The generative AI value chain: specialized players. (3)

The developers of generative AI models. – Alongside the major historical players of the digital economy—Google, Microsoft, Amazon, Meta, Nvidia, Apple—on whom they partly depend from an infrastructural, financial, and commercial standpoint—inasmuch as the latter control critical technologies (infrastructure and compute power), provide capital, and act as distribution channels—the generative AI sector includes a set of specialized developers whose core activity consists in designing, training, and continually improving generative AI models. These actors occupy a pivotal position in the value chain, insofar as they are the source of the foundation models on which a large share of generative AI uses and services rely. Hence a few conceptual and semantic clarifications that are important to keep in mind in order to understand precisely their role in the generative AI value chain.

A group of researchers affiliated with Stanford University defines a foundation model as: “(...) any model that is trained on broad data (generally using self-supervision at scale) and that can be adapted (e.g., fine-tuned) to a wide range of downstream tasks; current examples include BERT.” (Source: On the Opportunities and Risks of Foundation Models, published on arXiv in August 2022). A foundation model is therefore a type of AI model—which means that not all AI models are foundation models—defined by three main characteristics:

• Volume of heterogeneous data: They are trained on massive volumes of heterogeneous data (text, images, videos, sounds, computer code) drawn, generally, from a very significant portion of what is accessible on the Web. The term “heterogeneous” does not refer only to the multimodal nature of the data on which models are trained, but also to different cultural contexts, different people (teachers, researchers, political activists, journalists, writers, artists, the general public, conspiracy theorists, etc.), different languages and styles, different time periods, and also different intentions, such as informing, explaining, entertaining, teaching, deceiving, manipulating, persuading, and so on.

• Self-supervised learning: They are trained using an approach called “self-supervised learning,” as opposed to “supervised” learning. In other words, they learn by themselves, without any explicit rule being provided to them in advance. It is important to note that foundation models have no semantic understanding of the linguistic and non-linguistic symbols they manipulate; rather, during training they build a mathematical and statistical representation of how the multimodal content they have ingested functions.

• Adaptable to a wide variety of tasks: Finally, they are able to perform a very broad range of tasks that go beyond the strict scope of the AI subfield known as Natural Language Processing (NLP), since foundation models can be used in applications capable of processing not only text but also images, video, sound, and computer code.

It is important to note here that when one speaks of an LLM (Large Language Model), one is referring only to a subcategory of foundation models centered on language processing and perfectly suited to tasks such as text generation, commentary or explanation, multilingual translation, or answering questions formulated in natural language. One should not confuse Large Language Models (LLMs) with the more general category of foundation models, such as GPT (OpenAI), Claude (Anthropic), and Gemini (Google), which can be trained and fine-tuned—hence the term “foundation” used here—to generate not only text, but also images, video, code, and sound, depending on the types of content on which they were trained and the tasks for which they were specialized. In other words, a foundation model is defined above all by its versatility and potentially multi-application character, whereas an LLM is specialized. In both cases, they belong to the class of algorithmic models known as generative AI, in the sense that they can create original multimodal content.

Foundation models do not merely introduce a “paradigm shift” in the construction of AI systems, to use the very terms of the document cited above; they also bring forth new actors—OpenAI, Anthropic, Mistral AI, Cohere, Hugging Face, Aleph Alpha, Perplexity, etc.—who are contributing to a deep and almost unprecedented reconfiguration of the global digital environment (software, hardware, devices, interfaces, platforms, infrastructure, digital services, information systems, and so on). This reconfiguration is expressed first at the application layer, through the growing integration of generative AI models into existing, historically established applications—office suites, graphic creation software, customer relationship management (CRM) solutions, robotic process automation (RPA), enterprise resource planning (ERP) systems, search engines, software development platforms, etc.—but also through the emergence of new actors offering applications built natively on large language models (LLMs).

