TL;DR — Key Takeaways
– Open-weight models now account for 34% of enterprise AI token usage, up from 23% a year ago, with adoption expected to reach 41% over the next 12 months.
– Enterprise production adoption of open-weight models has risen to 42%, up from 31% in July, while another 43% of respondents are conducting pilots.
– Google’s Gemma has overtaken Meta’s Llama as the most widely deployed open-weight model family, with production usage reaching 55%.
Enterprise adoption of open-weight AI models is accelerating, driven largely by the desire for lower costs and greater control over AI deployments, according to new research from Enterprise Technology Research.
The September 2026 survey of 200 respondents found that open-weight models now account for 34% of enterprise AI token usage, compared with 23% a year ago. Respondents expect that share to reach 41% over the next 12 months.
The shift is particularly pronounced at some companies. Among enterprises running open-weight models in production, 24% now use them to process more than half of their AI tokens, compared with just 3% in July. Respondents expect that proportion to reach 34% within a year.
Currently, 42% of respondents have open-weight models in production, up from 31% in July. Another 43% are conducting pilots. Only 3% have yet to evaluate the technology, compared with 11% in the previous survey.
In an AI landscape in which high-profile proprietary models from the likes of Anthropic and OpenAI seem to get all the attention, the findings suggest that open-weight AI is moving beyond experimentation to become a core component of enterprise AI infrastructure.
However, the growth of open-weight AI does not necessarily come at the expense of proprietary models. Despite gaining a larger share of enterprise AI workloads, open-weight models still attract less enterprise spending than proprietary offerings.
In short, the open-weight vs. proprietary model issue offers a mixed picture. Among enterprises running open-weight models in production, 60% have transferred some workloads from proprietary offerings. However, 93% report that their proprietary model usage continues to grow.
The result is a more diverse enterprise AI environment in which organizations select models based on workload requirements rather than committing exclusively to a single provider.
Google’s Gemma Overtakes Meta’s Llama
While enterprise adoption of open-weight AI is growing overall, not every player in the sector is a winner. Google’s Gemma has overtaken Meta’s Llama as the most widely deployed open-weight model family, with production usage rising from 35% to 55% since July. Llama experienced an equally sharp decline, falling from 55% to 35%.
Gemma also leads enterprise pilots at 51%, while Microsoft’s Phi, included in the survey for the first time, has reached 47% in pilots and 39% in production.
The results indicate that Google and Microsoft are gaining significant enterprise traction in a market previously associated heavily with Meta’s Llama.
Chinese-origin open-weight models, meanwhile, saw declining adoption. DeepSeek fell to 25% in production, down seven percentage points, while its pilot usage declined 10 points to 28%. GLM experienced a steeper decline, dropping from 10% to 1% in production.
A possible reason? The survey found that export controls and sanctions exposure became three times more common as a major consideration in production deployments.
Cost Drives Adoption, Security Slows Deployment
The cost of AI inference is a key driver of enterprise interest in open-weight models. Cost savings were cited by 69% of production users and 67% of organizations conducting pilots. Among enterprises not currently using open-weight models, 74% said lower costs could persuade them to begin evaluating the technology.
The threshold for switching is relatively modest. Assuming comparable performance, 57% of organizations conducting pilots would transfer a workload to an open-weight model for savings of 25% or less.
Still, the potential savings must be weighed against the costs and risks of operating the technology. Among organizations piloting open-weight models, incomplete security and compliance reviews were the leading obstacle to production deployment, cited by 62%, up nine percentage points since July. Self-hosting costs followed at 50%, compared with 40% previously.
Model performance and quality, by contrast, accounted for just 10% of concerns.
Enterprise hosting preferences also show the desire for greater control of infrastructure. Self-managed cloud environments account for 46% of production hosting, followed by on-premises deployments at 42%.
Hyperscaler-native inference accounts for 29%, while managed inference providers represent 27%.
The bigger picture here is that enterprises are developing AI strategies built around multiple models and deployment environments. Nearly half of respondents, 49%, continuously evaluate new models as they become available.
