05 Evidence & Methods
Who Actually Uses Open Models?
Where each chart’s numbers come from, in one card per chart. The full calculations are in the files named on each card.
The charts
How much
- What it shows
- Open models’ share of the tokens, the companies, the developers and the spending, each source’s latest reading; developers by region.
- Source
- Menlo Ventures, Vercel AI Gateway, Ramp AI Index, ETR/Futurum and Futurum, as published; JetBrains Developer Ecosystem Survey 2025 (CC BY 4.0) and Stack Overflow Developer Survey 2025 (ODbL), our tabulations.
- How we calculated it
- One panel per measure; one row per source, with its population; JetBrains with its weights and 95% intervals.
- Main limit
- Rows are not comparable or combinable; no source represents all companies or all developers; Menlo’s shares are modelled and Ramp’s is a proxy that also counts closed-model sales.
- Files
analysis/adoption/p29/fig_5x_volume_money_firms.csv·analysis/adoption/p29/jetbrains2025_any_open.csv·analysis/adoption/p29/so2025_open_families.csv
Use and spending
- What it shows
- Open models’ share of US enterprises’ tokens and of their spending on models, September 2026.
- Source
- Menlo Ventures, State of Enterprise AI 2026, as published.
- How we calculated it
- Two published shares from one modelled estimate; the implied relative price of a token is their identity, (spend share / token share) / ((1 − spend share) / (1 − token share)).
- Main limit
- A model built on a survey of about 400 US decision-makers; “open-source” undefined; Menlo’s September 2025 values are not compared, because its 2025 report used another unit and sample; the implied price ratio is arithmetic on modelled shares.
- Files
analysis/adoption/p29/fig_5x_volume_money_firms.csv·analysis/adoption/p29/evidence_2025_2026.csv
Inside companies
- What it shows
- Stage of adoption, reasons cited and hosting, in one panel of 200 large companies.
- Source
- ETR/Futurum open-weight models survey, September 2026 wave, as published; Ramp AI Index (June and September 2026).
- How we calculated it
- Shares as published: stage of the 200; reasons of the companies piloting (several allowed); hosting of the 83 in production (several allowed).
- Main limit
- A non-random, mostly North American panel; question wording not published.
- Files
analysis/adoption/p29/evidence_2025_2026.csv
Sources and notes
Show the 24 notes
- 1Menlo Ventures, 2026: The State of Enterprise AI, published 3 November 2026; fieldwork by Morning Consult, 8–26 September 2026; about 400 US enterprise decision-makers (exact n not published). Modelled share of US enterprise tokens and model spend (survey combined with public data). The report also restates September 2025 as 11% of tokens and 6% of spend; Menlo's 2025 report (9 December 2025, n=495) had published 11% as a share of estimated dollars, a different metric with a different sample, so the two dates are not presented as a trend (Pass 26 decision,
research/editorial/scientific-corrections/26-freshness-and-factual-corrections.md). "Open-source" is not defined. Cited as published.analysis/adoption/p29/fig_5x_volume_money_firms.csv(rowsmenlo-2026);analysis/adoption/p29/evidence_2025_2026.csv(menlo-2026,menlo-2025). - 2Eurostat,
isoc_eb_ai(dissemination API, dataset updated 15 June 2026; checked 9 November 2026). EU-27 enterprises with ten or more employees, excluding finance (about 157,000). IndicatorE_AI_AOS(open-source AI software modified by the enterprise's staff) has values for 2021 and 2024 and none for 2025. 2024: 26.98% of AI-using enterprises, 3.43% of all enterprises. CC BY 4.0.analysis/adoption/p29/evidence_2025_2026.csv(eurostat-2025);sources.md§A. - 3Method note. Linux Foundation Research, generative-AI survey 2024, Q35 ("Do you use an open source model for inference?"), fieldwork August–September 2024; n=193 organisations using generative AI for inference, of 316 complete responses recruited through Linux Foundation channels; self-selected, one respondent per organisation, unweighted; "open source model" not defined. Microdata: Zenodo 10.5281/zenodo.21112596; recount in
pipelines/adoption/recompute.py. Not repeated in 2025 or 2026. Kept as a record only (ch05-adoption.md§0 and §8, E2). - 4Reuse status by source:
analysis/adoption/p29/evidence_2025_2026.csv, columnreuse_status. Menlo, Vercel, Ramp, ETR/Futurum and Futurum: quote with citation; each figure is attributed to its publisher. JetBrains raw data CC BY 4.0, © JetBrains s.r.o. 2025; Stack Overflow ODbL 1.0 / DbCL 1.0, attribution "Stack Overflow". Sources and URLs:research/editorial/extreme-edit-pass29/ch05-adoption/sources.md. - 5Ranges from
research/editorial/extreme-edit-pass29/ch05-adoption/ch05-adoption.md§2 andanalysis/adoption/p29/fig_5x_volume_money_firms.csv: tokens 19% (Menlo, September 2026) to 71.5% (Vercel, September 2026); money 4% (Menlo) to 23% (Vercel); firms 6.4% (Ramp, September 2026) and 42% (ETR/Futurum, September 2026); developers 18.0% (JetBrains 2025, excluding mainland China, weighted) to 38.2% (Stack Overflow 2025). Ranges are not combined (DEC-047, DEC-051). - 6Average price of an open-model token relative to a closed-model token = (spend share / token share) / ((1 − spend share) / (1 − token share)), from Menlo's published shares (note 1). September 2026, 19% and 4%: 0.18, about 1/5.6; with the shares at the edges of their rounding (18.5–19.5%, 3.5–4.5%), from 1/4.8 to 1/6.7. An identity of the two modelled shares, not an observed price; it does not say which models, tasks or prices produce it.