These new entrants in the tech sector generally take the form of AI research labs bringing together very high-level scientific skills, drawn both from academic research (Stanford, MIT, Berkeley, Harvard, Columbia, etc.) and from leading engineering schools (École Polytechnique, CentraleSupélec, Télécom Paris, etc.), as well as from private research (Google Brain, DeepMind, FAIR at Meta, Microsoft, etc.). Their main specificity lies in their ability to connect fundamental research and industrial development, to mobilize significant financial resources, highly qualified teams and substantial compute capacity, and to collect and process the massive volumes of data required to train models. These actors differ according to three main ways of making models available and distributing them, which reflect both different development philosophies and different underlying economic logics:

• Open source models: First, so-called “open source” models, whose code, parameters (weights), and training data are publicly accessible to the developer community. Developers can freely examine the source code—model architecture, layers, training logic, and so on—modify it if necessary, retrain the model on new data, and fine-tune it in order to make it perform specific tasks without having to start from scratch. Inspired by the free software tradition initiated under Richard Stallman and the Free Software Foundation (1985), the open source philosophy promotes an ideal of transparency, knowledge sharing, and collaboration within the scientific community, as well as the democratization of use for the benefit of the greatest number. The Open Source Initiative (OSI), a California-based non-profit founded in 1998, defines open source AI by drawing on the four freedoms associated with free software, namely: (1) the freedom to use the system for any purpose without having to request permission; (2) the freedom to study how it works and inspect its components; (3) the freedom to modify it for any purpose, including modifying its output; (4) the freedom to share it so others can use it with or without modifications, for any purpose. Today, French companies such as Hugging Face and Mistral AI offer foundation models that comply with the four freedoms set out by the OSI.

• Closed source models: Second, so-called “closed source” models, or proprietary models, for which neither the source code, nor the weights, nor the training data are made public. The company or organization publishing the model thus aims to fully control the model’s design, development, distribution, evolution, and monetization for commercial purposes, in a logic intended to build customer lock-in by offering top-tier performance. Actors such as OpenAI, Anthropic, and Google, with GPT, Claude, and Gemini respectively, favor the “closed source” model.

• Open-weights models: Finally, so-called “open weights” models, which sit in a sense at the intersection of open source and closed models. Here, only the model’s final parameters (weights)—hence the expression “open weights”—are made public. Making only the weights available, rather than the source code or the data used to train the model, allows researchers, developers, and companies to download the model, adjust its weights by fine-tuning on other data—potentially proprietary data, for example—then run it for specific tasks, or integrate it into another application such as a CMS (Content Management System) or a CRM (Customer Relationship Management). While this hybrid approach gives developers some flexibility, since they can adapt the model to business and functional constraints, it is the publishing company that controls the model’s overall evolution—i.e., updates, improvements, and roadmap—typically invoking security and intellectual property reasons. The actor that most prominently favors this approach today is Meta, with its foundation model Llama.

The monetization of products and services distributed by foundation model publishers is aligned with the business models practiced by most SaaS (Software as a Service) platform vendors for more than two decades. In other words, they rely on well-established approaches, widely proven and finely tuned, used by the major software industry players for over twenty years. Without claiming exhaustiveness, and without accounting for each actor’s specificities—which depend heavily on the market segments they prioritize and the positioning they adopt—one can identify six major pricing strategies, the first two being the most dominant today:

• SaaS subscription: The first monetization channel, and the most visible, consists in selling—via a monthly subscription—access to a conversational agent—ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Le Chat (Mistral AI), Copilot (Microsoft), Grok (SpaceX), Perplexity, etc.—through an out-of-the-box SaaS application that can be customized only to a limited extent. This type of offer, aimed primarily at individuals, can also be deployed within a company via license agreements and made available to employees to improve productivity or enable self-directed learning, for example.

• API and usage-based billing: The second channel, aimed primarily at companies and developers who want to build ad hoc applications or integrate the capabilities of a generative AI model—GPT, Claude, Gemini, Mistral, etc.—into existing applications, consists in charging for API calls each time a request is sent to the model to obtain a response. Pricing can be set per token—i.e., the number of processed units of text—or per request tier depending on volumes. This is also referred to as usage-based billing or pay-as-you-go.

