On January 27, 2025, a major surprise shook the world of artificial intelligence: Deepseek, an application developed by the Chinese company Hangzhou DeepSeek, reached the top of the free download charts on the App Store in both the United States and China, even surpassing OpenAI’s ChatGPT (Ng et al., 2025). This meteoric success is driven by DeepSeek-R1, a large open-source language model (LLM), launched only a week earlier. Its popularity is explained by an attractive performance-to-price ratio and by its permissive open source license (MIT), which allows free commercial use, unlimited modifications, and derivative development. This phenomenon illustrates a growing trend in artificial intelligence (AI): the rise of open-source LLMs. Unlike proprietary models like OpenAI’s GPT-4 or Google DeepMind’s Gemini, these models rely on transparency, collaboration, and free distribution. Yet behind the rhetoric of openness and community lies a calculated business strategy, one that leverages the language of open source to compete for market share, talent, and ecosystem dominance.

LLM, open source, proprietary: what are we talking about?

An LLM is an artificial intelligence model capable of understanding and generating text in an advanced way. It is trained on massive amounts of data and can be used for various tasks: answering questions, writing text, translating languages, coding, etc. These so-called LLMs can thus be either proprietary or open source:

• A proprietary LLM (like GPT-4, which was initially open source, or Gemini) is a closed model. Its code and weights (the parameters that allow it to function) are not accessible to the public. To use it, one must go through a paid application programming interface (API) or a platform controlled by the company that developed it (Pereira, 2023).

• An open-source LLM (like LLaMA 70B, Mistral, Falcon 180B or DeepSeek-R1) is a model whose weights and code are publicly accessible. Anyone can download, modify, and use it, sometimes without any commercial restrictions (ibid.).

As a result, the rise of open-source LLMs is disrupting the market, especially in a context where proprietary models are increasingly expensive to develop and use. Thus, companies like Meta (LLaMA 70B), Mistral AI (Mistral 7B, Mixtral), and Hangzhou DeepSeek (DeepSeek-R1) advocate for a more open vision of AI, but still face a major challenge: how to monetize a technology that everyone can download and use for free?

Why is open source becoming popular among stakeholders?

First of all, for regular ChatGPT users, one might wonder what the point of open-source LLMs is, knowing that the GPT-4 model works so well. It is important to know that open-source models are increasingly successful because they address a crucial need: making artificial intelligence more accessible, more flexible, and less expensive. Although they require costly graphics processing unit (GPU) infrastructure to operate, open-source LLMs are free with no license fees, whereas proprietary models follow a pay-per-token model, making large-scale use expensive (Patel, 2025). For instance, models with 8K of context length offered by OpenAI such as GPT-4 and GPT-4-0314 cost $0.03 for 1,000 input tokens and $0.06 for output (OpenAI, 2025). Next, open source enables total customization by adjusting data and parameters, whereas proprietary models, though immediately performant, remain black boxes that are difficult to adapt. Finally, open-source models also offer greater transparency. Indeed, access to the architecture allows, for example, compliance with current laws regarding ethics (European Commission, 2024) and enables cybersecurity experts to identify and fix vulnerabilities, making the model more robust. France, for example, invested in Mistral AI to promote sovereign AI, independent from American tech giants, particularly since Donald Trump’s reelection (Becel, 2025).

The business model: direct vs. indirect monetization

On one side, proprietary models like GPT-4 or Gemini fit into a classic model of rent and exclusivity. OpenAI, for example, adopts an "API-first" model, which means its model is not directly accessible but only usable via a paid API. This allows it to monetize every query, offer subscriptions or licenses, and avoid sharing its model, thereby ensuring total control over its use. This strategy guarantees a rapid return on investment, all the more crucial since the training costs for state-of-the-art models reach several hundred million dollars. For instance, GPT-4 is estimated to have required over $100 million in computing on supercomputers equipped with tens of thousands of GPUs (mainly Nvidia A100 or H100, with a unit price above $30,000) (Krim, 2025). On the other side, open-source LLMs, while seemingly free, rely on several economic strategies to generate revenue, by monetizing not the model itself, but everything surrounding it.