pipelines/web/ch05_2026.py,implied_price(). - 7Vercel, AI Gateway Production Index, September 2026 (published 17 September 2026, August data) and October 2026 (published 15 October 2026, September data). Population: traffic of Vercel AI Gateway customers, mainly developers and startups; spend estimated at list prices. September 2026: 71.5% of tokens, 23% of spend; DeepSeek V4.1 Flash nearly half of the tokens and 2.6% of the spend (
ch05-adoption.md§4). April 2026: 13% of tokens under the current, broader classification (11% under the old one). Cited as published; no licence stated.fig_5x_volume_money_firms.csv(rowsvercel-gateway-*). - 8Futurum AI Tools Pulse, October 2026, n=100 enterprises (not described): open weights 5% of metered AI consumption; two-thirds use open weights "in some form" (46% in August 2026). Cited as published.
evidence_2025_2026.csv(futurum-tools-pulse-2026-10). - 9ETR/Futurum, open-weight models survey, September 2026 wave, the 83 companies with open-weight models in production: open models 34% of their AI tokens, 23% "a year ago" as recalled by the same respondents; about a quarter (24%) say open models carry most of their tokens. Cited as published.
fig_5x_volume_money_firms.csv(rowsetr-futurum-ow-2026-09, paneltokens);ch05-adoption.md§2. - 10JetBrains Developer Ecosystem Survey 2025, raw data (CC BY 4.0; released 16 October 2025), fieldwork April–June 2025. Base: 8,269 employed developers answering the optional model block ("Which AI models / LLM do you regularly use for coding…"); DeepSeek or Llama at work via web, app or API. Weighted with JetBrains' weights; 27.6% (95% CI 26.2–29.0; unweighted 17.6%); excluding mainland China, n=7,774, 18.0% (16.8–19.3). Regions: mainland China 67.6% (n=495), Africa 39.6% (364), India 26.6% (303), Brazil 24.4% (345), United States 10.4% (829), Germany 10.2% (692), United Kingdom 8.9% (366), Japan 4.5% (229), South Korea 3.9% (381). Company size: 1–50 employees 30.9% (n=2,273), 51–1,000 27.8% (2,507), more than 1,000 23.5% (2,107). Our tabulation:
analysis/adoption/p29/jetbrains2025_any_open.csv; DEC-054, DEC-055. - 11Stack Overflow Developer Survey 2025 microdata (ODbL 1.0), published 29 July 2025, fieldwork 29 May – 23 June 2025; 49,191 responses, 16,281 named at least one model (
AIModelsHaveWorkedWith, "used for development work in the past year"). Self-selected, unweighted. Open families: DeepSeek, Llama, Qwen, Phi-4: 38.2% (6,217); 40.9% including Mistral; 3.0% only open families. By organisation size: freelancers 50.9% (n=643), 2–99 employees 37.0%, 100–999 34.3%, 1,000–9,999 34.0%, 10,000 or more 37.4%. Our tabulation:analysis/adoption/p29/so2025_open_families.csv. - 12Stack Overflow Developer Survey 2026 microdata (ODbL 1.0), published 6 October 2026, fieldwork 23 June – 5 August 2026; 30,903 responses, 11,637 named at least one vendor (
LLMHaveWorkedWith, vendors used for work in the past year). Open-weight-only vendors: DeepSeek, Moonshot, Zhipu, 01.AI: 20.7% (lower bound); 32.0% including vendors that sell both (Qwen, Mistral, Meta). Not comparable with 2025 (families against vendors). Our tabulation:analysis/adoption/p29/so2026_open_vendors.csv. - 13JetBrains: note 10 (France 8.0%, n=351). Stack Overflow 2025: EU-27 34.1% (n=5,892), United States 33.1% (n=3,135), United Kingdom 30.7%,
so2025_open_families.csv. Stack Overflow 2026, open-weight-only vendors: EU-27 19.0% (n=4,347), United States 13.9% (n=2,721),so2026_open_vendors.csv; DeepSeek alone: EU-27 17.5%, United States 12.7%,analysis/infra/europe_p29/so_mistral_by_region_2025_2026.csv. - 14Stack Overflow Developer Survey 2026 microdata (as note 12), option "Mistral AI" (vendor; open and closed products). EU-27 15.2% (95% Wilson CI 14.2–16.3; n=4,347), United States 6.2% (n=2,721), France 32.7% (n=419); DeepSeek in the EU-27 17.5% (16.4–18.6). Unweighted. In 2025 (family option): EU-27 13.9%, United States 10.3%, France 32.7%. Our tabulation:
analysis/infra/europe_p29/so_mistral_by_region_2025_2026.csv;research/editorial/extreme-edit-pass29/europe/europe.md. - 15Ramp AI Index, July 2026 (8 July) and September 2026 (9 September); July value (6.1%) via Fortune, 27 August 2026. US businesses on Ramp's card and bill-pay (n not published; skews to startups and technology-forward firms). Measure: share of AI-spending businesses paying a model-serving or routing platform, a proxy that misses self-hosting and cloud-marketplace purchases and includes closed-model sales. January 4.5%, June 5.8%, July 6.1%, September 6.4% (3.6% of all businesses). These businesses spend about 23 times the median AI spend per employee (
ch05-adoption.md§4). Cited as published.fig_5x_volume_money_firms.csv(rowsramp-*,fortune-2026-08). - 16ETR/Futurum, open-weight models survey, September 2026 wave (Techstrong, 22 September 2026; Futurum, 26 October 2026). n=200 enterprises, predominantly North American, from ETR's panel of IT decision-makers; not a probability sample; question wording not published. 42% in production (31% in July 2026), 43% piloting, 3% never evaluated. Futurum AI Tools Pulse: note 8. Cited as published.
evidence_2025_2026.csv(etr-futurum-ow-2026-09,futurum-tools-pulse-2026-10);fig_5x_volume_money_firms.csv. - 17ETR/Futurum, September 2026 wave (as note 16), companies piloting open-weight models (43% of the 200); several reasons may be cited. Savings 67%; security and compliance reviews not finished 62%; cost of self-hosting 50% (40% in July); quality 10%. Production users (n=83): 69% cite savings. Cited as published.
pipelines/web/ch05_2026.py(ETR_WHY);ch05-adoption.md§4. - 18Futurum: 52% of non-users lack the skills to self-host or fine-tune (base not published). Cited as published.
ch05-adoption.md§4, "Por qué no". - 19ETR/Futurum, September 2026 wave, production users (n=83): self-managed cloud 46%, on-premises 42%, managed providers 27%; several answers possible. Cited as published.
ch05-adoption.md§4, "Control y alojamiento propio";pipelines/web/ch05_2026.py(ETR_HOST). - 20ETR/Futurum, September 2026 wave, production users: low-complexity workloads 39% to 51% of open-model use (
ch05-adoption.md§2, "Para qué"). Linux Foundation Research, 2026 Enterprise Transitions to the Agentic Era (September 2026; interviews; CC BY-ND 4.0): practitioners at IBM and Adyen describe routing routine agent work to small local open models and complex reasoning to frontier models. Qualitative; no prevalence figure.evidence_2025_2026.csv(lf-agentic-2026). - 21ETR/Futurum, September 2026 wave, production users: origin "important or decisive" 47% (35% earlier); exposure to export controls an important consideration 30% (10% earlier). Cited as published.
ch05-adoption.md§4, "Soberanía y origen" and "Origen chino". - 22Ramp AI Index, July 2026 (data to June 2026): 96.4% of AI-spending businesses paying a model-serving platform also pay OpenAI or Anthropic. ETR/Futurum, September 2026, production users (n=83): 93% say proprietary-model use is still growing. Stack Overflow 2025: note 11.
evidence_2025_2026.csv(ramp-ai-index-2026-07);ch05-adoption.md§2, "Para qué". - 23
ch05-adoption.md§2, "Por qué discrepan" (unit, population, threshold, observed or declared, what each source sees, definition, sponsor's interest). - 24Sponsor interests: Menlo's portfolio includes Fireworks, Modal and OpenRouter (
sources.md§B); Vercel sells the gateway it reports on; ETR/Futurum sell data to investors. Futurum Group press release on the acquisition of Aptiviti Inc., parent of Enterprise Technology Research (May 2026; closing 5 June 2026).sources.md§B;evidence_2025_2026.csv, columnpublisherofetr-futurum-ow-2026-09.