• Enterprise licenses and private integrations: The third major channel consists in licensing contracts tailored to companies operating in sensitive, highly regulated sectors—banks, insurance companies, public administrations, etc.—that do not accept that their data transit through public servers, but that also wish to host, on internal servers, models they have previously customized based on specific use cases and business applications (fine-tuning). Example: a bank that wants to build a sensitive internal document analysis application without any data leaving its own information system.

• Integration into third-party platforms: The most typical case of integration into third-party platforms or applications—the fourth monetization channel—is Microsoft Copilot—which is in reality OpenAI’s GPT—integrated and accessible directly within Microsoft 365 (Word, Excel, PowerPoint, Outlook, Teams), with the aim of automating drafting tasks or producing summaries in Word, data analysis in Excel, or generating PowerPoint presentations from text written in Word, for example. In this context, content generation inside Microsoft’s suite relies on API calls in a way that is fully transparent to the user. Independently of the integration agreement between OpenAI and Microsoft, the revenue model relies primarily on usage-based billing.

• Marketplaces and platform models: A fifth monetization mode is that of marketplaces, for which OpenRouter is the most obvious example. The American company, founded by Alex Atallah and Louis Vichy in 2023, provides an API aggregation platform for AI models and acts as an intermediary between, on the one hand, developers who want access to many different models and, on the other, foundation model providers such as OpenAI, Anthropic, Google, or Mistral, as well as open source actors. In other words, OpenRouter standardizes—via a single hub—technical access to a very large number of models and takes a margin on each API request, which it then routes to the chosen model.

• Advertising and attention monetization: Finally, the sixth and last monetization channel—which is not, as of today, deployed and which some actors, including OpenAI, have begun to explore—is advertising, or what some call sponsored recommendations. This is the well-known model of sponsored links in search engines such as Google or Microsoft Bing, which OpenAI would reportedly have begun testing in the United States for users of the free version of ChatGPT. The aim is not only to fund free access to the application, diversify revenue sources, but also encourage users to subscribe to the paid version. It remains to be determined under what rules an ad would be displayed, in response to what kind of prompt, where in the user interface, under what billing model, and what this implies for the protection of users’ privacy. Some competitors such as Anthropic have openly criticized Altman’s proposals, arguing that it would undermine user trust and divert the primary function of AI assistants, whose role is first and foremost to serve users, not advertisers.

In a context marked by the diversification of business models and of leading development strategies, OpenAI occupies a particular place among developers of generative AI models. Founded in 2015 by Elon Musk, Sam Altman, Greg Brockman and other figures from the tech world, with the primary aim of developing artificial intelligence that benefits humanity, the company has largely contributed to the large-scale democratization of a mass-market conversational agent (chatbot) from November 2022 onward. Initially created as a non-profit organization intended to prevent the formation of monopolies and to guarantee fair access to the technology, the company adopted, from 2019, a hybrid model, with the creation of a capped for-profit entity designed to enable it to raise the funds necessary for its development.

The first GPT model, GPT-1, published in 2018, was mainly disseminated in the form of a research publication. It laid the foundations of an approach developed by OpenAI, based on pre-training the model on very large text corpora, followed by a supervised fine-tuning phase for specific tasks. GPT-1 relied on a Transformer-type architecture, introduced shortly beforehand by a Google research team in a 2017 paper entitled Attention Is All You Need. OpenAI’s teams aimed to demonstrate that a language model built on this specific neural network architecture was capable of understanding the context of a sentence and generating text coherently.

Published and made available to developers and researchers in open source form—code and weights—via the GitHub platform in November 2019, version 2 (GPT-2) largely confirmed the hopes raised by version 1, with a far more powerful model containing nearly 1.5 billion parameters, compared with roughly 100 million for GPT-1. However, it was only with version 3, unveiled in May 2020 and integrating more than 175 billion parameters, that OpenAI paved the way for a first form of commercialization by exposing its API (Application Programming Interface) to developers, who now had the ability to create applications leveraging the model’s exceptional capacity to process text formulated in natural language.

