In this new episode of MuzzOsphère, Antoine Gigomas welcomes Ludovic Peran, formerly at Google DeepMind and now in Technical Strategy at OpenAI.
On MuzzOsphère, we give the floor to the people who build growth: inside organisations, in structural decisions, often far from the spotlight. Today, we turn to a subject that is profoundly transforming value creation: artificial intelligence. Beyond the theory, this conversation explores very concrete questions: how to make the right decisions, structure your organisation and scale effectively in a world increasingly driven by AI.
what OpenAI's growth looks like from the inside
Antoine Gigomas opens with a direct question: what is actually happening at OpenAI today? Ludovic Peran's answer comes down to one word: effervescence. The company has published figures it rarely shares. As Ludovic Peran points out in the episode, OpenAI generated $20 billion in revenue in 2025, up from $6 billion in 2024, with the acceleration driven mainly by B2B. One million businesses now use its services, and he cites 19% week-on-week growth in its API revenue.
What strikes the guest is the ratio between that impact and the size of the organisation. By his account, OpenAI has around 3,000 employees, fewer than the legal department at Google, his former employer. The internal engine of the moment: agentification, with agents like Codex deployed at scale and what he describes as very strong traction with businesses.
structuring an organisation in hypergrowth: “well-organised chaos”
Asked what hypergrowth really looks like (extreme clarity or disorder?), Ludovic Peran settles on “well-organised chaos”. Chaos, because teams have wide latitude and the “everybody is a leader” principle pushes everyone to spot opportunities and seize them, without waiting for a manager's approval. Organised, because leadership sets clear directions, repeated regularly, and small, agile teams move forward on constant feedback from their peers.
This freedom has one non-negotiable limit: safety. Ludovic Peran, who led work on these issues, insists on one point: nothing leaves OpenAI without going through a rigorous validation process, a series of review committees that goes all the way up to the board. Everything else can change overnight; this anchor point does not move. He describes a “preparedness framework”, the document that details the full set of tests (capabilities in biology, biochemistry and cyber) that every model goes through before release.
the “capability overhang”: the real barrier to AI adoption in business
For Ludovic Peran, the most consequential decision in the face of AI is not technological: it is a matter of governance and leadership. He describes what he calls the “capability overhang”, the growing gap between what models can do and the way the average user actually uses them. He illustrates this gap with one figure: the most advanced users draw on seven times more reasoning capacity than average.
Even if we froze the models today, it would still take a few years before many business users were using them to their full capacity.
Yet the models are far from frozen: they improve every month, and the gap keeps widening. Antoine Gigomas illustrates a way forward with a concrete example: at Muzzo, ChatGPT checks every application that goes through the platform against the job description and the client's brief, flagging for the recruiter any gaps with what the employer asked for. Ludovic Peran sees it as a good example of AI raising the quality of a service, and suggests going even further on high-value tasks.
attracting the best people and scaling in an AI-driven world
OpenAI does not fit any of the classic boxes (product-led, sales-led). Ludovic Peran coins a category of his own: “research-led”. Fundamental research is the core of the engine; without it, he says, there would be no OpenAI. It is a hard model to replicate, because it requires attracting rare profiles (the guest even turns the hiring question back to his host, who runs Muzzo). His message on attractiveness is clear: pay and computing power are necessary conditions, never sufficient ones. What tips the balance is a clear mission and vision, and the scale of the impact (products used, by his account, by 800 million users a week).
One anecdote sums up how he relates to these environments. Ludovic Peran quotes one of his former managers:
When you are offered a seat in a rocket ship, you do not ask which seat.
He adds a word of caution for anyone dreaming of copying that trajectory: OpenAI is 10 years old, ChatGPT only 3; fundamental research is a marathon. His advice for deciding and scaling under uncertainty comes down to two things: agentification, and pushing your use of AI to the max, never superficially. In practice, that means giving models long, complex tasks that create value, rather than just automating what already exists: that is how you join the small group of companies “at the frontier” of their industry.
the guest: Ludovic Peran
Ludovic Peran has built his career around a single question: the impact of cutting-edge technology on society. He started out in politics, working for a member of parliament who was rapporteur of the French Digital Republic Act, before joining Google (first in policy, then in product management at Google Health in Palo Alto). He then worked on AI at Google Brain, in Responsible AI, and later on open-model safety at Google DeepMind. At OpenAI for a little over a year, he was safety lead for GPT-OSS, led the launch of Safeguard (the company's first open safety model) and now works, back in France, on the principles applied to products, within the Technical Strategy team.
key takeaways
- OpenAI generated $20 billion in revenue in 2025 (up from $6 billion in 2024) with around 3,000 employees: a rare impact-to-size ratio, according to Ludovic Peran.
- The internal model is “well-organised chaos”: everyone takes initiative under clear direction, with safety as the only non-negotiable anchor point.
- The real barrier to AI is not technological: it is the “capability overhang”, the gap between what models can do and how companies actually use them.
- To attract the best people, mission matters more than pay; building a research lab is still a marathon (OpenAI is 10 years old).
- The advice for scaling: agentification and advanced use of AI on long, complex tasks, not just automation.
