Market development: open versus closed models
The market for large language models has changed from a clear-cut field with a handful of pioneers into a layered ecosystem. Where the discussion previously focused primarily on performance differences between proprietary APIs and freely accessible weights, the emphasis is now shifting to commercial interests, infrastructure, and total cost of ownership. The interplay between closed and open models determines not only how vendors recoup their investments, but also how organizations structure their IT architecture.
In this analysis, we look at the economic dynamics behind this market development. We analyze which revenue models form the basis of the current AI market, how the presence of open weights creates pricing pressure on APIs, and which strategic considerations factor into making a sustainable technology choice.
Market Structure and Positioning of Players
Providers of language models can broadly be divided into three categories based on their go-to-market strategy:
- Pure closed providers: These parties make their models available exclusively through managed APIs or closed cloud environments. They target the absolute top of the market and invest heavily in large-scale training clusters. The intellectual property surrounding the model weights and training data remains strictly confidential.
- Open-weight providers: These developers make the model parameters publicly available. Users can download these weights and run them locally or in their own cloud infrastructure. These parties often do not earn their money directly from the weights themselves, but from additional services, enterprise support, or infrastructure. To find suitable models within this category, you can consult overviews in the directory for open source models.
- Hybrid players: Large platform and cloud vendors combine both approaches. They offer proprietary models for complex tasks, while simultaneously publishing lighter open models or supporting a wide range of open-source alternatives on their own cloud platforms.
Underlying Revenue Models
Building and maintaining advanced language models requires capital-intensive investments in computing power, data centers, and specialized talent. The way developers recoup these investments varies greatly by market segment.
| Revenue model | Primary revenue source | Main advantage for provider |
|---|---|---|
| API revenue per token | Consumption of input and output tokens | Direct link between usage and revenue |
| SaaS subscriptions | Fixed amount per user per month | Predictable recurring revenue |
| Cloud infrastructure | Rental of compute (GPU/TPU) and storage | Margin on the underlying computing power |
| Ecosystem lock-in | Sale of additional software and hardware | Increasing switching costs for customers |
For parties that depend solely on API revenue per token, margins are under pressure. Once comparable performance is delivered by cheaper or open alternatives, prices per token must fall to retain market share. Cloud providers are in a stronger position: they earn from the computing power needed to run both closed APIs and self-hosted open models. The extensive analysis on the hardware and chips race goes deeper into the impact of compute costs on total market value.
Why Parties Release Weights Without Direct Revenue
The at-first-glance paradoxical behavior of making powerful models available for free rests on clear strategic mechanisms. In the software industry, offering an underlying layer for free is a proven way to shift value toward one's own core activities.
First, the principle of 'commoditizing the complementary layer' plays a role. If a company earns its main income from hardware or cloud infrastructure, it benefits from the software that runs on that hardware being as cheap and accessible as possible. By releasing high-quality model weights, the marginal value of proprietary API models erodes, while demand for the computing power needed to run those open models actually increases.
Second, opening up weights serves as a powerful tool for talent recruitment and establishing an industry standard. Developers and researchers prefer to build their workflows and tooling around models they can inspect and modify without restrictions. This creates a flywheel effect in which the ecosystem around a specific model grows, helping the developer attract the best engineers.
Pricing Pressure and the Floor in the API Market
The continuous stream of high-quality open models has a direct effect on the rates of proprietary APIs. Once an open model reaches a performance level that is 'good enough' for standard applications, many organizations lose their willingness to pay a premium for a closed API.
This phenomenon pulls a price floor under the market. Providers of closed APIs are forced to periodically lower their prices per token or offer more functionality for the same price. This creates a clear split in demand:
Price-sensitive segment: Tasks such as text summarization, simple classification, information extraction, and second-line customer service are increasingly shifted to efficient open models or specialized small language models. For a deeper look at this category, see the article on small language models and their deployability.
Price-insensitive segment: Complex reasoning tasks, advanced code generation, and applications where the highest accuracy is required remain the domain of the largest closed models. Organizations deploying these models pay for the marginal performance gain that makes the difference between successful and failed automation.