1. Integrating an Ecosystem and Imposing a Standard

One of the most effective levers for open-source players is their ability to impose a technological standard by integrating their models into a broader ecosystem. This strategy doesn’t aim for immediate return on investment, but long-term dominance of the technological landscape. Take Meta and its LLaMA 2 model: instead of directly commercializing its LLM, the company distributes it under a restricted license that prohibits certain commercial uses while allowing research and development. The goal is to encourage massive adoption by developers, startups, and academic institutions. By facilitating application creation around LLaMA, Meta stimulates external innovation while capturing benefits indirectly and reducing hallucination risks, for example. These innovations can then be reintegrated into Meta’s services to optimize ad targeting, boost user engagement, or improve automated moderation. LLaMA was recently integrated into WhatsApp. More importantly, this targeted distribution allows Meta to progressively impose its technological standards. The more developers adopt LLaMA as the basis of their projects, the more the ecosystem structures itself around its architectural choices, APIs, or data formats. This creates technological lock-in, where alternatives become costly or incompatible, consolidating Meta’s dominant position. Mark Zuckerberg admits: "I don’t expect LLaMA to generate a large amount of revenue in the short term, but in the long term, it could become something important." (Habibi, 2025). The company gains visibility, attracts talent to its technical infrastructure (via AWS SageMaker or NVIDIA NeMo), and reduces development costs by sharing effort with third-party partners (Jullien & Viseur, 2021). In short, imposing a standard via open source accelerates adoption and structures innovation around its tools. It is a subtle but effective way to consolidate leadership in artificial intelligence, without relying on direct monetization.

2. Selling Premium Services

The monetization of value-added services and the B2B market are two essential pillars of the open-source LLM business model. Although these models are free, using them effectively remains costly in terms of infrastructure, software optimization, and maintenance. Companies have thus taken advantage of this reality by offering paid APIs, allowing developers to access the models without managing the servers themselves (Subramanian, 2025). This allows for custom optimizations to reduce execution costs. This approach is particularly relevant to the B2B market, where large companies seek reliable, secure, turnkey solutions. Rather than downloading an open-source model and handling its hosting and maintenance themselves, they prefer to pay for API access with technical support and continuous updates. For example, a company wishing to integrate Mistral 7B or Mixtral into a chatbot or virtual assistant can choose to deploy it internally or use Mistral AI’s API, which guarantees an optimized service and technical support.

3. Fine-Tuning

Next, another important aspect of monetizing open-source LLMs lies in offering fine-tuning and optimization services. Fine-tuning allows LLMs to be adapted to specific use cases without starting from scratch or investing significant resources to adjust an existing model. This is especially attractive for companies that lack the technical expertise or means to carry out this customization themselves (Vake et al., 2025). For example, Mixtral 8x7B, a model launched in 2023 by Mistral, relies on a Mixture of Experts (MoE) architecture. This model has 46.7 billion parameters, but by activating only a fraction of them per request, it reduces inference costs while maintaining high-level performance (Mistral AI, 2024). This approach allows companies to not only access a powerful model, but also to optimize its usage, thus reducing the operational costs associated with running these models. The fine-tuning services offered by companies like Mistral therefore allow users to personalize their models without having to manage the complex infrastructure required to do so.

4. Finding a Niche

Finally, as open-source LLMs multiply, some stand out by adopting a specialized approach, targeting specific use cases rather than a generalist model. On February 21, 2024, Google launched Gemma, an open-source model designed for researchers and developers, based on the same work as Gemini. Google immediately positioned Gemma by emphasizing its optimization for scientific and technical applications, while incorporating commercial incentives such as free Google Cloud credits to boost its adoption, again with the goal of integrating this open-source LLM into an ecosystem (Driss, 2024). This differentiation allows open-source LLMs to find lucrative niches, where customization and sector-specific expertise are essential. Rather than directly competing with giants like OpenAI, these specialized models meet specific needs, attracting companies and institutions willing to pay for optimized, ready-to-use AI in their field.

To conclude, the monetization of open-source LLMs reveals a deeper truth: openness, in today’s AI economy, is less a value than a market tactic. By offering free models, developers drive adoption, shape standards, and monetize through services or influence. As this model scales, proprietary players may be forced to rethink closed systems or double down on exclusivity. The emergence of open-source players like DeepSeek is not just a technical achievement, it’s a geopolitical maneuver. By open-sourcing high-performing models, actors beyond the U.S.-dominated AI sphere gain leverage over foundational structures of the data economy. It’s not merely about access; it’s about asserting control over the standards, architectures, and flows of intelligence that define global power. Open-source LLMs are reshaping how we think about ownership, innovation, and influence. The real divide may no longer be open vs. closed, but between those who engineer the ecosystem and those subject to it. Recognizing this shift is essential to understanding where power and agency will lie in the next phase of AI.