The economics of artificial intelligence once appeared to favor a small group of companies able to train very large proprietary models. Open models have complicated that picture. Developers can now download strong model weights, run them on their own infrastructure, adapt them to a domain, and choose among competing hosting providers. The result is not the end of proprietary AI. It is a broader market in which control of a model is separated from access to a model service.
That distinction starts with vocabulary. Open source traditionally means that users can inspect, modify, and redistribute source code under a qualifying license. An AI release may expose weights while withholding training data, data-cleaning code, or the full training recipe. The Stanford AI Index therefore cautions: “Open-weight models, though more accessible than closed-weight models, are not necessarily fully open source.” Calling every downloadable model open source hides meaningful differences in licenses, transparency, and reproducibility.
Performance is becoming less scarce
Open-weight releases have reduced the premium attached to raw model access. Stanford's 2025 AI Index found that the gap between the leading closed- and open-weight models on the Chatbot Arena leaderboard narrowed from 8.0% in early January 2024 to 1.7% by February 2025. On MMLU, the gap fell from 15.9 percentage points in late 2023 to 0.1 point by the end of 2024. Benchmarks do not measure every production requirement, but such convergence changes buyers' negotiating position.
Distribution reinforces that effect. Meta reported in March 2025 that its Llama family had passed one billion downloads since the first release in 2023. A download is not a unique user or a production deployment, and Meta has a commercial interest in emphasizing adoption. Even so, the scale documented in Meta's announcement shows how quickly weights can become common infrastructure rather than a metered endpoint.
Organizational data points in the same direction. A Linux Foundation literature review and survey synthesis found 89% of AI-adopting organizations used some open source in their AI stack, with open components averaging 41% of that infrastructure. Sixty-three percent reported using an open model. The study was commissioned by Meta, so its sponsorship should inform interpretation, but its disclosed figures help describe a market that is already hybrid rather than purely open or closed.
Value moves up and down the stack
When base weights are widely available, companies can compete elsewhere. Chipmakers and cloud providers sell the computation needed for training and inference. Model hosts sell reliability, scaling, security controls, and managed updates. Specialist firms sell fine-tuning, evaluation, retrieval systems, and integration. Application developers can substitute one compatible model for another, reducing lock-in—provided they have built portable data and evaluation layers.
Open weights also change the build-versus-buy calculation. An organization with sensitive data may run a model in its own environment; another may pay for a managed open model because operating GPUs is not its core competence. Access to free weights does not make inference free. Hardware, electricity, engineers, monitoring, and incident response remain real costs, and reproducing a training run can still be far beyond the budget of most firms.
The competitive pressure reaches proprietary vendors as well. Closed providers must justify their margins through higher quality, distinctive tools, service guarantees, or safety and compliance features. Meanwhile, open-model vendors can monetize hosting and enterprise support. The model itself increasingly functions as one layer in a product bundle.
Openness creates governance choices
The regulatory treatment of open releases recognizes both their benefits and limits. The European Union's AI Act gives certain transparency exceptions to general-purpose models released under free and open-source licenses when parameters, architecture information, and usage information are public. But the exceptions do not remove obligations to publish a training-content summary or maintain a copyright-compliance policy, and they do not cover models classified as presenting systemic risk.
Security presents a similar dual effect. External researchers can inspect and test an accessible model, potentially finding faults faster. The same accessibility allows malicious users to remove safeguards or adapt capabilities. Governance therefore cannot rest on a binary label. Buyers need model cards, license review, provenance records, vulnerability reporting, reproducible evaluations, and clarity about who will maintain a release.
A market of complements
Open models are unlikely to produce a single winner. They encourage specialization: smaller regional-language models, domain adaptations, local inference, and ensembles that route tasks by cost and difficulty. Proprietary frontier systems may continue to lead on difficult capabilities, while open alternatives put a ceiling on what customers will pay for routine ones.
The deepest economic change is that model intelligence is becoming more substitutable. Scarcity shifts toward trusted data, distribution, compute, integration, and the ability to operate systems safely. In that environment, openness is not merely a licensing philosophy. It is a force that redistributes bargaining power across the AI supply chain.