Evidence Shows Enterprises Use of Open Weight Models is Mainstream

About three weeks ago I wanted to find evidence to support the claim that more and more enterprises are moving to open-source, open-weight models and away from public models. My trained research assistant, Chuck, told me to publish that – there was not enough evidence. Times have changed; yes, that quickly.

Chuck writes…

The evidence has now crossed the threshold decisively. The important development since the last check is that we now have enterprise-specific longitudinal evidence, rather than having to infer enterprise behavior primarily from developer/gateway traffic.

Enterprise Technology Research data published September 22 is the strongest addition. Its September survey of 200 enterprise respondents reports that open-weight models now account for 34% of enterprise AI token usage, up from 23% a year ago. Among enterprises already running open-weight models in production, 24% now send more than half their tokens through them, versus only 3% in July. Production adoption increased from 31% in July to 42% in September, with another 43% piloting. Techstrong.ai

More important for the thesis, this is finally evidence of workload migration rather than merely additional adoption. Among organizations running open-weight models in production, 60% report transferring some workloads from proprietary models. Yet 93% say their proprietary-model usage is also increasing. That substantially strengthens the bifurcation argument: enterprises are not abandoning frontier providers; rapidly expanding AI consumption is allowing open/private inference to capture workloads inside a growing overall market. Techstrong.ai

The placement data is equally useful. Among those production open-weight deployments, 46% use self-managed cloud environments and 42% use on-premises infrastructure, compared with 29% using hyperscaler-native inference and 27% managed inference providers. Cost savings were cited by 69% of production users. Security/compliance remains the leading obstacle among pilots at 62%, while self-hosting cost has actually become a larger concern, rising from 40% to 50% since July. Techstrong.ai

That last number is valuable counterevidence. Private inference is not simply winning because self-hosting is cheaper. Enterprises appear increasingly willing to accept infrastructure complexity where economics, control, sovereignty or workload characteristics justify it.

A separate September 23 study provides another useful enterprise-level signal. The Tokenomics Foundation surveyed 472 organizations across 11 industries representing $4.6 trillion in aggregate revenue. 51% currently describe their model mix as heavily frontier-weighted; only 24% expect to remain that frontier-heavy in twelve months. The most frontier-dependent organizations anticipate the largest shift. Already, 29% report private hardware deployment, 31% rented GPU infrastructure and 32% edge/local inference. Tokenomics Foundation

The connection to AI investment decisions is becoming clearer as well. EY found in July that 98% of senior leaders using token-based AI said token usage and costs had caused their organizations to reconsider their approach. This is not evidence specifically of private migration, but it establishes the economic pressure capable of driving routing and placement decisions as AI scales. EY

And Accenture’s August open-weight analysis supplies an important correlation: roughly one-quarter of organizations deliberately balance open-weight and proprietary models, and those organizations are more than twice as likely to report AI operating at full production scale across functions—61% versus 29% for organizations relying on a single model tier. This does not establish causality, but it connects multi-model architecture with enterprise AI commitment and production maturity. Accenture Newsroom

The original thesis still needs one correction. I would not write that enterprises are migrating from frontier APIs to private inference. The data supports something more interesting:

Enterprise inference is fragmenting. As AI consumption scales, enterprises are retaining frontier models for workloads where premium intelligence earns its cost while moving a growing share of production volume toward open-weight models and infrastructure they can operate or control.

That distinction makes the argument more original. The story is not open versus closed. It is the emergence of an enterprise inference portfolio in which model capability, workload economics, data sensitivity and control increasingly determine where each unit of intelligence runs.

And now back to me…

I initially saw four signals, that have now converged: token share is moving; actual proprietary workloads are being transferred; privately controlled infrastructure is receiving meaningful production deployment; and AI economics are becoming important enough to alter enterprise investment decisions. The thesis is becoming reality: more and more enterprises are adopting open weight because they get both better results, and better economics. They increasingly distribute inference among frontier APIs, hosted open models, private cloud, dedicated infrastructure and on-premises models. Alas, the value is not the multi-model strategy, it is the enterprise's ability to determine dynamically which model runs which workload, with what context, under what economics and conditions.

And now, we focus our efforts in promoting Private Platforms to manage that routing and multi-model usage. Stay tuned for that content…