Growth in spending on the largest AI training runs will slow substantially during the 2020s.
Assessment
Credible evidence or argument exists on multiple sides.
The prediction originates in early-2020s forecasting work, notably Epoch AI's 2023 cost-trend analysis, which judged the recent pace of growth in frontier training spending unsustainable and expected it to slow greatly before 2030. Through the middle of the decade the prediction has not come true: estimates place the cost of frontier training runs on a 2 to 3x per year growth path, with the multi-fold annual growth rate persisting through the mid-2020s, frontier run budgets crossing roughly half a billion dollars in 2025, and hyperscaler infrastructure spending accelerating rather than retrenching into 2026.
Whether a substantial slowdown still arrives in the decade's remaining years is genuinely disputed. The case that it must rests on finance and physics: continued multi-fold growth in training spending is argued to be financially unsustainable through 2030, with a widely cited gap of hundreds of billions of dollars between AI infrastructure spending and AI revenues, and multi-gigawatt power requirements looming for the largest runs. The case against notes that the growth trend has held for over a decade despite similar past predictions, and that constraint analyses find training runs of around 2e29 FLOP feasible by 2030, meaning any slowdown before then would come from choices about spending rather than hard limits.
The question resolves by the end of the decade: a sharp deceleration in frontier training budgets during 2027 to 2029 would vindicate the prediction, while continued multi-fold annual growth to 2030 would refute it. On the evidence so far, the burden has shifted toward those expecting the slowdown.
Full reasoning: the evidence and decisions behind this verdict
The claim entered the graph from Epoch AI's analysis of dollar training costs (epoch.ai/blog/trends-in-the-dollar-training-cost-of-machine-learning-systems), which affirmed that spending growth "will likely slow down greatly during the 2020s" given the apparent unsustainability of the trend, aligning with earlier CSET and Cotra forecasts.
Evidence against, so far dominant on the observed record: Epoch AI's 2024 cost study (epoch.ai/blog/how-much-does-it-cost-to-train-frontier-ai-models) finds amortized hardware and energy costs of frontier training runs growing at 2.4x per year since 2016 (95% CI 2.0x to 3.1x), with no observed break in the trend, projecting billion-dollar runs by 2027; this grounds the subclaim that costs kept growing at more than 2x per year through the mid-2020s. Epoch's constraint analysis (epoch.ai/blog/can-ai-scaling-continue-through-2030) concludes runs of about 2e29 FLOP are likely feasible by 2030, power binding first, so no external constraint forces a slowdown within the decade. Reporting on hyperscaler budgets shows combined AI capital expenditure guidance for 2026 near $700 billion, up roughly 75 percent year over year, an acceleration rather than the predicted deceleration.
Evidence for a coming slowdown: the financing arithmetic is widely flagged as strained, with a commonly cited gap of around $600 billion between AI infrastructure spending and AI-attributable revenues (Sequoia's David Cahn, echoed in 2026 coverage, e.g. www.forbes.com/sites/jasonkirsch/2026/06/02/the-ai-capex-to-revenue-gap-is-widening---and-markets-are-starting-to-notice/); Goldman Sachs reports analyst expectations of capex growth slowing sharply through 2026 (www.goldmansachs.com/insights/articles/why-ai-companies-may-invest-more-than-500-billion-in-2026); and industry analyses assert a power- and capital-driven deceleration in frontier scaling (deluair.com/consultancy/insights/frontier-ai-training-cost-2026). This is the live content of the financial unsustainability subclaim, which remains unresolved: capex growth-rate moderation from 75 percent toward 25 percent per year would still be growth, and slower overall capex growth does not directly entail substantially slower growth in the largest individual training runs.
Weighing: the claim's instances point both ways (the originating Epoch blog and the Deluair analysis affirm; the descriptive record leans against). More than half the decade has elapsed without the predicted substantial slowdown, which is why the credence sits below even odds, near 0.35: the claim now requires a sharp break during 2027 to 2029. But the sustainability argument is credible and actively pressed by serious analysts, so the disagreement is real rather than fringe, which makes contested the right status rather than contradicted. What would change the verdict: evidence that frontier training run budgets flattened or grew well below 2x per year across 2026 to 2028 would move the claim toward supported; continued 2 to 3x annual growth into 2028 to 2029 would move it toward contradicted. The verdict is also mildly sensitive to reading "slow substantially": measured against the fastest early-2020s rates a modest moderation (for example 3.2x to 2.4x per year) has arguably occurred, but the sources asserting the claim plainly meant a much larger deceleration, and that has not happened.
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
Because the dollar cost of final training runs grew about half an order of magnitude per year through 2022, continuing at that rate would put individual runs in the tens of billions of dollars before 2030; and given that growth at recent multi-fold annual rates is financially unsustainable through 2030, the growth rate must fall substantially before the decade ends.
The extrapolation is sound: continuing the documented trend does reach sums in the tens of billions per run before 2030, so the inference goes through if its forward premise holds. The argument lives or dies on whether growth at recent rates is financially unsustainable through 2030, which remains an open dispute between revenue-gap arithmetic and the observed willingness of hyperscalers to keep accelerating investment. A caveat: even granting unsustainability at some point, the slowdown must arrive before 2030 for the conclusion as stated, and the timing is exactly what the premise does not fix.
Because frontier training run costs continued to grow at more than 2x per year through the mid-2020s, more than half the decade has passed without the predicted deceleration; and given that training runs of around 2e29 FLOP will be feasible by 2030, no physical or industrial constraint forces a slowdown before the decade ends.
The observed record carries most of the weight: continued multi-fold annual cost growth through the mid-2020s is well evidenced and directly undercuts a prediction of substantial slowing during the decade, since most of the decade has now elapsed without it. The feasibility premise, that runs of around 2e29 FLOP will be feasible by 2030, adds that no hard constraint forces a late-decade break. The caveat is that neither premise rules the claim out: a sharp, voluntary or finance-driven deceleration in 2027 to 2029 would still satisfy it, so the argument shifts the burden rather than closing the question.
Provenance
Where this claim has been said, linked to its canonical form.
recent growth spending will likely slow down greatly during the 2020s, given the current willingness of leading AI developers to spend on training, and given that the recent overall growth rate seems unsustainable
This is consistent with the direction of predictions in AI and Compute (from CSET), and Cotra’s “Forecasting TAI with biological anchors”—namely, that recent growth spending will likely slow down greatly during the 2020s.
Frontier model training compute scaled at roughly 4x to 5x per year between 2018 and 2024, sat near 10x per year for the leading lab releases, and now confronts a deceleration driven by power, capital, and data, not by silicon.
An industry analysis of frontier AI training cost trajectories arguing that pretraining budgets, having crossed roughly half a billion dollars in 2025, face a power- and capital-driven deceleration in scaling before the next order of magnitude.
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by extractor · Aug 11, 2026. Every judgment on this page is accompanied by a reasoning trace.