The landscape of players (June 2026)

Table 7.1. Leading frontier laboratories as of mid-2026 (non-exhaustive list).

Three dynamics are worth highlighting. The United States concentrates the best-funded players and the widest diversity of strategies, including a wave of "neo-labs" born of high-profile departures (see the gallery of portraits). China is advancing at a remarkable pace despite American restrictions on access to the most advanced chips (Chapter 8), betting heavily on openness and cost efficiency. Europe, long on the sidelines, sees in Mistral its standard-bearer for a "sovereign AI," supported by public authorities and, notably, by the Dutch semiconductor manufacturer ASML, which became its principal shareholder at the end of 2025.

The DeepSeek moment, in early 2025, made a strong impression: an open model from China demonstrated that a near-frontier level could be reached for a fraction of the cost, upending the entire economics of the sector.

We detailed the benchmarks in Chapter 4 (MMLU, GPQA, SWE-bench, ARC-AGI, human-preference arenas), along with their limits (saturation, contamination, Goodhart effect). Let us recall their central lesson: as of mid-2026, no model dominates across the board, and advanced organizations practice routing (assigning each task to the best-suited model). Any ranking is only a snapshot, valid for a few weeks.

This layered architecture sheds light on what follows: the infrastructure layer is the subject of Chapter 8 (chips, data centers) and Chapter 9 (local AI). The application layer deploys across every sector (health, work, law, addressed in Part V).

This layer accompanies a new business model: selling not a piece of software, but an autonomous agent presented as a "digital worker" dedicated to a function (sales prospecting, customer support, accounting, legal). French-speaking platforms like Limova or Agentova illustrate this, alongside numerous sales agents (AI SDRs) internationally. Billing then shifts from per-user subscription toward usage (per token), then toward outcome (per task completed). The promise of "full autonomy," however, often remains oversold: without human supervision, these agents quickly drift, which loops back to the previous point.

This ferment can be read in several phenomena. First, a global entrepreneurial explosion: nearly 60 percent of the startups in the American accelerator Y Combinator are now focused on AI, cohort after cohort, and the most recent ones no longer sell features but coworkers, going so far as to rebuild entire enterprise software, or even consulting or law firms, in the form of agents. The phenomenon is truly planetary. Its center of gravity remains the San Francisco Bay Area, but major hubs exist everywhere: China (Beijing, Shenzhen, Hangzhou), London, Tel Aviv, Bangalore, Singapore or the Gulf states (Section 7.1), not to mention emerging ecosystems in Africa (Lagos, Nairobi) and Latin America (São Paulo). France is just one example among others, with venues like Station F in Paris (the world's largest startup campus) or EuraTechnologies in Lille. Nor are accelerators limited to Y Combinator: Techstars, Entrepreneur First or Antler operate across several continents.

The projects themselves are highly diverse: there are not only agents sold as SaaS, but also vertical applications (health, law, finance), developer tools, research labs and a host of open-source projects. This is the second facet, the culture of "building in public": developers publish their code as open source (GitHub stars and models deposited on Hugging Face serving as reputation currency) and show off their prototypes through countless demos on X and Reddit. Examples abound in every category: engines for running models yourself, locally (llama.cpp, Ollama, already mentioned in Chapter 9) or on a high-throughput server (vLLM); frameworks for building agents and retrieval-augmented search (LangChain, LlamaIndex); or visual creation interfaces for image and video (ComfyUI). On the products and companies side, the sweep is just as global: image generation with Germany's Black Forest Labs (FLUX models), synthetic video and avatars with Britain's Synthesia, translation with Germany's DeepL, medical research with France's Owkin, law with Legora, customer relations with America's Sierra, defense with Germany's Helsing, or enterprise sovereignty with Germany's Aleph Alpha (merged in 2026 with Canada's Cohere). One detailed example: rtk, an open-source utility (a binary in the Rust language) that cuts the token consumption of AI coding tools by 60 to 90 percent, by filtering command output before it reaches the model; carried by a handful of contributors and boasting tens of thousands of GitHub stars, it shows how a useful project can be born from a simple public repository.

The flip side warrants a word of caution. This abundance mixes the solid and the hollow: alongside those who deliver concrete, measurable products, many are merely stirring air, rebranding themselves "AI" without changing anything at the core, or promising an autonomy they fail to deliver. Knowing how to tell signal from noise (a product that works, customers and verifiable figures, as opposed to mere marketing packaging) has become a skill in its own right. As the observers themselves note, a market whose main customer becomes another piece of software is as exhilarating as it is fragile.

