The compute cost of the most expensive AI training run will reach about $230 billion (one percent of 2021 US GDP) around 2040.
3 events · 1 assessment · 1 decision
Structured and assessed first pass
First pass (structure_and_assess). Decomposed into two named arguments. FOR (Trend extrapolation): minted two novel subclaims after match_claim confirmed novelty — the 2025 base cost (~several hundred million dollars; seeded 0.8, importance 0.3) and the continuation crux (rapid cost growth persists through the 2030s; seeded 0.35, importance 0.55, contestation 0.8) — and linked two existing trend claims (0.2 OOM/yr since 2015; >2x/yr through mid-2020s) as supports. AGAINST (Limits to sustained cost growth): linked existing claims on financial unsustainability through 2030 and a 2020s spending slowdown as contradicts. Declined to mint a proposed "power/chip/capital constraints will halt scaling" claim as redundant with the existing constraint claims. Verified the source (Epoch AI, Cottier 2023) and cross-checked the base point (AI Index GPT-4/Gemini Ultra figures) and the feasibility literature (Epoch "Can AI scaling continue through 2030?") via three web searches; no new instances recorded since the only assertions found were the originator (already recorded) and its own cross-posts, and no source asserts the negation of the dated claim specifically. Improved the canonical form to pin the dollar figure and GDP base year ($230B, one percent of 2021 US GDP), since "one percent of US GDP around 2040" misread as 2040-vintage GDP; identity unchanged, same considerations bear. Set importance 0.45 / contestation 0.6 (notable: the 1%-of-GDP threshold figures in AI-timelines reasoning, but the claim is a single-org extrapolation with a three-decade CI and no graph dependents). Assessed CONTESTED (confidence 0.6, credence 0.25, marginal_yield 0.3): premises well anchored, crux genuinely disputed, and the date can miss in both directions (observed growth faster than the extrapolated rate implies pre-2040 arrival; a financial break implies much later or never). No dependents exist, so no notification sent.
Assessed Contested
verdict confidence 0.60 · credence 0.25
The claim originates in a 2023 Epoch AI analysis by Ben Cottier, which extrapolated the dollar cost of the largest machine-learning training runs and forecast that a single run would cost about $233 billion, one percent of 2021 US GDP, around 2040, with a wide stated uncertainty spanning 2033 to 2062. The extrapolation's ingredients hold up individually: the most expensive run in 2025 cost on the order of several hundred million dollars, and costs for the largest runs have grown steadily since 2015, a pace that remained multi-fold per year through the mid-2020s. What the forecast turns on is whether that growth continues through the 2030s, and this is genuinely disputed. Sustaining the trend to 2040 would require developers to spend tens and then hundreds of billions of dollars on individual runs, which critics argue is financially unsustainable given the gap between AI infrastructure spending and AI revenues; a related prediction that spending growth will slow substantially during the 2020s is itself contested. On the other side, the trend has repeatedly outlasted predictions of its end, and feasibility analyses find no hard technical barrier through at least 2030. The date could also miss in the other direction: costs through the mid-2020s have grown faster than the rate the forecast extrapolates, and at that faster pace the milestone would arrive in the early-to-mid 2030s rather than around 2040. The question resolves progressively as the decade unfolds; a clear break or persistence in frontier training budgets through the late 2020s would substantially narrow the range.
Claim entered the graph