Masteria, cabinet IA à Lyon : audit, conseil, développement et formation
Certifié QualiopiFinançable OPCOEurope · États-Unis · Inde
Français
Wednesday 18 November 2026 edition

AI Watch, Wednesday 18 November 2026AI's own champions doubt its revenue

Masteria editorial team, led by Mathias Nizan · , 9:25

Nvidia doubles revenue to $115.8 billion and rules out Jensen Huang's return, yet on the same day Baidu, Xiaomi and the IMF warn that AI revenue is lagging behind the investment.

  • OVHcloud reveals its first French model, OpenAI launches GPT-6.2, and Microsoft wants Copilot to be the workplace where humans and agents work together.
13 stories13 sources7 min read
Top story
Cited in the analysis

Nvidia beats expectations at $115.8 billion and shuts the door on Jensen Huang

For its first quarterly results since its co-founder's departure on 12 November, the chipmaker doubled its revenue year on year, above the $108.9 billion Wall Street expected, and raised its guidance. A few hours earlier, Jensen Huang had said he would be "honoured" to return if the board asked him.

The answer came during the results presentation: "The board accepted his resignation unanimously and has no intention of reinstating him," said chairman Harvey Jones. Interim chief executive Colette Kress announced that Nvidia is halting chip sales to China, despite Washington's green light, and confirmed the acquisition of Hugging Face.

In the background, the SEC has reportedly opened an investigation, and the group has set up a national security and export control committee. SCMP, Bloomberg,

Europe and France

OVHcloud reveals Olala, its first AI model

The French hosting company has unveiled its first family of large language models, named Olala. The first in the line, Olala mini 7A1B, is a compact mixture-of-experts model (an architecture that activates only part of its parameters on each request, to cut compute cost), able to run on a consumer GPU.

The model and its weights will be published this week on Hugging Face under an Apache 2.0 licence, so freely reusable, including in business. Octave Klaba, OVHcloud's CEO, is targeting 100 million users. The wager remains to find a place between the American giants and the low-cost Chinese open models.

Le Monde,

SourceNext.ink
Cited in the analysis

59% of French employers think AI will have little or no effect within five years

A study by the Observatoire de France Travail (the research arm of France's public employment service), which surveyed 16,000 establishments, measures a clear gap between the announced shock of AI on jobs and how employers see it. Its use remains a minority practice, and on average its automation capacity is not judged decisive for now. Wide differences exist across sectors.

The figure rounds out recent studies that already measure a drop in hiring in the most automatable jobs.

Cited in the analysis

An AI audit is judged on the work, not on the technology

The Journal du Net publishes a five-question grid for a small business before buying a tool or launching a development project: where does AI pay off, on which tasks, at what cost, with what data, and how to measure the gain. The through-line: a useful audit starts from the real work and its frustrations, not from the available technology.

International

OpenAI launches GPT-6.2 Astra and Sol, better and cheaper

The lab is rolling out two new models in ChatGPT, Codex and its API. Sol targets everyday work, with the best speed/cost balance for most uses; Astra is the frontier model, which stands out as soon as you hand it an entire case file.

OpenAI points to higher-quality professional outputs, stronger code and more reliable computer control, all at an inference cost (the price of running a model once it is trained) below the previous generation. The falling price per token confirms a deep trend that weighs on 2027 budgets. Next.ink,

SourceClubic

Microsoft wants to make Copilot "the operating system of work."

At its Ignite 2026 conference, Microsoft presents Copilot as a space where employees and AI agents work together, linked to the company's data and software.

The group adds tools to secure and, above all, measure the cost of agents, new Azure virtual machines, and seven cheaper in-house models, from code to voice via cybersecurity and health, presented as doing almost as well as OpenAI or Anthropic for a fraction of the price. Microsoft also announces Copilotforce with Salesforce, to inject customer context into its tools.

The battle is being fought over where the working day is spent, and over price. ZDNet,

SourceFrandroid

Isomorphic Labs raises $5 billion to move into clinical trials

The Google DeepMind spin-off born from AlphaFold, the system that predicts protein structure, has closed a $5 billion Series C round. The money is meant to fund moving its AI-designed drug candidates to human trials. The amount is among the largest ever raised by a drug-discovery start-up.

AI applied to biology enters its expensive phase, that of clinical trials and the regulatory calendar.

China and Asia

Cited in the analysis

The heads of Baidu and Xiaomi warn about the gap between AI's promises and its revenue

At the APEC CEO summit in Shenzhen on 17 November, Robin Li, founder of Baidu, and Lei Jun, founder of Xiaomi, called on the industry to focus on concrete applications, warning that the massive investments far exceed the commercial returns. The warning comes amid worries about an AI bubble and the day after Jensen Huang's resignation.