This evolution naturally led, in November 2022, to the launch of an optimized version of GPT-3, called GPT-3.5, which definitively popularized the use of generative AI by enabling any user—without any particular technical skills—to interact with the model through a user-friendly interface accessible to all, either via a simple web browser or through a mobile application. In other words, from a generative AI model usable mainly by developers through API connections, GPT-3 became ChatGPT-3.5, turning a language model into a conversational assistant capable of answering naturally to questions of all kinds, or generating text in more than 80 languages and in varied styles.

Since then, OpenAI has continued to accelerate the release pace of its new models, with, in March 2023, the announcement of GPT-4, integrating multimodal capabilities for combined processing of text, images, video, or audio files; then GPT-4.5, in February 2025, focusing on reducing hallucinations as well as on more fluid and natural interactions; and finally, in August 2025, the release of GPT-5, described as more robust and higher-performing, but above all capable of accomplishing a wide range of very high-level tasks, such as advanced reasoning (logic and understanding of abstract problems), the solving of complex mathematical problems, writing, correcting and optimizing code in several programming languages, as well as multimodal capabilities—introduced with version 4—further strengthened.

In early 2026, the conversational agent would reportedly count around 800 to 900 million weekly active users, which would make it one of the most used digital applications in the world. By way of comparison, only a few social platforms surpass it in audience, such as Facebook (3 billion), YouTube (2.5 billion), Instagram (2 billion), WhatsApp (2 billion) or TikTok (1.6 billion). This user base would have almost doubled in one year, reflecting both massive and rapid adoption. Estimates of the number of paying subscribers vary by source, but converge around 30 to 35 million users, i.e., roughly 3.5% of the weekly active base.

At the end of 2025, OpenAI’s estimated revenue would reach around 13 billion dollars, versus 2.7 billion in 2024, i.e., an almost fivefold increase in the space of one year. Since the company does not publish an official detailed breakdown of these revenues, it remains difficult to assess the respective share coming from the “consumer” segment and the “enterprise” segment. By the 2030 horizon, projections point to around 2.6 billion active users, i.e., a tripling compared with 2026, with 220 million paying subscribers expected, representing around 8.5% of the active user base, versus 3.5% in 2026. According to certain unofficial estimates, projected revenues for fiscal year 2026 could fall between 26 and 29 billion dollars, i.e., roughly double those estimated for 2025. Finally, it should be recalled that OpenAI, not being a publicly listed company, is not subject to the same financial transparency obligations as public companies.

Along with Google, Anthropic is today the most serious competitor to OpenAI. The company was created in January 2021 by Dario Amodai and his sister Daniela, both former OpenAI employees. Their departure was motivated by deep disagreements with the direction taken under Sam Altman’s leadership. Taking shape in the context of the exceptional enthusiasm within the community of engineers and researchers for the very promising performances of the first generative AI models—such as the developments of GPT-2 or GPT-3 to which they contributed—while also being conscious of the associated risks, the founders gave Anthropic the central mission of designing AI models that are reliable and aligned with human values.

The founding vision of the two co-founders, soon joined by other OpenAI researchers and engineers from the creation of the organization, rests on three key principles: (1) designing generative AI systems aimed at minimizing risks of undesirable behaviors, which is expressed in particular through the implementation of guardrails intended to prevent the generation of harmful content, such as instructions relating to the manufacture of dangerous substances; (2) working toward the design and deployment of AI systems that are understandable and controllable, enabling users to understand, at least partially, how an automated decision is produced, within a perspective of explainable AI; and finally, (3) ensuring that the responses produced by models are aligned with widely shared ethical and social norms—according to an ethics by design logic—such as benevolence, respect for individuals’ autonomy, and social justice.

It is within this framework that Anthropic devised a new alignment method called “Constitutional AI,” intended to remedy the limits of the reference method that had until then been dominant, “reinforcement learning from human feedback” (Reinforcement Learning from Human Feedback, or RLHF), and above all to guarantee that increasingly complex and high-performing AI models remain useful, safe, and aligned with principles judged acceptable by humans.

