In almost every wave of computing, open source eventually wins. Linux for servers, Android for phones, Kubernetes for cloud, PyTorch for machine learning and so on. Yet in AI the story so far looks inverted. The headlines are dominated by OpenAI, Anthropic, and Google. The models that capture most mindshare are closed, expensive, and tightly controlled. But the open model ecosystem is evolving faster than many realize, raising the question: will AI be the first domain where open source doesn’t prevail, or just the next one where it inevitably does?

In their 2024 State of Generative AI in the Enterprise, Menlo Ventures found “Closed-source solutions underpin the vast majority of usage with 81% market share, while open-source alternatives (led by Meta’s Llama 3) hold steady at 19%, dropping just one percentage point from 2023.”

‍

‍

That data already feels dated. In 2025, the open model ecosystem has accelerated dramatically. As Nathan Lambert recently summarized, one of the year’s biggest storylines has been the rise of highly capable Chinese open models like Qwen, DeepSeek, and Kimi.

‍

Which led to this recent flip: Chinese open models overtaking Western models in downloads.

‍

Market share

What remains murky is relative market share: are open models gaining or losing ground? Once a developer downloads a model from Hugging Face, usage becomes private by design. While download counts don’t reveal absolute usage, they do show directionally strong growth. Cumulative open-source model downloads on Hugging Face rose from under 50 million at the start of 2024 to over 1 billion by October 2025.

Over the same period, OpenAI’s weekly active users grew from 100 million to 800 million. By that metric, open models are keeping pace with overall AI growth.

‍

‍

Closed-open performance gap

Performance gaps between open and closed models have remained remarkably stable. The top open models typically trail the frontier labs by 3-6 months. Given the billions spent by the frontier labs - and an intense global talent war - the persistence of that narrow gap is striking.

‍

Enterprise adoption

The latest Chinese models Qwen, Deepseek and Kimi are quickly closing the gap to the Western frontier labs. Omer Cheema posted that “80% of Bay Area startups are building on Chinese open source models”. 

For large enterprises, the picture looks different. Many lack the in-house talent to deploy full-stack AI solutions and are wary of Chinese models. They either hire OpenAI or Anthropic for bespoke deployments or bring in system integrators like Deloitte or Accenture. 

How are enterprises choosing between open vs closed models today? Here is a basic decision flow:

• Testing a new AI use case? → Use a frontier lab API for quick experimentation.
• Need frontier capabilities? → Stick with the top closed model for best results.
• Scaling a commercial use case? → Rising inference costs drive interest in open stacks.
• Data privacy a key concern? → Open models enable on-premise or private cloud deployment.
• Control and predictability? → Closed models evolve rapidly, often without notice: API behavior, costs, or output quality can shift overnight. Open models offer greater stability and control.
• In-house talent available? → Open model stack for scaling use cases is viable and cheaper.
• Deploying on edge devices? → Open models are essentially the only option.

As deployment tooling matures and AI engineering knowledge spreads beyond a few experts, open models will increasingly become a serious enterprise alternative.

‍

What history teaches us

The AI software stack is still unsettled. Capabilities evolve weekly, and the big frontier labs dominate both media and enterprise mindshare. But history has shown over and over again that open source is a powerful force and hugely popular with developers.

Let’s look at well-known closed developer ecosystems:

• Apple (iOS/macOS, Swift, Xcode)
• Microsoft (.NET, Visual Studio)
• Salesforce (Apex)
• Adobe (Creative Suite)
• SAP (ABAP, HANA)
• Industrial Automation (Siemens, Rockwell)

Compared to prominent open ecosystems:

• Cloud/DevOps (Linux, Kubernetes, Docker, Terraform)
• Web (React, Vue, Node.js, Rails)
• Data infrastructure (PostgreSQL, MySQL, Apache Spark, Airflow)
• Servers (Linux)
• Languages (Python, Go, Rust, Java)
• Crypto (Bitcoin, Ethereum)

The pattern is clear: when a platform controls the customer base, it can enforce a closed stack. When the developer community controls innovation, open ecosystems win.

‍

The future of open models

The future market share of open models depends on who controls the user experience. If generative AI becomes a general-purpose layer—like cloud computing or databases—then open architectures will steadily gain ground. But if end-user experiences are dominated by sticky, closed applications (ChatGPT, Claude, Google Workspace, Windows), proprietary models will retain their market share lead.

OpenAI and Anthropic understand this. Their true moat isn’t the model, it’s the interface. That’s why OpenAI is pushing deeper into the enterprise stack. It’s why they released a browser and are working on a consumer device. And why Anthropic is building full agentic coding environments rather than just selling tokens to apps like Cursor.

1. Engineering knowledge diffusion and stack maturity - AI deployment needs to become as productized and accessible as web development.
2. Maturing enterprise use cases - inference costs, avoiding lock-in and desire for control and data privacy will push companies toward open infrastructure.
3. Open models performance - Open models must continue to stay close to frontier lab performance. If the frontier starts slowing down, open models will catch up quicker and vice versa.

The first two are nearly inevitable but will take a few years to play out. The third is the one to watch.

Closed models may define the frontier today, but open models will play a foundational role, quietly powering startups, products, and edge devices the way Linux powers the internet today. In the end, open models might not win the public attention, but they’ll win the infrastructure.

‍