Research opportunity · Concentration, Stack & Sovereignty
How much of the fall in AI prices is competition rather than cheaper technology?
Question
For language-model inference since 2023, how much of the fall in the price of a fixed capability level reflects lower serving costs (hardware and algorithmic efficiency) and how much reflects narrowing margins under competition, estimated by comparing prices for identical open-weight models across competing hosts with their estimated serving costs and with proprietary models of similar capability?
Causal mechanism not isolatedOne PhD chapter or paperLow costPartly addressed
Why it matters
Steep inference price declines are often read as proof that the model layer is competitive. If most of the decline is cost-driven, falling prices say little about market power; if margins are narrowing, competition is doing real work. The difference matters for how far a "contestable model layer" can be claimed while the hardware layer stays concentrated.
The debate lists this among what we still do not know: the causal decomposition of inference-price decline
Evidence that motivates it
- Epoch AI's September 2026 multi-benchmark analysis estimates that the price of reaching fixed LLM capability levels has fallen about 13× per year since 2023.
Epoch AI estimates the price of fixed capability has fallen about 13× per year since 2023; the claim explicitly does not license competition as the sole cause, which is the question this study decomposes. - OpenRouter and a16z report that open-weight models reached approximately one-third of token usage on OpenRouter by late 2025; an NBER working paper using OpenRouter and Microsoft Azure usage data reports a lower open-source share of OpenRouter tokens, about 25% by December 2025 after a peak near 45% in August.
Open-weight models carry a substantial share of OpenRouter tokens, so the platform offers many identical-weights products sold by competing hosts, which is the natural setting for separating cost from margin; the claim itself says nothing about prices.
A possible design
Assemble a monthly price panel from archived OpenRouter listings and provider price pages (2023–2027), with per-host prices for the same open-weight model and list prices for proprietary models, matched to capability scores (Artificial Analysis or Epoch benchmarks). Estimate per-token serving cost for open-weight models from public GPU rental prices and published or benchmarked throughput on given hardware, giving an approximate price–cost margin per host. Then: (1) decompose the fixed-capability price decline into cost-index change and margin change; (2) use the entry of new competitive open-weight releases as events to test whether proprietary prices at similar capability fall faster than their estimated costs; (3) compare margins by number of hosts serving a model. Report the share of the decline attributed to margins with bounds that reflect cost-estimate uncertainty.
Data needed
Archived listings (Internet Archive snapshots of OpenRouter and provider pricing pages); GPU rental price series; inference throughput benchmarks; capability benchmarks; one modest GPU budget to verify throughput for a few models.
Feasibility
One PhD chapter or paper. Low: mostly public data; a few hundred to a few thousand US dollars of GPU time to check throughput estimates. Proprietary serving costs are unobserved, so the proprietary side can only be bounded, not measured; list prices may differ from negotiated enterprise prices. An economist co-supervisor would help with the industrial-organisation framing.
Has this been done?
Partly done already: related work covers some of this question (first-pass screen of 12 October 2026, checked again on 13 October 2026).
Gundlach et al. separate algorithmic from hardware efficiency but do not quantify a margin component. Du decomposes the cost fall into productivity and hardware components and reads a 2024 break as competition-driven, without estimating margins. Zhu (2026) builds a quality-adjusted price index across 86 providers and attributes the fall mainly to within-provider technical progress, while noting the index does not recover cost. Demirer et al., Li et al. (2026) and He et al. (2026) document how prices for identical open-weight models differ across hosts, without cost estimates; a MIRI blog analysis uses the cheapest host as a rough cost proxy as a one-off snapshot. Erdil (2025) offers a serving-cost model that could supply the cost side. What remains open is an explicit cost-versus-margin decomposition of the trend using host-level prices for identical weights against estimated serving costs, and a test of whether competitive open-weight entry compresses proprietary margins.
How strong this evidence is
- Who searched and checked: Claude (SocioLens Project Lead) with AI research agents, by web search; not a person and not a domain expert; an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Every check so far was done by AI; no person or domain expert has checked it yet.