For Anthropic’s researchers, the RLHF method (see my article: L’éthique à l’épreuve de l’IA : les limites de l’éthique des machines) suffers from at least five major flaws (source: Constitutional AI : Harmlessness from AI Feedback): (1) in the trade-off between usefulness and harmlessness (helpfulness and harmlessness), human annotators mechanically tend to privilege the safest answers, at the expense of the model’s real usefulness; in other words, when in doubt, to favor “I can’t answer” type responses rather than risk a potentially harmful answer; (2) the RLHF approach mobilizes thousands, even tens of thousands, of annotators, which introduces disparities of evaluation (cultural differences, fatigue, variability of judgments) while generating high logistical and financial costs; (3) it produces a form of normative drift, since decisions rest more on aggregated individual preferences than on an explicit reference framework of clearly formulated principles; (4) it raises ethical issues linked to working conditions, annotators being exposed to potentially offensive or dangerous content; finally, (5) it struggles to scale in a context where models become ever more powerful and autonomous, which makes it necessary to establish a more systematic, explicit, and homogeneous normative framework.

By contrast, the “Constitutional AI” approach proposed by Anthropic, which constitutes the development and deployment framework for the Claude family of models, aims to integrate into the model’s parameters (weights) a set of behaviors learned during training, which should reflect as well as possible the values, objectives, and ethical framework explicitly described in a foundational document—available on Anthropic’s website—called “Claude’s Constitution.” Just as a constitution, in the legal sense of the term, indicates a set of principles situated at the top of the hierarchy of norms, Claude’s “Constitution” refers to a kind of founding ethical charter intended to guide model training, in order to frame and orient their behaviors in sensitive contexts. The principal author of the document, Amanda Askell (ex OpenAI), PhD in philosophy from New York University and a specialist in moral philosophy, explicitly refers here to the notion of “constitutionalism,” i.e., the idea—at once legal, political, and philosophical—that the best protection against arbitrary power consists in framing its exercise through a fundamental text. The idea of “Constitutional AI” at Anthropic thus transposes into the field of engineering and artificial intelligence concepts hitherto associated with legal, political, and philosophical thought.

If I have taken the trouble to explain in detail this notion of “Constitutional AI,” it is because it lies at the heart of the scientific and entrepreneurial project of Anthropic’s co-founders, notably defended publicly by Dario Amodei—particularly on his personal blog—who regularly insists on the risks associated with AI systems and on the necessity of deploying a development philosophy that guarantees safe, ethical, transparent, and responsible use of AI. This original multidisciplinary approach mobilizes engineering proper, but also ethical philosophy, the social sciences (economics, political science, law, sociology), and technological governance, in order to deliberately move away from an exclusively technicist vision of AI—while nonetheless always linking it to Anthropic’s positioning and development strategy.

Dario Amodai’s company develops a whole series of large language models (LLMs) designed to meet differentiated needs and user profiles—including the Claude Opus series, oriented toward complex problem-solving and agentic uses; Claude Sonnet, particularly well suited to software development and document analysis; and Claude Haiku, favored for fast applications, personal assistance, and content generation—with, as its core target and strategic growth priority, the enterprise world (B2B), and more particularly large public or private organizations, without excluding the “consumer” market (B2C).

According to several financial estimates and projections relayed by the specialized press, the company’s revenue would thus have reached around 9 billion dollars in 2025, i.e., a ninefold increase compared with fiscal year 2024, while Anthropic would be aiming to double its revenues in 2026, with an upwardly revised forecast of around 18 billion dollars. This exceptional growth would be driven mainly by two revenue sources: (1) enterprise adoption of the Claude model family via API subscriptions; (2) the rise of offers dedicated to software development, including the Claude Code product, which would alone have generated more than one billion dollars in revenue according to certain sources.

Internal documents and financial projections cited by several media outlets also mention revenue scenarios that could reach respectively 55 billion, 70 billion, and 148 billion dollars by the 2027, 2028, and 2029 horizons, with positive cash flow envisaged before 2028. The company nevertheless communicates no official forecast regarding a precise profitability target, invoking in particular the scale of infrastructure investments required for training and operating its AI models.

