// Chapter 05 · Adoption

Who Actually Uses Open Models?

In US enterprises, open models processed 19% of the tokens sent to AI models but received only 4% of the spending on those models, in Menlo Ventures’ estimate for September 2026.

Data: Menlo Ventures, Vercel AI Gateway, Ramp AI Index, ETR/Futurum and Futurum (2025-2026, cited as published); JetBrains Developer Ecosystem Survey 2025 (CC BY 4.0) and Stack Overflow Developer Survey 2025 (ODbL), our tabulations.

4%

of US enterprises’ spending on AI models went to open models in September 2026, while they processed 19% of the tokens (Menlo Ventures).

Menlo Ventures 2026.

Illustration: layered landscape with a pale sun
Adoption

Where open models are already in use, they grow fast. On Vercel’s AI Gateway, used mostly by developers and startups, they went from 13% of the tokens in April 2026 to 71.5% in September. In a panel of 200 large companies, those with open models in production rose from 31% to 42% between July and September. But this adoption adds to closed models rather than replacing them, and counted in companies it is still small: only 6.4% of US businesses that pay for AI pay a platform that serves them.

01 How much

Open models take between a fifth and most of the tokens, depending on the source you look at

Menlo Ventures puts open models at 19% of the tokens US enterprises sent to AI models in September 2026. Among companies that already run them in production the figure rises to 34% of their own tokens, and on Vercel’s gateway, with developer and startup traffic, to 71.5%.

How much of AI use goes to open models: four measures, each source on its own line

Share of the tokensVercel AI Gatewaydevelopers and startups, Sept. 202671.5%ETR/Futurumtheir own tokens, companies with open models in production34%Menlo VenturesUS enterprises, modelled, Sept. 202619%Share of the developersStack Overflowused at least one open family, 202538.2%JetBrainsuse DeepSeek or Llama at work, 202527.6%Share of the companiesETR/Futurum200 large companies, in production, Sept. 202642%Ramp AI IndexUS businesses paying for AI, Sept. 2026 index6.4%Share of the spendingVercel AI Gatewayat list prices, Sept. 202623%Futurum100 enterprises, Oct. 20265%Menlo VenturesUS enterprises, modelled, Sept. 20264%0%25%50%75%100%

Share of the tokens

  • Vercel AI Gatewaydevelopers and startups, Sept. 202671.5%
  • ETR/Futurumtheir own tokens, companies with open models in production34%
  • Menlo VenturesUS enterprises, modelled, Sept. 202619%

Share of the developers

  • Stack Overflowused at least one open family, 202538.2%
  • JetBrainsuse DeepSeek or Llama at work, 202527.6%

Share of the companies

  • ETR/Futurum200 large companies, in production, Sept. 202642%
  • Ramp AI IndexUS businesses paying for AI, Sept. 2026 index6.4%

Share of the spending

  • Vercel AI Gatewayat list prices, Sept. 202623%
  • Futurum100 enterprises, Oct. 20265%
  • Menlo VenturesUS enterprises, modelled, Sept. 20264%
Sources: Vercel AI Gateway; ETR/Futurum; Menlo Ventures; Ramp AI Index; Futurum AI Tools Pulse; JetBrains Developer Ecosystem Survey 2025, with JetBrains’ weights (© JetBrains s.r.o. 2025, CC BY 4.0), and Stack Overflow Developer Survey 2025 (ODbL): our tabulations of their microdata.

Two patterns recur. The closer a population is to developers, the higher the share: 38% of Stack Overflow’s developers used an open family, against 6.4% of US businesses that pay for AI. And in the two sources that measure both tokens and spending, open models’ share of the tokens is much larger than their share of the spending.

02 Use and spending

Open models took 19% of the tokens but only 4% of the spending

In Menlo Ventures’ estimate, that means an open-model token cost on average between a fifth and a seventh of a closed one. It is a division of two modelled shares, not an observed price. Vercel’s gateway, with another population, shows the same distance: 71.5% of the tokens and 23% of the spending.

