net judgment. The welfare conclusion for investors and the public needs to be derived from a full accounting.
• Build a common risk framework — Compare overbuild, credit stress, demand shortfall, geopolitical disruption and cyber on probability and severity over a stated horizon. It would justify the claim that cyber is the most plausible near-term bear case and resolve the severity-versus-likelihood mismatch.
• Support the historical analogy — Compare the current cycle to constrained and unconstrained historical buildouts (fiber, semiconductors, shale, cloud), including cases where scarcity coexisted with a bust. It would test whether constraints have ever dampened overbuild, instead of relying on a favorable counterfactual.

Scenario Tests

• Power, advanced packaging and fab capacity expansions all come online in the same 12–24 month window while AI revenue growth flattens. (Challenges) — The brake disappears just as demand softens, and the delayed capacity produces a sharp overshoot. This is the central failure mode for the 'slower and smoother' claim.
• Constraints remain binding for several years, hyperscaler capex tracks monetization, and rental prices for compute stay firm. (Supports) — Rationed supply keeps capacity near demand and the buildout proceeds without a glut, consistent with the thesis, though it would not show that constraints were the cause.
• A wave of defaults among leveraged neoclouds and data-center developers follows widening credit spreads, with spillover to private credit lenders. (Challenges) — Capital scarcity acts as the trigger of a bust, and the cyber event is not the most plausible near-term bear case. It uses the argument's own A1.
• A zero-rate environment returns and foundry and equipment supply expands rapidly with no shortages. (Supports) — This matches P4's counterfactual and the historical pattern of telecom and ZIRP-era excess, supporting the claim that constraints are a meaningful safeguard.
• A major AI-enabled cyberattack disrupts a concentrated hyperscaler or foundry node. (Supports) — It would validate cyber as a significant near-term tail risk, though it would show little about the constraints thesis, and the concentration that scarcity encourages could itself make the target more attractive.
• Efficiency breakthroughs sharply reduce compute needed per task. (Challenges) — Demand falls relative to even constrained capacity, so overbuild occurs regardless of supply pace and shows that supply speed is not the main determinant.

Coherence & Relevance

The argument has two loosely joined parts. The first is a coherent contrastive case that scarcity reduces the pace of buildout and hence some overbuild risk relative to an unconstrained world. It is plausible but incomplete: it lacks a demand-side analysis, does not address whether constraints persist, and does not weigh the costs of scarcity that A1 itself concedes. The step from lower overbuild risk to being better for investors and the public is not supported by the premises. The second part, the cyber claim, is an add-on that neither supports nor depends on the first and is asserted without comparison to other risks. Overall the argument is a reasonable, thoughtful framing of the matter that needs additional premises to carry its full conclusion.

• The worst outcome for AI infrastructure would be an actual overbuild. (Moderate) — It sets the stakes but is a value claim. It does not show that overbuild is worse than other risks, including the cyber risk in P5, and it is not tied to public welfare.
• The industry is structurally short of power, wafers and, increasingly, capital. (Strong) — It is the factual foundation of the argument, but 'structural' is not distinguished from cyclical, and the shortage is not linked to the soundness of end-demand.
• Those constraints slow and smooth the buildout. (Strong) — It is the key causal step. It does not show that slower means lower overbuild risk, and it omits timing effects, over-ordering and the credit stress acknowledged in A1.
• With zero interest rates, much larger foundry expansion and no equipment shortages, an overbuild and bubble would be likely. (Moderate) — It establishes a contrast case, but it holds demand fixed, is not quantified, and shows only relative risk reduction.
• A severe cyber event is the most plausible near-term bear case, though AI cyber defense is improving. (Weak) — It is logically independent of the constraints argument and lacks a comparison set. It uses a different risk dimension (plausibility) than P1 (severity).

View this argument on LogicFirst.ai