Research opportunity · Concentration, Stack & Sovereignty
How large is open-weight model use beyond a single routing platform?
Question
How does the open-weight share of language-model use estimated from OpenRouter compare with shares estimated from other observable sources (other inference marketplaces and hosts, model downloads, self-hosting tool registries, and a survey of organisations' production deployments), and which population does each source actually represent?
Untested beyond the original settingOne PhD chapter or paperNeeds a budget (medium or higher)Partly addressed
Why it matters
The claim that open-weight models reduce dependence on proprietary vendors rests heavily on usage data from one routing platform whose users (developers, hobbyists, price-sensitive and roleplay traffic) may differ sharply from enterprises and public bodies. Without a second, independent measurement, nobody can say whether "about one-third" is a market fact or a platform fact.
The debate lists this among what we still do not know: global open-weight share beyond OpenRouter
Evidence that motivates it
- 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.
The one-third open-weight token share is measured on OpenRouter only, and the claim does not license a global, enterprise or infrastructure-independence reading; this study tests how far it travels. - Menlo Ventures estimated late-2025 enterprise LLM API shares of Anthropic 40%, OpenAI 27% and Google 21%, compared with OpenAI 50% and Anthropic 12% in 2023.
The only enterprise-side provider shares cited come from a single venture-investor survey and are rated contested; an independent organisational survey would add a second, replicable enterprise measure. - Concentration claims in AI must identify the stack layer, market definition, metric, geography and period; concentration at one layer does not establish concentration or power at another.
Open-weight "share" means different things under different denominators (tokens, downloads, deployments, spend); the design keeps them as separate estimands.
A possible design
Build a multi-source panel for 2025–2027 with one row per model family per month: (1) token shares on OpenRouter from its public rankings; (2) usage or listing signals from other inference marketplaces and hosts that publish them; (3) Hugging Face download histories (using the released Longpre et al. dataset as a base); (4) pull counts or listings from self-hosting registries where public; and (5) a fielded survey of 500–1,000 organisations (stratified by size and sector, in two or three countries) asking which models run in production, how they are hosted, and roughly what share of their language-model workload each takes. Treat each as a separate estimand, report open-weight shares with intervals per source, and test how well platform-based shares predict survey-based shares. A short validation component asks a few hosting firms or a university IT department for aggregate logs to check the survey against records.
Data needed
Public OpenRouter rankings (archived snapshots exist); Hugging Face metadata and the Longpre et al. dataset; registry data where public; a commissioned organisational survey (or a module on an existing business survey); optional data-sharing agreements with hosts.
Feasibility
One PhD chapter or paper. Medium: public data is free; an organisational survey costs roughly US$5,000–20,000 depending on country and panel provider, less if added to an existing business survey. Hosts may not share logs; the study is still informative from public sources plus the survey. Downloads are not usage, and the write-up must not convert them into usage shares.
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).
Demirer, Fradkin, Tadelis and Peng pair OpenRouter with Microsoft Azure AI Foundry data for pricing and demand rather than a population share. Nagle and Yue (2025) and the ATOM Report (2026) both rely on OpenRouter for usage share, and ATOM covers open models only. Longpre et al. map Hugging Face downloads (distribution, not use), and Dunivin (2026) measures open-weight use in scientific papers, a separate population. What remains open is a deliberately multi-source estimate that treats each source as a different population, includes self-hosted and enterprise production use measured by a replicable organisational survey, and tests whether the OpenRouter share generalises.
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: 1 of the 7 nearest works below was judged from an abstract, summary or search listing only, and 1 has details that could not be confirmed; it 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, paywalled journals and commercial survey reports (for example industry analyst subscriptions) were not reachable. The Demirer et al. details were checked against the NBER working-paper PDF. Absence of a hit is not proof that the work does not exist. Literature check of 2026-10-13: SSRN page for Nagle and Yue could not be fetched; its description relies on a Linux Foundation blog summary. Commercial enterprise surveys (vendor and analyst reports) surfaced in search but were not treated as scholarly evidence and were not opened. Full texts of ATOM and the Ollama measurement paper were read only partly or via abstract.