Coming from two Chinese heavyweights, this caution echoes the doubts of American markets on the very day of Nvidia's record results.

SourceSCMP

China launches its first "token export" cluster on Huawei chips

The country has brought into service in Shantou, on the southern coast, a first data centre aimed at foreign customers, driven by Huawei's latest Ascend 950DT chips. The business model, called "token export," consists of selling compute power and model responses rather than the chips themselves, which US controls prevent from being exported.

Beijing thus wants to become a global supplier of AI capacity despite the sanctions, relying on its domestic hardware.

SourceSCMP
Research

Research and papers

For technical teams

A language model can rewrite its own weights to improve itself

A study posted on arXiv on 18 November shows that a 27-billion-parameter model can, from its own weights, scored responses and the ability to run itself, diagnose its failures and modify its parameters to correct itself, with no gradient and no new training. The authors call this capacity "self-rewriting."

The result extends work where a frontier model edits the weights of a smaller model, and bears directly on governance: a system that modifies itself complicates traceability and control.

SourcearXiv
The rest of the news

In brief

  • OpenAI has signed a deal with seventeen French media outlets, including Le Parisien and Les Échos, to feature their content in ChatGPT and roll out its tools, including ChatGPT Enterprise, in newsrooms; no amount was disclosed.

  • ChatGPT can now settle purchases with your bank card, while Google extends Gemini Tasks, able to make calls and buy things on your behalf: agents gain a spending power that shifts the question from trust to control. Siècle Digital,

  • More than one in three Gen Z young people rely on AI for companionship, but those whose friendships are mostly online or with an AI report greater loneliness, according to a Gallup survey. Wired,

    Source404 Media
The Masteria read

When AI's sellers doubt the revenue, audit the task where the gain already shows

Baidu, Xiaomi, the IMF and 59% of French employers all said the same thing on the same day: AI revenue is lagging. Before buying a tool, audit the task where the gain already shows, not the technology.

Robin Li, head of Baidu, and Lei Jun, founder of Xiaomi, said the same thing on Tuesday 17 November from the APEC podium in Shenzhen: the money poured into AI far exceeds what it brings in, and the industry would do better to aim for concrete applications. These are not professional sceptics. They are two of the biggest AI sellers on the planet.

On the same day, the signals converge. Nvidia reports a quarter at $115.8 billion, yet becomes "more disciplined and selective" in its investments. The IMF describes a world economy standing on a single leg, AI's, and sounds the alarm on the cross-financing between players in the sector.

Elon Musk admits that Grok costs more than it earns. And in France, 59% of the employers surveyed by the Observatoire de France Travail, across 16,000 establishments, think AI will have little or no effect within five years. Nearly six employers in ten, on the ground, do not yet see any sign of the promise.

In this chorus we read a simple shift. The burden of proof is changing sides. For three years, saying "we are deploying AI" stood in for a strategy and was enough to reassure a leadership team.

When the champions themselves acknowledge the gap, that sentence no longer protects anyone. In 2027, the question put to a project manager will not be "did you install the tool," but "show me where it pays."

The answer does not run through a new purchase. It starts with an audit of the work, not of the technology, as the Journal du Net reminds us. Take a specific task, put a figure on the time it costs today, measure what a tool saves on it over a month, and decide.

The 59% of employers who see no effect are often right about their own reality: AI was placed there on the most visible tasks, rarely on the ones that weigh on the accounts. The gain exists, but somewhere other than where it was sought.

The question is no longer whether AI works. It works. The question is where it pays, for you, on which task, and how soon. Those who have mapped that before buying will spend less and prove more.

The Masteria editorial team.

Mathias Nizan, founder of Masteria
Mathias Nizan
Editor of AI Watch
Method

175 articles read, 13 kept

Selected on the morning of 18 November from 38 news feeds. Our method (in French)

Today's sourcesTechCrunchNext.inkLa TribuneJournal du NetClubicFrandroidSCMParXivSiècle Digital404 Media

Next edition: Thursday 19 November.

What the 18 November edition means for your teams

Nvidia beats expectations at $115.8 billion and shuts the door on Jensen Huang. Does it touch your teams? Let's talk.

Contact us

On compte les visites, pas les visiteurs.

En acceptant, vous aidez un cabinet indépendant à savoir quelles pages vous sont utiles et à améliorer les autres. Aucune publicité, aucune revente, aucun suivi d'un site à l'autre, et vous pouvez changer d'avis à tout moment.