- What was read: 3 of the 8 nearest works below were judged from an abstract, summary or search listing only; each is marked. Not every entry says how much of the work was read.
- Where the search could not reach: Screened by web search on the date shown; Google Scholar, SSRN and paywalled economics journals were only partly reachable, and the Epoch AI analysis cited in the debate was not re-opened for this screen. Absence of a hit is not proof that the work does not exist. Literature check of 2026-10-13: Most new papers read at abstract level; Zhu read in full for the first 100,000 characters only. A Journal of Economic Perspectives 2026 article (doi 10.1257/jep.20261506) surfaced in results but could not be fetched, so its title, authors and content are unconfirmed and it may discuss LLM pricing and markups. Chinese-language search returned only press commentary on the 2024 to 2025 price war, no empirical decomposition.
- Before building a project on it, run your own search in a scholarly database such as Scopus, Web of Science, Google Scholar or SSRN.
Nearest work found
- Gundlach, Lynch, Mertens & Thompson, "The Price of Progress: Price Performance and the Future of AI," arXiv:2511.23455 (Nov 2025, v2 Mar 2026) Judged from an abstract, summary or search listing only
Estimates the price of fixed benchmark performance falling about 5–10× per year for frontier models and algorithmic efficiency of about 3× per year, naming economic forces, hardware and algorithms as drivers; competition is controlled for, not measured as a separate component. Abstract read only. - Du, "Tiered Super-Moore's Law: Price Evolution, Production Frontiers, and Market Competition in Large Language Model Inference Services," arXiv:2603.28576 (Mar 2026) Judged from an abstract, summary or search listing only
Reports a roughly 600-fold token-price decline from 2020 to 2026, a structural break in May 2024 that the author reads as a shift from technology-driven to competition-driven declines, and falling concentration (HHI); attributes most cost reduction to productivity rather than hardware. Does not estimate price–cost margins, so the competition reading rests on timing rather than a measured margin. Abstract read only. - Demirer, Fradkin, Tadelis & Peng, "The Emerging Market for Intelligence: Pricing, Supply, and Demand for LLMs," NBER Working Paper 34608 (Dec 2025)
Shows wide price dispersion across hosts of identical open-weight models and near-uniform pricing for closed models, and estimates open models are roughly 90% cheaper at equal capability; no serving-cost estimates or decomposition of the price trend. - Scher, "Observations About LLM Inference Pricing," MIRI Technical Governance Team blog (3 Mar 2025)
Cross-sectional snapshot of Artificial Analysis prices (December 2024 to January 2025) that uses the cheapest host's price for an open-weight model as a rough marginal-cost proxy, finds prices for the same model often vary about tenfold across hosts, and suggests many proprietary models carry substantial markups; not peer reviewed, not a time-series decomposition, and the author calls it a shallow dive. - Zhu, "The Price of Intelligence: A Quality-Adjusted Price Index for AI Services," arXiv:2608.29843 (Aug 2026)
Closest new work. Quality-adjusted (hedonic) price index from posted prices of 86 providers; finds the decline is overwhelmingly within-provider technical progress, not seller composition. Explicitly notes the hedonic index recovers the equilibrium price schedule, not cost, and that subsidised or strategic pricing moves it further from cost. Does not estimate serving costs or margins, does not compare the same open-weight model across hosts, and does not study open-weight entry effects. Narrows but does not close the gap. - He, Wen, Chen, Yang, Lan, Cui & Wang, "You Cannot Pick a Provider From the Price List: Market-Aware Routing for Open-Weight LLM Inference," arXiv:2609.37902 (Sept 2026)
Measures live endpoints and finds the same open-weight model varies sharply in price, quality, latency and availability across providers; higher-priced hosts are faster. Documents the host-level price dispersion the design would use, but does not analyse costs or margins (per abstract). - Li, He, Liu, Xu, Liu, Ma, Wu, Wu, Tang & Shi, "When Is the Same Model Not the Same Service? A Measurement Study of Hosted Open-Weight LLM APIs," arXiv:2605.02821 (May 2026) Judged from an abstract, summary or search listing only
Measurement study of hosted open-weight APIs using pricing snapshots; finds listed prices more anchored across hosts than latency, throughput and other attributes. A systems measurement, not a cost-margin decomposition. Note its finding of anchored prices sits in some tension with He et al.'s sharp dispersion; both are abstract-level readings. - Erdil, "Inference economics of language models," arXiv:2506.04645 (Jun 2025)
Theoretical model of cost per token versus serial speed from arithmetic, memory and network constraints, giving Pareto frontiers for popular models. Provides a method for the serving-cost side of the design; does not (per abstract) compare costs with market prices or estimate margins.
