Scholars Challenge Viral Chart Claiming 100,000x Drop in AI Prices

Starting September 23, Oxford philosopher Toby Ord publicly challenged a viral chart claiming AI prices fell more than 100,000x between 2021 and 2026, with economist David Roodman and researcher David Manheim adding methodological and structural critiques in discussions related to Epoch AI. Their shared conclusion: the chart conflates capability gains with cost declines into a single cost curve, so the 100,000x claim does not hold; a more accurate description is that prices within each capability tier fell roughly 100 to 1000x cumulatively.

Confirmed

• Ord argued that it is hard to name anything that actually got 100,000x cheaper; the figure resembles an aggregate or proxy statistic unsuitable for presenting price trends directly.
• In discussion with @ramez, Ord noted that per capability tier, price declines are not exponential but flatten quickly even measured in monthly percentage terms, so the headline 100,000x figure does not hold.
• Ord said the chart mixed capability improvements into the price curve, confusing the statistical basis; economist David Roodman made a similar critique, arguing that combining rapid capability gains with cost declines to derive a pure cost conclusion is methodologically infeasible.
• In his corrected account, Ord offered two points: prices within each capability tier fall fast then slow, cumulatively declining about 100–1000x, while model quality available at a given price keeps rising rapidly.
• In follow-up discussion with Roodman and Epoch AI, Ord raised two reservations about claims of steeply falling AI costs: beyond how the decline is presented, he is unsure of its exact scope—how many parts of AI services have not seen similar declines—questioning whether Epoch's cited 725x service price decline represents the whole picture.
• David Manheim acknowledged that distilled small models do deliver slightly lower quality at much lower prices, but the sweeping conclusion of a dramatically falling overall cost curve does not follow; he added that actual costs have not fallen proportionally to "GPU price declines + capability gains" because of fixed costs.

Unconfirmed

• The original source and exact data definitions of the chart were not given in the posts.
• The scope of AI service components that have not seen similar price declines remains uncertain, including for Ord himself.

Why it matters

• Manheim highlighted a structural constraint on compute prices: users cannot buy just "10 flops"—the same price buys a quadrillion times more compute—so the reverse scenario (everything getting a quadrillion times cheaper) is impossible due to fixed-cost bottlenecks, implying a de facto price floor for compute.
• Such cost-decline curves are often used in AI trend forecasting and policy arguments; if the methodology is flawed, they could systematically overstate the pace of AI price declines and distort judgments about AI adoption and economic impact.

2026-09-23 ~ 2026-09-24 · 10 related posts

Full story(2 episodes)→

• Episode 1: Epoch AI: Cost of Equivalent AI Performance Falls 47% Per Quarter(2026-09-23, 13 posts)
• Episode 2: Scholars Challenge Viral Chart Claiming 100,000x Drop in AI Prices(2026-09-23, 10 posts)

• [source] Toby Ord questions chart claiming AI prices fell 100,000x since 2021 — tobyordoxford · 2026-09-23
• [source] LLM cost curves questioned: distillation gains vs the $0.01/mtok price floor — davidmanheim · 2026-09-23
• Why compute prices can't collapse: you can't buy 10 flops, only trillion-fold bundles — davidmanheim · 2026-09-23
• Economist challenges Epoch's cost-decline curve: conflating capability gains with cost reduction — davidroodman · 2026-09-23
• Toby Ord: Each AI Capability Tier Falls ~100-1000x in Cost While Quality Keeps Climbing — tobyordoxford · 2026-09-24
• [source] Toby Ord questions Epoch's 725x cost-decline figure: what isn't getting cheaper? — tobyordoxford · 2026-09-24
• Epoch AI debate: fixed costs complicate the '100,000x compute price drop' narrative — davidmanheim · 2026-09-24

3 near-duplicate retellings: tobyordoxford · tobyordoxford · tobyordoxford