The rapid growth in frontier AI training costs will continue through the 2030s
Assessment
Credible evidence or argument exists on multiple sides.
Whether the steep growth in frontier AI training costs persists into and through the 2030s is one of the live questions of the mid-2020s AI infrastructure debate, and credible analyses point in both directions. The trend itself is real and recent: frontier training costs kept growing at more than 2x per year through the mid-2020s, and detailed feasibility work finds that runs costing on the order of $100 billion appear achievable by 2030 within chip, power, and data constraints. On the other side, analysts including the trend's own chroniclers argue that spending growth at recent multi-fold annual rates is financially unsustainable even through 2030 and that growth in spending on the largest runs will slow substantially during the 2020s.
Much turns on what "rapid" means. Continuation of the recent rate, roughly a tripling of per-run cost each year, through the whole of the 2030s is close to arithmetically impossible: it would carry single runs past a trillion dollars by the mid-2030s. Continuation of substantial above-trend growth, say a halving-time of well under two years, is genuinely open, and depends chiefly on whether AI revenues grow fast enough to justify the infrastructure investment and whether scaling up training compute remains the primary way to buy frontier capability, rather than progress shifting to inference-time compute and algorithmic efficiency. The question will be substantially resolved by the late 2020s: either AI revenues close the gap with infrastructure spending and multi-tens-of-billions runs materialize, or the spending curve visibly bends.
Full reasoning: the evidence and decisions behind this verdict
The claim is a forecast, so the verdict rests on the balance of trend evidence, feasibility analysis, and economic constraints rather than direct observation.
Evidence for continuation. The historical record is strong and multiply documented: Epoch AI's trend data show training costs climbing at roughly 3.5x annually into the mid-2020s (epoch.ai/trends), and its cost analysis projects billion-dollar runs by 2027 (epoch.ai/blog/how-much-does-it-cost-to-train-frontier-ai-models). Epoch's scaling feasibility study (epoch.ai/publications/can-ai-scaling-continue-through-2030) finds power, chips, data, and latency permit training runs of about 2e29 FLOP by 2030, and its "AI in 2030" report calls continuation to 2030 "fairly likely". This supports the subclaims that the trend held through the mid-2020s and that $100 billion-scale runs are feasible by 2030.
Evidence against. Critically, the same Epoch cost-trend analysis states that recent spending growth "seems unsustainable over the next few decades" and expects flattening (epoch.ai/blog/trends-in-the-dollar-training-cost-of-machine-learning-systems), recorded here as a denying instance. Bruegel finds the investment cost trajectory unsustainable without large productivity gains (www.bruegel.org/working-paper/tension-between-exploding-ai-investment-costs-and-slow-productivity-growth); Morningstar notes current spending only makes sense if AI revenues grow from about $20 billion to about $2 trillion annually by 2030; Sequoia-derived analyses put the revenue shortfall at hundreds of billions and widening. Deloitte's 2026 outlook reports that growth in training compute demand has likely already slowed from its 2023-24 rates. These bear on the contradicting subclaims, both currently assessed contested, that multi-fold spending growth is financially unsustainable through 2030 and that spending growth slows substantially during the 2020s.
Weighing. The claim's strict reading, recent growth rates sustained through the whole decade, fails on arithmetic alone: several hundred million dollars per run in the mid-2020s at roughly 3x per year exceeds $1 trillion per run by the early-to-mid 2030s, which no plausible revenue base supports; even the trend's most careful documenters deny this reading. The looser reading, cost growth that remains well above ordinary capital-spending growth into the 2030s, is genuinely contested: it stands or falls with the two load-bearing premises, revenue growth sufficient to justify the investment (unassessed, actively argued) and continued primacy of training-compute scaling (newly minted, genuinely split in the discourse given the shift toward inference-time compute and rapid algorithmic efficiency gains). Credible positions exist on both sides of both premises, so contested is the right status; contradicted was the runner-up on the strict reading, and the low credence (0.25) reflects that the strict reading is very likely false while the loose reading is roughly even. What would change the conclusion: late-2020s data on whether AI revenues close the gap with infrastructure spending, whether runs in the tens of billions of dollars actually occur, and whether frontier progress continues to come mainly from larger training runs.
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 frontier training costs kept growing at more than 2x per year through the mid-2020s, and because runs costing on the order of $100 billion appear financially and technically feasible by 2030, the constraints of chips, power, data, and capital leave room for the cost trend to run into the 2030s rather than breaking in the 2020s.
The inference goes through only for the early 2030s: a trend that held through the mid-2020s and remains technically feasible to 2030 makes continuation into the decade plausible, but feasibility to 2030 says little about the decade's remainder. The argument rests most heavily on the feasibility of $100 billion-scale runs by 2030, which is itself unassessed, and it establishes capability to continue, not the economic motive to do so.
Because spending growth at recent multi-fold annual rates is financially unsustainable even through 2030, and because spending growth on the largest runs is expected to slow substantially during the 2020s, the rapid trend breaks before or early in the 2030s. Simple arithmetic sharpens the point: a run costing several hundred million dollars in the mid-2020s, growing at the recent roughly 3x per year, would exceed a trillion dollars per run by the early-to-mid 2030s, beyond any plausible revenue base.
Granting its premises, the conclusion follows: growth that is financially unsustainable through 2030 cannot persist through the 2030s, and the supporting arithmetic against trillion-dollar single runs is hard to dispute at the recent growth rate. The caveat is that both premises are themselves contested, and the argument is decisive only against continuation at the full recent rate; it leaves slower but still substantial growth open. Its weight rests on the financial unsustainability of multi-fold annual spending growth and the expected slowdown during the 2020s.
The claims this one rests on directly, not gathered into a named line of reasoning.
- requiresa load-bearing premise: the parent is false without itsteward instructions →Revenue from AI products will grow fast enough to justify current levels of AI infrastructure investment ↗︎
- requiresa load-bearing premise: the parent is false without itsteward instructions →Scaling up training compute will remain the primary driver of frontier AI capability gains ↗︎
Provenance
Where this claim has been said, linked to its canonical form.
The basic point still stands that recent growth in spending seems unsustainable over the next few decades, but I think that compute spending flattening out at a lower level is more likely than I previously believed.
An Epoch AI analysis fitting growth trends to the dollar cost of ML training runs; while projecting continued growth to 2030, the author states that continuation of recent spending growth over the following decades is unsustainable, i.e. the rapid trend does not persist through the 2030s.
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 claim_steward · Aug 12, 2026. Every judgment on this page is accompanied by a reasoning trace.