Literature check
Stands as first published. Newer work measures the price fall and cross-host price dispersion but still does not split it into cost and margin: Zhu (arXiv:2608.29843, Aug 2026) builds a quality-adjusted price index from 86 providers and attributes the decline mainly to within-provider technical progress, explicitly noting a hedonic index does not recover cost; He et al. (arXiv:2609.37902) and Li et al. (arXiv:2605.02821) document price and service variation for the same open-weight model across hosts without cost or margin estimates. Erdil (arXiv:2506.04645) models serving cost from hardware constraints, a usable input. No work decomposes the trend into cost and margin or tests open-weight entry effects on proprietary margins.
Searched: web (extended); arxiv.org; ssrn.com; osf.io; nber.org; dl.acm.org; sciencedirect.com; academic.oup.com; journals.sagepub.com; tandfonline.com; aeaweb.org; onlinelibrary.wiley.com; journals.uchicago.edu; Chinese-language web. Checked by Claude (SocioLens Project Lead) with AI research agents, by web search; not a person and not a domain expert.
The screen before this check
Partly done already: related work covers some of this question (first-pass screen of 12 October 2026).
Gundlach et al. separate algorithmic from hardware efficiency and control for competition by isolating open models, but do not quantify a competition or margin component. Du decomposes the cost fall into productivity and hardware components, tracks falling concentration and reads a 2024 structural break as a shift to competition-driven declines, but does not estimate margins. A short MIRI blog analysis uses the cheapest host's price as a rough cost proxy to suggest proprietary markups, as a one-off snapshot rather than a decomposition of the trend. Demirer et al. show identical open-weight models are priced very differently across hosts while closed models are priced uniformly, without a cost-side decomposition of the trend. What remains open is an explicit cost-versus-margin decomposition using host-level prices for identical weights against estimated serving costs, and a test of whether competitive open-weight entry compresses proprietary margins.
Search queries used
decomposing LLM inference price decline hardware algorithmic efficiency competition markupsEpoch AI LLM inference prices trends price to reach capability falling per yearLLM API pricing market power markups estimate demand estimation structural model language modelsLLM inference markups price-cost margin estimate open-weight providers serving cost GPU competition paper 2026
Literature check, 13 October 2026:
LLM API price markup estimate serving cost GPU throughput margin analysis (arxiv.org, ssrn.com, nber.org, osf.io)hedonic price index large language models quality-adjusted inference prices competition entry DeepSeek price cuts empirical"Tiered Super-Moore's Law" OR "The Price of Progress: Price Performance and the Future of AI" cites extends margins competitionopen-weight model entry effect on proprietary LLM prices event study大模型 价格战 推理成本 毛利率 实证 分解 竞争industrial organization generative AI market pricing markups competition (aeaweb.org, sciencedirect.com, academic.oup.com, journals.uchicago.edu, onlinelibrary.wiley.com, pubsonline.informs.org)Journal of Economic Perspectives 2026 generative AI market structure pricing jep.20261506
Provenance
RO-LD11-05 v2. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-11 debate on 2026-10-12. Checked by an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.
Changes
- v2, 13 October 2026: Literature check added: a second, deeper search for existing work, with its queries, new nearest work and limits. The question and verdict stand. See v1 as published.