The Actual Performance Gap
The gap between the absolute top of proprietary models and the best-performing open models is not a static given. There is a wave-like motion. When a new flagship model from a closed provider appears, a clear capability gap temporarily emerges. In the months that follow, open models close much of that gap, supported by more efficient training methods and better data quality.
It is important to distinguish between benchmarks and practical production value. A closed model may score a few percentage points higher on standardized tests, but in a company-specific context, a smaller, fine-tuned open model may perform better on the specific domain task. A broader overview of this development is described in the article on trends in open-source LLMs.
What 'Open' Means in Commercial Practice
In the commercial debate, the term 'open' is often used imprecisely. Making model weights available is fundamentally different from traditional open-source software.
With open weights, the user receives the trained parameters of the network. This makes it possible to run the model locally and fine-tune it. However, the source code of the training pipeline, the architectural design, and — above all — the exact training data are rarely made public. As a result, the model cannot be reproduced or audited from the ground up.
In addition, many open-weight developers impose usage restrictions through customized licenses. For example, limits are sometimes placed on the number of monthly active users, or it is prohibited to use the model's output to train competing models. For a thorough legal interpretation of these license terms, we refer to the overview on the licensing debate around open models.
Regulation and Sovereignty in Europe and the Netherlands
For European and Dutch organizations, legislation and data sovereignty play an increasingly important role in the choice between open and closed structures. The European AI Act sets specific requirements for transparency, risk management, and documentation. The obligations differ depending on the type of model and how it is deployed.
When using American API services, organizations face complex issues surrounding data transfer outside the European Economic Area (EEA), data processing agreements, and the risk of unauthorized access by third parties. This places strict demands on compliance with privacy legislation.
Open models hosted within an organization's own IT infrastructure or with a European cloud provider offer maximum control in this respect. The data does not leave the organization's own environment, which is essential for government, healthcare, and the financial sector. This need for digital sovereignty is a strong driver for the adoption of open weights in the region.
Decision Factors for Organizations
When determining the right strategy, IT decision-makers face a trade-off between convenience and control. The choice between a closed API and a self-hosted open model is rarely an all-or-nothing decision; many organizations adopt a hybrid approach.
To determine which route is most suitable, an organization can apply the following criteria:
- Speed of implementation: Closed APIs require no infrastructure management and offer direct access through simple integrations. This is ideal for validating concepts and pilots.
- Data privacy and confidentiality: When strict confidentiality is required or when privacy-sensitive data is processed, a locally hosted open model offers superior guarantees.
- Customizability: Open models can be deeply adapted to specific business processes through fine-tuning and adjustment of the model architecture.
- Cost predictability: With a high and continuous volume of requests, the token costs of APIs become substantial. Self-hosting on fixed GPU infrastructure can be more cost-effective at sufficient scale.
A detailed calculation of the infrastructure and management costs can be found in the guide on TCO of open versus closed models. For strategic guidance in defending these choices, you can consult the guidelines on building versus buying AI solutions.
Sober Outlook and Market Assumptions
The market for language models is developing rapidly, but a number of structural trends and uncertainties can be clearly formulated.
Structural Developments
- Price erosion for basic tasks: The cost of standard inference (generating responses) will continue to fall as open models and more efficient hardware push down the cost per token.
- Specialization of models: Instead of one universal model, IT landscapes will consist of a combination of small, specialized models for specific tasks and one large model for complex questions.
- Increased attention to governance: Documentation, model cards, and regulatory compliance are becoming an integral part of the selection process. Practical guidance on this can be found in the manual on reading model cards and licenses.
Explicit Market Assumptions and Uncertainties
- Assumption: That the performance gap between open and closed models remains small for everyday tasks. If closed providers make a qualitative leap through a new breakthrough that open models cannot keep up with, the market could once again escalate toward closed APIs.
- Assumption: That large tech companies continue to subsidize open weights for strategic reasons. Should shareholder pressure for direct profitability of AI divisions increase, the willingness to share advanced weights free of charge could diminish.
- Assumption: That the supply of hardware and energy grows sufficiently to meet rising demand for inference and training. Restrictions on the distribution of advanced chips could drastically change the balance of the market.
