ED ZITRON: Demand for open-weight models has not been shown to be large enough to justify current AI infrastructure spending

The Gist

Ed Zitron asks whether anyone has seen enough demand for free models to justify the huge data center spend. He implies no one has, and that the theory needs evidence. This is a steelman: each side is reconstructed in its strongest fair form, not endorsed, and no winner is declared.

Conclusion

Unless open-weight demand is shown at a scale comparable to current infrastructure spending, it cannot carry the justification for that spending.

Premises

1. Baker's case that open-weight models support infrastructure requires substantial paying demand for those models.
2. Open-source models were reported to be under 5% of business AI spend.
3. Infrastructure spending in 2026 runs at hundreds of billions of dollars a year.
4. A claim that a demand source justifies that spending should be supported by evidence of demand at comparable scale.

Assumptions

• The post is a question; it is steelmanned as the implied claim that the evidence is missing.
• Baker's reply chain does not answer the question directly; rising GPU rental prices bear on total demand, not on open-weight demand specifically.

Analysis

Argument type: Inductive.

Potential Fallacies

• Proxy/measurement mismatch (Premise 2 (to the conclusion)) — Share of business spend on models is used as a stand-in for open-weight demand on infrastructure. Because open-weight models cost little or nothing to license, their footprint appears in compute, hosting, and cloud budgets rather than in model spend. A low spend share can therefore coexist with a large compute share.
• Part-whole comparison / all-or-nothing standard (Premises 3 and 4) — Requiring one demand source to match total infrastructure spending treats justification as non-additive. A source can justify part of the spending alongside others without covering the whole. Baker's claim may only be that open-weight models support the spending, not that they alone justify it.
• Modal shift / drift from 'not shown' to 'cannot' (Conclusion) — The premises support the claim that the case is currently unsupported by evidence. The conclusion's 'cannot carry the justification' reads as a stronger claim about incapacity. The conditional wording ('unless shown') largely limits this drift, so the effect is mild.
• Argument from absence of evidence (mild) (Conclusion and Assumptions 1-2) — Missing evidence of large demand is treated as a reason to doubt it. This is a legitimate burden-of-proof move, but absence of evidence is weak evidence of absence when the relevant usage is decentralized and hard to observe, and when the lack of a reply comes from a social media thread.
• Terminological conflation (Premise 2 versus title and conclusion) — The evidence concerns 'open-source' models while the claim concerns 'open-weight' models. The categories overlap heavily but are not identical, and the argument does not flag the shift.

Counterarguments

These challenges are not ranked.

• Premise 2 (High impact) — Open-weight models are free to download, so their demand shows up as GPU rental, inference-hosting revenue, and self-hosted compute rather than as model spend. A low share of model spend says little about their share of tokens or GPU-hours. Usage data from inference hosts and routing platforms could show a far larger workload share.
• Premise 4 (High impact) — Infrastructure is justified by the aggregate of many demand streams (closed APIs, enterprise, consumer, agents), so open-weight demand need only contribute. Requiring any single source to match total spending would make almost every source, and almost every early-stage buildout, fail the test.
• Premise 1 (Medium impact) — Baker may not have claimed that open-weight models alone justify the spending, or that they do so through direct paying demand. If the claim was narrower, the argument attacks a position Baker does not hold.
• Conclusion (Medium impact) — Capex is built ahead of demand and justified by projections, so a snapshot of current spend share does not test an investment thesis. Open-weight share may be growing quickly.
• Assumption 2 (Low impact) — Rising GPU rental prices are not fully separable from open-weight demand, since rented GPUs are where many open-weight users buy compute. The signal is not proof, but it may be weak indirect evidence rather than irrelevant.

Suggested Improvements

• Metric choice — Replace or supplement the share-of-business-spend figure with compute-based measures: token-volume share, GPU-hours for open-weight inference and fine-tuning, and revenue of hosts serving open models. This measures the quantity that matters for infrastructure and answers the strongest objection to P2.
• Standard of proof — Replace 'comparable scale' with a defined threshold, such as open-weight-attributable compute spend as a share of annual capex, and state that the test is for sufficiency rather than any contribution. A defined threshold makes the claim falsifiable and avoids the per-source fallacy of division.
• Source transparency — Name the report behind the under-5% figure, with its sample, time period, and definition of 'open-source', and align terminology with 'open-weight'. Without provenance and consistent terms, the main empirical anchor cannot be evaluated.
• Representation of the opponent — Quote Baker's actual claim and address the GPU-rental-price reply on its merits, including what workload decomposition would show. This protects against strawman charges and makes the 'not shown' claim fairer.
• Conclusion wording — State the conclusion as 'the claim that open-weight demand justifies current spending is currently unsupported' and acknowledge forward-looking and partial-contribution cases. This matches the conclusion to what the premises can establish and removes the drift from 'unshown' to 'incapable'.

Scenario Tests

• Inference-provider and routing-platform data show open-weight models account for a large share of tokens and rented GPU-hours despite small model-licensing revenue. (Challenges) — The under-5% figure would be shown to measure the wrong quantity, and the claim that demand has not been shown would be undermined, though a revenue gap could remain.
• Disaggregated data show open-weight workloads and hosting revenue are small in both dollars and compute relative to capex. (Supports) — Open-weight demand could not carry a meaningful part of the justification, and the burden would move to other demand sources such as closed-model enterprise revenue.
• Baker clarifies that open-weight models are one of several demand drivers, not the sole justification. (Challenges) — The 'comparable scale' standard would be aimed at a position Baker does not hold, and the argument would shrink to a point about partial contribution.
• The under-5% figure comes from a survey skewed toward large enterprises using closed APIs, and other surveys give much higher open-model shares. (Challenges) — The only quantitative anchor would weaken, leaving only the burden-of-proof framing.
• Late-1990s fiber buildout, in which critics demanded segment-level revenue at scale and were largely right on aggregate overbuild, though some demand later materialized. (Neutral) — Historical analogies cut both ways, so burden-of-proof demands are reasonable but cannot settle whether current demand will grow into the spending.

Coherence & Relevance

The argument is coherent as a conditional burden-of-proof challenge: it asks for scale-appropriate evidence and notes that the evidence offered is thin. It is weakest where the evidence and the standard misalign. A spend-share statistic is used to assess infrastructure-relevant demand, and a per-source parity test is applied to what is probably a cumulative justification. Within its stated assumptions the modest reading ('not shown') is reasonably supported, while the stronger reading ('cannot carry') goes beyond what the premises establish.

• Baker's case that open-weight models support infrastructure requires substantial paying demand for those models. (Strong) — It defines the test the rest of the argument applies, but assumes that paying demand for the models is the only route to supporting infrastructure, and rests on an unverified paraphrase of Baker.
• Open-source models were reported to be under 5% of business AI spend. (Moderate) — It is directionally relevant but measures model-vendor spend rather than compute demand, covers only business users, and is not directly commensurable with an absolute infrastructure figure without a bridging premise that business spend proxies total paying demand.
• Infrastructure spending in 2026 runs at hundreds of billions of dollars a year. (Moderate) — It sets the scale of the comparison but treats infrastructure spending as one pool, though it is heterogeneous (training, closed-model inference, enterprise, consumer).
• A claim that a demand source justifies that spending should be supported by evidence of demand at comparable scale. (Strong) — It is the bridge to the conclusion but is asserted rather than argued, and 'comparable scale' is ambiguous between sole sufficiency and meaningful contribution. Read strictly, it proves too much.

View this argument on LogicFirst.ai