Open models’ share of the tokens and of the spending, in the two sources that measure both

Menlo Ventures US enterprises, modelled, Sept. 2026

Tokens19%

Spending4%

Vercel AI Gateway developers and startups, Sept. 2026

Tokens71.5%

Spending23%

Menlo Ventures: a modelled estimate built on a survey of about 400 US decision-makers (Morning Consult, September 2026). Vercel: tokens through its AI Gateway and spending estimated at list prices. Figures as published.

Volume concentrates in cheap models. In the latest index of Vercel’s gateway a single inexpensive model, DeepSeek V4.1 Flash, carried nearly half of the tokens for 2.6% of the spending. And in ETR/Futurum’s panel, the companies with open models in production say a growing part of that use is low-complexity work. The sources draw a split: a growing part of the volume goes to open models, and most of the money still goes to closed ones.

03 Inside companies

In one panel of 200 large companies, 42% had open models in production and 43% were piloting them

Companies that adopt open models do not drop closed ones. Of the US businesses on Ramp that paid a platform serving open models in June, 96% also paid OpenAI or Anthropic, and 93% of the companies with open models in production in ETR/Futurum’s panel say their use of closed models is still growing. Using open weights does not mean running them in-house either: companies in production run them in cloud accounts they manage themselves, on their own premises or through managed providers, and some combine several routes.

Those who adopt open models keep paying for closed ones

Two different populations, never added, and neither represents all companies.

  1. 96.4%

    of the US businesses on Ramp that paid an open-model platform also paid OpenAI or Anthropic

    Ramp AI Index, June 2026: businesses paying for AI by Ramp card or bill

  2. 93%

    of the companies with open models in production say their use of closed models is still growing

    ETR/Futurum, September 2026: 83 large companies with open models in production

Ramp observes payments to platforms that also sell closed models; ETR/Futurum records what one person answers for each company, in a panel that is not random. Figures as published.

Inside one panel of 200 large companies, September 2026

Where the 200 are

  • In production 42%
  • Piloting 43%
  • Neither yet 15%

Why, for the companies piloting them

For Against

  • Savings67%
  • Unfinished security reviews62%
  • Cost of hosting them50%
  • Model quality10%

Where those in production run them

  • Own cloud account46%
  • Own premises42%
  • Managed provider27%
ETR/Futurum open-weight models survey, September 2026 wave: 200 large, mostly North American companies, one respondent each; 3% had never evaluated open models; not a random sample. Reasons and hosting allow several answers. Figures as published.

What holds back the move from pilots to production is mostly the work around the model: unfinished security and compliance reviews and the cost of hosting weigh far more than quality, while savings are the main reason to try it. Counted in companies that pay, adoption is still small: Ramp counts 6.4% of US businesses that pay for AI also paying a platform that serves open models, up from 4.5% in January.

Chapter 05 · Closing

How much companies use open models

The published sources agree on three facts. Open models process a meaningful share of the tokens (a fifth in US enterprises according to Menlo Ventures, most of them on Vercel’s gateway) and receive a much smaller share of the spending. Their use grows where they have been tried. And on Ramp, almost every company that pays a platform serving them keeps paying OpenAI or Anthropic too. Our reading: for now, open models are adopted alongside closed ones, not instead of them, and they take mostly the cheap volume. In ETR/Futurum’s panel, the firms still trying them out cite unfinished security reviews and the cost of hosting far more often than quality.

Use and spending
Open models process a meaningful share of the tokens (a fifth in US enterprises according to Menlo Ventures, most of them on Vercel’s gateway) and receive a much smaller share of the spending.
Alongside closed models
Companies that adopt open models do not drop closed ones. Of the US businesses on Ramp that paid a platform serving open models in June, 96% also paid OpenAI or Anthropic.
From pilots to production
What holds back the move from pilots to production is mostly the work around the model: unfinished security and compliance reviews and the cost of hosting weigh far more than quality.