Open source AI models are becoming more mainstream among enterprises, governments, and startups, drawn by the promise of lower costs, more control, and more flexibility. But as increasingly capable models become freely available, a harder question follows: if anyone can download the model, who actually makes money — and what does building on someone else’s open model really cost?
This article answers that question. It explains what “open” really means when applied to AI models, why adoption is accelerating, who builds the leading open models, how companies profit when the model itself is free, and where the risks and hidden costs sit.
What open source AI actually means
The word “open” covers several different things, and the distinction matters. Most of what the industry calls open source AI is better described as open-weight AI: the trained model parameters — the weights — are published so anyone can download, modify, and run the model themselves. Downloadable weights alone do not establish open-source status, and licence conditions vary from model to model.
That distinction matters for control, safety, and reuse. Downloadable weights give a team real independence: the model can run on its own servers and be fine-tuned on private data, subject to the licence conditions that apply to that model. But weights alone do not reveal how the model was trained, what data it learned from, or whether it can be fully reproduced. Several of the largest models marketed as open carry custom licences with conditions — for example, extra terms for managed services above certain revenue thresholds — rather than a plain permissive licence such as MIT or Apache 2.0.
Hugging Face sits at the centre of this ecosystem as a model distribution hub. It is where models are published, discovered, downloaded, and discussed, and where derivative work — fine-tuned models, adapters, benchmarks, and applications — accumulates. Its scale — more than 2 million public models — makes it a useful window on which open models get used, alongside model quality.
Why open models are booming now
Three drivers explain the adoption wave: lower costs, more control, and more flexibility. A team that runs an open model pays for its own compute rather than per-token API fees, can choose exactly where its data goes, and can adapt the model to its own use case instead of accepting a vendor’s one-size-fits-all behaviour.
The scale of the shift is visible in the hub’s own numbers. In 2025, Hugging Face grew to 13 million users, more than 2 million public models, and over 500,000 public datasets, with users, models, and datasets all close to doubling. More tellingly, users are increasingly creating derivative artifacts — fine-tuned models, adapters, and applications — rather than only consuming pre-trained systems. That is the sign of an ecosystem being built on, not just browsed.
Adoption has reached the largest companies: over 30 percent of the Fortune 500 now maintain verified accounts on Hugging Face. At the same time, the ecosystem remains highly concentrated — the top 200 most downloaded models, just 0.01 percent of models on the platform, account for 49.6 percent of all downloads. Open source AI is booming, but attention still pools around a small number of winners.
Who builds the open models
The competitive field has a clear shape. On one side are American and European open models such as Meta’s Llama and Google’s Gemma, alongside specialist labs such as Mistral. On the other side are Chinese labs — DeepSeek, Alibaba’s Qwen, and Moonshot AI’s Kimi — whose models have become, in the words of one policy analysis, the global default for open-weight AI.
Developers who want a model they can download, modify, and run themselves increasingly start with Qwen or DeepSeek. DeepSeek’s models carry a permissive MIT licence. Qwen, meanwhile, is the other name developers increasingly start with, alongside DeepSeek.
A chief reason for this lead is cadence. American open releases are voluntary and one-off, which leaves developers unable to plan around future releases.
How anyone profits when the model is free
The answer is that the money has moved from the model to everything around it. The value is shifting away from the model itself and toward execution, customization, and distribution. Companies convert open models into revenue in several concrete ways.
First, hosted services: users can run a model themselves, but they pay a company for hosted access, managed inference, and uptime guarantees. Second, cloud deployment: cloud providers and specialist platforms charge for the managed infrastructure — optimized compute, scaling, and compliance — that lets enterprises run models inside their own secured environments. Third, enterprise support and governance: firms pay for reliability, security scanning, access controls, audit logs, and priority support, the features that turn a public download into a production system a risk committee will approve.
Fourth, fine-tuning and specialization: customers pay to adapt a general model to their domain, data, and workflows. Fifth, commercial licences: some labs charge for certain rights or enterprise terms even while the base model is free. And sixth, consulting: businesses pay for implementation and customization of open models.
The common thread is friction. Moving a high-performing open model from a public repository into a secure, compliant production environment is hard, and companies pay to have that friction removed. When the model is free, the infrastructure for running it becomes the valuable asset.
The costs, risks, and trade-offs
Building on open models is not free, and the costs are easy to underestimate. Running, maintaining, and securing models in production requires GPU infrastructure, DevOps expertise, and ongoing model management — costs that teams must assess before committing to self-hosting. Self-hosting shifts the bill from API fees to hardware, engineering time, and operations.
Licensing is a second trap. Most models called open-source do not meet that definition in full, and several carry conditional terms. A team that builds a commercial service on a model with revenue-based conditions may owe fees or face restrictions it did not anticipate. Reading the licence file, not the marketing, is a necessary step.
Policy and security risks are a third consideration, sharpest for Chinese models. In October 2026, the US Department of Homeland Security barred DeepSeek from federal agencies despite its MIT licence — a reminder that a permissive licence does not settle the regulatory question. Congress has also investigated companies using Qwen and Moonshot models.
Finally, there is a capability question to manage: the supplied research does not establish current comparative capability between open and frontier proprietary models. A team should validate an open model against its own data before committing, rather than trusting a benchmark chart.
What to watch next
Three questions will decide whether the open source AI boom lasts. First, can American players keep pace with Chinese open models — not just on benchmarks, but on release cadence and developer trust? Second, will open models stay competitive as frontier proprietary models advance, or will the gap widen again? And third, can open source AI become sustainable rather than subsidised — funded by real revenue around the models, rather than by venture capital or state support?
What is already clear is that companies seek revenue through execution, customization, and distribution around open models. The model itself is increasingly a shared resource, and how the eventual profits from the boom are distributed remains unresolved.