If OpenAI and Anthropic, which one could describe as “pure players,” seem to represent, at present, the two main specialized developers of foundation models—excluding more integrated or diversified historical companies such as Google, Microsoft, Amazon, or Meta—other actors, notably European ones, which can be considered “very serious challengers” given their smaller size, must also be mentioned, first among them Mistral AI. Founded in April 2023 by three French researchers from DeepMind and Meta, Arthur Mensch, Guillaume Lample and Timothée, the company specializes in designing foundation models grouped under the name Mistral, as well as in developing a conversational agent called Le Chat, intended for both professional and mass-market use.

Mistral AI positions itself as a very serious alternative actor, capable of competing, on certain market segments, with the models developed by the major American players, notably OpenAI and Anthropic. Unlike these two competitors, the initial intention of the three founders, who met during their studies at École Polytechnique, is to build high-performing language models accessible to as many people as possible by favoring a so-called “open source” approach, intended to promote and foster collaborative innovation. The aim is also to bring European actors onto the scene in order to counterbalance American hegemony and to guarantee, in the long run, Europe’s technological independence (the issue of technological sovereignty). Mistral AI’s development plan is mainly centered on companies, public institutions, and the developer community, with an openly stated desire, from the outset, to expand internationally. In early 2026, the company would have reached an estimated annual recurring revenue (ARR) of 400–450 billion dollars, with a revenue objective close to one billion dollars by the 2026 horizon according to certain projections.

Like Mistral AI, which claims an open approach as a credible alternative to closed models, Hugging Face also occupies a singular place in the generative AI value chain. Founded in 2016 by three French entrepreneurs—Clément Delangue, Julien Chaumond and Thomas Wolf—the Franco-American company provides a hosting and distribution platform for open source or open-weights models—the Hugging Face Hub—for the community of researchers, developers, engineers, academics, startups and companies wishing to publish, create, fine-tune, share and deploy generative AI models. To date, Hugging Face offers more than two million open source or open-weights models, making it by far the world’s largest public repository of artificial intelligence models, freely accessible, downloadable and reusable. Its business model does not rely on selling licenses to use proprietary models (closed models), but on subscriptions and infrastructure services, as well as support and professional services (Professional Services) enabling companies to develop, deploy and maintain applications based on open source or open-weights models. Between 2021 and 2022, the company also initiated the BigScience project, bringing together more than one thousand researchers from more than sixty countries. This led, in July 2022, to the launch of a 176-billion-parameter model distributed under an open source license, called BLOOM, trained on the Jean Zay supercomputer located in Saclay, near Paris.

Finally, beyond the most mediatized actors already mentioned, there exists a multitude of other companies that contribute, each in its own way, to the dynamics of a rapidly expanding sector. Some, situated upstream in the value chain alongside OpenAI, Anthropic or Mistral AI, develop their own generative AI models, such as the Canadian company Cohere (development of models specialized in Natural Language Processing—NLP), the German company Aleph Alpha (development of sovereign language models), the British company Stability AI (publisher of Stable Diffusion for photorealistic image generation), or the American company Midjourney (artistic image generation). Others, positioned more at the application layer, offer multimodal tools accessible through mass-market web interfaces, such as Runway (audiovisual creation), HeyGen (video generation with digital avatars), or ElevenLabs (AI voice synthesis). Finally, a new generation of companies focuses on task automation and workflow orchestration through approaches based on AI agents, such as Dust, Make, or n8n. These are symptomatic of a movement whereby the market and value progressively shift from simple automated content generation toward the automation of automatically coordinated tasks, which is at the same time a movement of expansion of automated work over human work.

This rapidly changing ecosystem thus illustrates the emergence of a wave of innovations combining multimodal models, conversational interfaces and new software engineering practices, often designated by terms announcing a new phase in the development of generative AI, such as agentic AI—an approach combining language models and APIs in order to build systems capable of acting autonomously or semi-autonomously—or vibe coding, a new practice of developing software applications from instructions formulated in natural language that a generative AI translates into computer code.