- 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
- Aubakirova, Atallah, Clark, Summerville & Midha, "State of AI: An Empirical 100 Trillion Token Study with OpenRouter," arXiv:2601.10088 (Jan 2026)
The source of the open-weight usage pattern cited in the debate; uses OpenRouter data only, so it cannot say how representative that platform is. - Demirer, Fradkin, Tadelis & Peng, "The Emerging Market for Intelligence: Pricing, Supply, and Demand for LLMs," NBER Working Paper 34608 (Dec 2025)
Uses scraped OpenRouter data plus firm-day usage from Azure AI Foundry (a restricted set of models used by selected Azure customers, July 2023 to June 2025); reports open-source models about 90% cheaper than closed models of comparable capability, and an open-source OpenRouter token share that rose from about 25% in January 2025 to about 45% in August before falling back to about 25% by December 2025 (the introduction elsewhere says it stayed below 30%). The authors caution that OpenRouter is not representative of the global market. Two platforms, but no survey of production use or self-hosting. - Longpre, Akiki, Lund, Kulkarni, Chen, Solaiman, Ghosh, Jernite & Kaffee, "Economies of Open Intelligence: Tracing Power & Participation in the Model Ecosystem," arXiv:2512.03073 (Nov 2025)
Weekly download histories for 851,000 Hugging Face models (2.2 billion downloads, 2020–2025) showing shifting developer concentration; downloads only, no inference or production usage, and no comparison with proprietary model use. - Nagle & Yue, "The Latent Role of Open Models in the AI Economy," SSRN working paper 5767103 (2025; year from SSRN listing, page not opened) Judged from an abstract, summary or search listing onlySome details not confirmed
Per the Linux Foundation summary of the paper (SSRN page could not be fetched), uses OpenRouter usage, price and performance data and reports closed models take about 80% of usage and 96% of revenue despite costing about six times more. It is a second single-platform estimate from OpenRouter, so it strengthens the need to test whether OpenRouter shares generalise; it does not combine sources or survey organisations. - Lambert & Brand, "The ATOM Report: Measuring the Open Language Model Ecosystem," arXiv:2604.07190 (v2 May 2026)
Measures the open-model ecosystem with Hugging Face downloads, derivatives, OpenRouter inference token share and benchmarks. States inference share is drawn from OpenRouter only (top 10 per month) and that available usage data correlate strongly with Hugging Face downloads. Covers open models only, so it gives no open versus closed share, and does not survey production deployments; it is a partial cross-source check among open models. - Dunivin, "Who Uses Open-Weight Models? China and the Shifting Geography of AI in Science," arXiv:2608.11090 (Aug 2026)
Extracts model use from 21 million full-text papers (S2ORC) and finds open-weight use in single-family papers reaching 44.0% in 2026, driven by Chinese models. A distinct source and population (scientific research) that could serve as one panel row; it does not cover commercial or organisational production use and does not compare against OpenRouter. - Xu, Li, Qiu & Sun, "Ollama in the Wild: A Longitudinal Measurement of Exposed Ollama LLM Endpoints at Internet Scale," arXiv:2609.07115 (Sept 2026)
Internet-scale measurement of exposed self-hosted Ollama endpoints over a year (exposure, versions, geography, networks). The abstract does not report which model families are served, so it is a possible method for a self-hosting signal rather than an estimate of open-weight use share.
Literature check
Stands as first published. New 2025 to 2026 work adds further single-source estimates but no multi-source comparison: Nagle and Yue (SSRN 5767103) estimate open versus closed usage and revenue shares from OpenRouter only; the ATOM Report (Lambert and Brand, arXiv:2604.07190) uses OpenRouter for inference share and notes it correlates with Hugging Face downloads but does not estimate an open versus closed share; Dunivin (arXiv:2608.11090) measures open-weight use in 21 million scientific papers. None fields an organisational survey or treats sources as distinct populations, so the opportunity stands as written.
Searched: web (extended); arxiv.org; ssrn.com; osf.io; nber.org; dl.acm.org; sciencedirect.com; academic.oup.com; journals.sagepub.com; tandfonline.com; hbs.edu; aeaweb.org; linuxfoundation.org. 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).
Demirer, Fradkin, Tadelis & Peng already pair OpenRouter data with Microsoft Azure AI Foundry usage, so a second platform has been used, though Azure is a narrower, enterprise-oriented setting and the paper's aim is pricing and demand rather than a population share. Longpre et al. map open-weight downloads on Hugging Face, which measure distribution, not use. What remains open is a deliberately multi-source estimate that treats each source as a different population, includes self-hosted and enterprise production use measured by a replicable survey, and tests whether the OpenRouter share generalises.
Search queries used
open-weight model share of token usage across inference platforms measurement study beyond OpenRouter"State of AI" OpenRouter a16z 100 trillion token study open-weight one thirdHugging Face downloads open model adoption economics paper open vs closed model usage share Nagle
Literature check, 13 October 2026:
survey enterprises production LLM deployment open-source vs proprietary models share of workloads 2026open-weight models adoption measurement usage share inference providers (arxiv.org, ssrn.com, osf.io)"OpenRouter" data representativeness LLM market shares bias developers sampleopen source AI economic value usage closed vs open models demand estimation Nagle Yue 2025 working paper (nber.org, hbs.edu, ssrn.com, aeaweb.org, sciencedirect.com)Ollama pulls self-hosted LLM usage measurement study local inference adoption"Economies of Open Intelligence" Longpre extends OR builds on 2026 open model usageNagle Yue "The Latent Role of Open Models in the AI Economy" OpenRouter share open models
Provenance
RO-LD11-02 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.