The builders (founders and executives)

Sam Altman leads OpenAI, which he co-founded in 2015. A flagship figure of the "AGI" ambition, he has publicly stated that he aims for a "superintelligence" and a "glorious future." His path was marked by a spectacular governance crisis in November 2023 (his ouster and then return within a few days). He embodies the strategy of rapidly deploying ever more capable models.

Dario Amodei and Daniela Amodei co-founded Anthropic in 2021, after leaving OpenAI, with the idea of placing safety at the heart of development (hence the "helpful, harmless, honest" framework and the so-called "constitutional AI" method). Dario Amodei is also the author of a much-noted essay, Machines of Loving Grace (2024), describing the potential benefits of a powerful but controlled AI.

Demis Hassabis leads Google DeepMind. A former chess prodigy and neuroscientist by training, he co-received the 2024 Nobel Prize in Chemistry for the work on AlphaFold (protein folding, Chapter 14). He champions a path to AGI by way of solving major scientific problems, and sees world models as an essential step.

Elon Musk founded xAI (the Grok model, the Colossus supercomputer) and runs Tesla and SpaceX. A co-founder of OpenAI in 2015, he has since distanced himself from it. His position is singular: he has long warned publicly about the existential risks of AI, while himself running a frantic race to develop it.

Arthur Mensch is the co-founder and chief executive of Mistral (France), which he launched in 2023 with Guillaume Lample and Timothée Lacroix, after a stint at DeepMind. He has become the face of European "sovereign AI," advocating a partly open approach and the continent's technological independence.

Jensen Huang leads NVIDIA, the maker of the chips that train and run nearly all the major models (Chapter 8). The "arms dealer" of the AI rush, he popularized the idea that tokens are "the language and the currency" of this new economy.

Mark Zuckerberg leads Meta, which has bet on open weights with the Llama family and is investing massively in a "superintelligence" laboratory, to the point of triggering a genuine war for talent in 2025-2026 (top-dollar hires of researchers from rival labs).

Liang Wenfeng founded DeepSeek, backed by his quantitative hedge fund High-Flyer. He has become the champion of open and low-cost AI on the Chinese side, whose R1 reasoning model marked a turning point in early 2025.

The scientific pioneers (the "godfathers")

Geoffrey Hinton, one of the "godfathers" of deep learning (Chapter 1), co-received the 2024 Nobel Prize in Physics. In a striking move, he left Google in 2023 in order to warn freely about the dangers of an AI that would surpass humans, becoming one of the great voices of caution.

Yann LeCun, French, a Turing Award laureate and long the chief AI scientist at Meta, is a pioneer of convolutional networks. Skeptical both of catastrophism and of the idea that LLMs alone would lead to intelligence, he champions world models (Chapter 5); he left Meta at the end of 2025 to found AMI Labs in Paris.

Yoshua Bengio, Canadian, also a Turing Award laureate, has established himself as a major voice for safety, notably chairing an international report on AI risks. He argues for great caution and global coordination.

Ilya Sutskever, co-founder and former chief scientist of OpenAI (and co-author of AlexNet, Chapter 2), founded Safe Superintelligence (SSI) in 2024, valued at around 32 billion dollars with no public product whatsoever, with a single goal: to build a safe superintelligence. At the end of 2025 he theorized the shift from "the age of scaling" to "the age of research."

Fei-Fei Li, a Stanford professor and "godmother" of AI for having created the ImageNet image database (Chapter 2), founded World Labs and carries the concept of spatial intelligence (Chapter 5).

The thinkers of safety and ethics

Nick Bostrom, a philosopher at Oxford, popularized with his book Superintelligence (2014) the control problem and the "paperclip maximizer" thought experiment, at the heart of Chapter 24.

Stuart Russell, a professor at Berkeley and co-author of the reference textbook in AI, argues in Human Compatible (2019) for a "provably beneficial" AI, designed from the outset to remain under human control.

Eliezer Yudkowsky, a self-taught researcher and co-founder of the Machine Intelligence Research Institute (MIRI), is the most radical figure of the existential-risk camp. A pioneer of thinking on alignment (Chapter 24) and a leading figure of the so-called "rationalist" community, he has for years defended a bleak thesis: with current methods, building a superintelligence would almost certainly lead to catastrophe. He has carried this position so far as to call for a global halt to frontier AI development, and summed it up in 2025 in a book with an unambiguous title, co-written with Nate Soares: If Anyone Builds It, Everyone Dies. His detractors judge his scenarios speculative and unverifiable; his supporters see in them a salutary warning. He embodies the pole that his opponents call, often to mock it, that of the "doomers."

