The formula is well known and has been the guiding principle of the digital economy for 20 years: “if it’s free, you’re the product“. This is the rule of the game in Web 2.0, with platforms offering services and content for free in exchange for personal data and behavior that can be exploited for advertising purposes. The deal was clear, even if it wasn’t always well understood by users, at least in the beginning.
But the emergence of generative AI is challenging this equation because even when users pay, they continue to be integrated into the economic model in a more profound and systemic way
The user remains a product, but not quite in the sense we understood it yesterday. They are becoming an active link in the value creation process, and that is the real breakthrough.
In short:
- Generative AI is changing the digital business model: even when paying, users remain a source of value by contributing to model training through their interactions.
- Each interaction with AI constitutes learning data that feeds into a cycle of continuous improvement, accelerating model performance through the collection of human knowledge and reasoning.
- This dynamic promotes increased market concentration, driven by the combined effects of data, massive computing resources, and lock-in mechanisms.
- Current regulatory frameworks (GDPR, antitrust) are ill-suited to these challenges, as AI exploits qualitative micro-learning and creates technological gaps that are difficult to bridge without dominating market share.
- Unlike traditional SaaS, generative AI integrates the user into a continuous R&D loop, reinforcing lock-in and making it increasingly complex to switch providers over time.
A change of facade, no logic
Generative artificial intelligence is not based on free business models financed by advertising, which are not without their problems given the specific nature of this sector (Does OpenAI want to, should it, and can it become the new Google?).
Users now pay subscriptions for services such as ChatGPT Plus, Claude Pro, and Copilot, but contrary to what one might think, this does not remove them from the value capture system.
Every interaction with an AI model becomes a contribution; every question asked, every correction made, every reformulation, every iteration of a prompt feeds the AI. The user is not just a customer, they are also the one who produces the raw material, and it is their daily cognitive work that fuels the continuous improvement of the product.
We have entered a hybrid model where the user finances access to the service while simultaneously enriching its quality: not only do they pay, but they also train it.
The AI economy: a self-reinforcing learning loop
This is the economic model of AI: each interaction not only satisfies a specific user need, but also provides additional data that improves the system.
The more users there are, the more the model learns. The more it learns, the better it performs. The better it performs, the more new users it attracts. This is the self-reinforcing cycle of generative AI, often referred to as the data flywheel.
This mechanism is all the more powerful when the collected data is rich. Unlike the clicks and likes of Web 2.0, the prompts and outputs of an LLM capture intentions, reasoning, business context, and specialized language. We no longer collect passive behavioral traces, but directly absorb human cognition.
It is the user who, often without realizing it, injects fragments of their thoughts into the machine every day, which is also why these models are progressing so quickly.
A dynamic of market concentration
This logic is likely to automatically produce much greater market concentration than that observed in Web 2.0, as several effects come into play and reinforce each other.
First, the data effect: the larger a player’s user base, the faster the quality of its model improves. It then becomes increasingly difficult for a new entrant to catch up in terms of accumulated experience.
Next, there is the “compute” effect: training these models requires massive computing resources, which are reserved for a few businesses capable of investing billions in specialized infrastructure. Microsoft, Google, Amazon, and Meta are locking down access to the computing power that is essential for AI learning.
Finally, there is the lock-in effect: each use creates a history, preferences, fine-tuning, and customizations that make it increasingly complex for the user to change models. The exit cost, although invisible, is very real and increases over time.
The result is an ecosystem where dominant positions are not simply due to current market share but to an accumulation of learning that is virtually irreversible. It is no longer just a race for audience share but a race for intelligence.
But beware: just because de facto quasi-monopolies may emerge does not mean that the issue of the sector’s profitability is resolved (Generative AI: a bubble, a crash, or a turning point?).
Outdated regulation
But the current rules, inherited from the traditional digital economy, are not designed to regulate this new model.
The GDPR and antitrust laws focus primarily on two things: the protection of personal data (such as user identity or preferences) and the monitoring of dominant positions in customer relations (such as distribution or advertising). However, generative AI operates on a different level: it does not simply collect data that could potentially identify individuals, but absorbs the very content of exchanges with users and their intentions. Each interaction provides reasoning, knowledge, and ideas, which are reused to refine models, resulting in a much more subtle and invisible exploitation of knowledge and information.
The real capture of value no longer depends on the volume of transactions, but on the gradual accumulation of these millions of micro-learnings.
The problem is similar when it comes to competition rules. Traditional antitrust tools still measure market share, whereas with AI, it is not necessary to dominate a market to create a huge advantage: all you need to do is accumulate a qualitative lead. This lead, fueled by everyday use, eventually creates a technological gap that becomes very difficult, if not impossible, to close.
A more profound disruption than SaaS
One common mistake is to apply economic reasoning from traditional SaaS to AI, even though the underlying dynamics are very different.
In traditional SaaS, users purchase software with a largely fixed functional scope. Its use does not directly improve the product, except through feedback gathered over time.
In generative AI, on the contrary, every daily interaction contributes to training the model. Each user is integrated into the R&D loop, and improvement is continuous and distributed. The scale of competitive advantage becomes exponential with use, rather than simply proportional to the number of customers.
It is this integration of the user into value creation that produces much more powerful lock-in effects than those of traditional SaaS, and changing AI models will become, over time, much more complex than changing CRM.
Bottom line
The generative AI economy establishes a new implicit contract between users and platform operators. Customers are no longer content to finance the product through their subscription, they also become a driving force that continuously fuels the model’s performance.
This dynamic is shifting the boundaries and blurring the traditional markers of regulation, heralding a level of market concentration unprecedented in the recent history of digital technology.
But, in short, even when we pay, we remain integrated into the business model and now even into product development. We have all become, in a sense, premium test subjects, paying and willing participants.
Related articles:
AI heading for an economic dead end?
Wrappers, deeptechs, and generative AI: a profitable but fragile house of cards
Generative AI: a bubble, a crash, or a turning point?
OpenAI acquires iO for $6.5 billion: strategic pivot or headlong rush?
Is generative AI doomed to never be profitable?
Does OpenAI want to, should it, and can it become the new Google?