Emily Bender and Timnit Gebru embody another facet of caution, centered not on distant risks but on the present harms of AI (bias, data exploitation, environmental costs, concentration of power); we owe them the phrase "stochastic parrots" (Chapter 4). Their presence is a reminder that the debate on risks does not reduce to the question of superintelligence.

The heralds of acceleration

In contrast to the voices of caution, a current assumes an accelerationist position: the main danger would not be AI, but slowing it down.

Guillaume Verdon, a physicist and quantum computing researcher (formerly of Google, founder of the hardware startup Extropic), is, under the pseudonym Beff Jezos, at the origin of the effective accelerationism movement (in French accélérationnisme efficace, or e/acc), which emerged in 2022. His credo, formulated in a provocative register and a vocabulary borrowed from thermodynamics, comes down to a few ideas: technological progress would be a quasi-cosmic process that must be accelerated without hindrance, the market would be its best engine, and any attempt to slow it down (regulation, the precautionary principle) would entail more risk than it would avoid. His identity, long anonymous, was revealed by the press at the end of 2023 (Chapter 24).

Marc Andreessen, co-founder of the Netscape browser turned a leading Silicon Valley investor (the Andreessen Horowitz fund), is its most influential mouthpiece. In his Techno-Optimist Manifesto (2023), he ranks "existential risk" among the "enemies" of progress and asserts that AI will not destroy the world but could save it. These widely circulated positions are also fiercely contested: their critics see in them the defense of economic interests as much as a philosophy.

The voices of public debate

Beyond the builders and the scientists, essayists and intellectuals shape the way societies receive AI. The French debate offers contrasting figures, whose theses, often provocative, are themselves disputed.

Luc Julia, a Franco-American engineer associated with the creation of the Siri voice assistant (an authorship he himself qualifies, calling himself rather the "grandfather of Siri"), former head of innovation at Samsung then chief scientist at Renault until 2026, and a member of the Académie des technologies, is the author of L'intelligence artificielle n'existe pas (2019). He defends a deflationist thesis: the term "artificial intelligence" would be a misnomer; one should speak of "augmented intelligence," because the machine would merely assist humans, without consciousness or intelligence of its own, and a general AI "capable of doing everything" would not exist. Much publicized, his arguments are also contested by other specialists, who point out technical inaccuracies in them.

Laurent Alexandre, a surgeon by training, founder of the Doctissimo website and an essayist, has long defended a transhumanist and alarmist vision: in La Guerre des intelligences (2017), he presents AI as an inevitable upheaval requiring a rethinking of education and the "augmentation" of humans, with pointed warnings about employment and inequalities of "cognitive capital." His provocative formulations elicit as much echo as criticism.

Olivier Babeau, an economist and founder of the Institut Sapiens think tank, co-signed with Laurent Alexandre Ne faites plus d'études (2025), an essay with a deliberately provocative title: with intelligence becoming "free" thanks to AI, classical studies would lose their value, and one would have to learn differently, in a more demanding and permanent way (a debate developed in Chapter 15). The book is careful to insist it is not an indictment of knowledge.

These voices, optimistic or worried, deflationist or catastrophist, are a reminder that AI is also an object of societal debate, where the most clear-cut positions must be read as such: stances taken, not established truths.

---

Key takeaways (Chapter 7)

• The state of the art is spread across three hubs: the United States (the best-funded and the most diverse), China (fast, open, cost-efficient despite chip restrictions) and Europe (Mistral, "sovereign AI").
• The great strategic divide pits proprietary models (closed, sold via API: OpenAI, Anthropic, Google) against open-weight models (Llama, DeepSeek, Qwen, Mistral). Openness democratizes but complicates control.
• The ecosystem is tiered into layers: hardware, foundation models, applications, no-code tools.
• A services layer (freelancers, agencies, integrators) helps companies deploy and supervise AI; in parallel emerges a "digital worker" model (an autonomous agent billed by usage or by outcome), driven by a wave of "agentic" startups.
• The people behind AI are spread across a spectrum ranging from builders in a hurry to voices of caution, with no sharp boundary: Altman, the Amodeis, Hassabis, Musk, Mensch, Huang, Zuckerberg, Liang on one side; Hinton, LeCun, Bengio, Sutskever, Li, Bostrom, Russell, Yudkowsky, Bender and Gebru bringing, each in their own way, science and vigilance, while the heralds of acceleration (Verdon, alias Beff Jezos, and Andreessen) defend the opposite thesis.

Thus ends Part II. We have understood what AI is, how it works and who makes it. Part III descends into the engine room: the hardware, the infrastructure and the energy that make all of this possible.

---