Masteria, cabinet IA à Lyon : audit, conseil, développement et formation
Certifié QualiopiFinançable OPCOEurope · États-Unis · Inde
Français
Thursday 19 November 2026 edition

AI Watch, Thursday 19 November 2026Anthropic goes public, record debut

Masteria editorial team, led by Mathias Nizan · , 8:35

Anthropic goes public at $1.5 trillion, the first listing of a frontier AI lab and the largest fundraising ever completed.

  • The same day, Anthropic and Google cut to 30 days the time allowed to fix a flaw, a study puts at 25,000 the job openings lost in France, and Tencent sextuples its investment to catch up in the Chinese AI race.
14 stories14 sources7 min read
Top story
Cited in the analysis

Anthropic sets its IPO at $83 a share and is worth $1.5 trillion

The maker of the Claude model has fixed its share price at $83, above the announced range of $70 to $78, for a valuation of around $1.5 trillion. With its associated private placement, the company and its seven co-founders raise $107.3 billion, the largest such deal in history and the first stock market debut of a frontier AI lab.

The money is meant to fund long-term commitments on computing power and data centres. Amazon (14.7%) and Google (13.7%) will together hold nearly 28% of the capital, through non-voting shares. Masayoshi Son, head of SoftBank, called missing the investment a "mistake".

Le Monde, Bloomberg, TechCrunch,

SourceClubic

Europe and France

Le Figaro signs with OpenAI in turn

The Figaro group becomes the third French press publisher to formalise an agreement with OpenAI, after Le Monde in 2024 and the media outlets of the LVMH group in mid-November. When answering its users, ChatGPT will offer summaries and excerpts from the group's titles, along with links to the original articles.

The contract authorises OpenAI to use the journalistic content of the group's brands. No amount was disclosed. Le Monde,

SourceNext.ink

Qwant, Shadow and OVHcloud seal an alliance for sovereign AI

Qwant, the Shadow cloud service, European Search Perspective and OVHcloud have formalised a partnership for a "European and sovereign" AI, hosted and operated in Europe end to end, from web search to computing power. First concrete result: at Qwant, instant answers now rely on Olala mini, a French model that handles three million queries a day.

The whole effort revolves around Octave Klaba, the founder of OVHcloud. The bet is still to hold a place between the American giants and the low-cost Chinese open models. Next.ink,

SourceClubic

A study puts at 25,000 the job openings lost in France

Credit insurer Coface, in a study unveiled on 17 November, measures 25,000 fewer openings than expected on France Travail (the French public employment service) in the third quarter of 2026. The drop is concentrated in the jobs most exposed to generative AI. According to France Travail, 41% of employers expect a direct impact of AI on their business within five years.

The study gives a first tangible figure to an effect so far mostly forecast. Siècle Digital,

Source01net

International

Banks refuse to cover customers who entrust their finances to an AI

Citi joins the major institutions that decline any liability towards customers using AI agents to manage their money, even in the event of hacking: the loss stays with the user. The same day, the Monetary Authority of Singapore (MAS) warns that AI agents can move deposits and funds faster than a human, creating new liquidity and financial stability risks.

A Meta assistant, Muse, also hallucinated transactions and read broadly into its users' data access permissions. The spending power handed to agents is arriving faster than the safeguards meant to frame it. TechCrunch, Bloomberg,

SourceNext.ink
Cited in the analysis

Anthropic and Google shorten the flaw disclosure window

Anthropic will now give publishers only 30 days to fix a serious flaw it reports to them, down from 90. Three days earlier, Google's Project Zero team announced a comparable reduction, phased in through 2028. The reason fits in one sentence: AI uncovers vulnerabilities faster than teams fix them.

Google estimates that autonomous attacks, carried out end to end by an AI on its own, are already hitting French websites. The pace shows on the patch side too, where Google rolled out in eight days its third Chrome security update, closing 204 flaws. Numerama, 01net,

Satya Nadella explains the buyout offer made to 60,000 employees

Microsoft's CEO details why the group offered 60,000 employees in October the chance to leave voluntarily, rather than launching a new round of layoffs. He ties the decision to AI. Microsoft is looking for teams able to work with agents, and the issue is as much about the skills expected as about the number of posts.

The move shows how far the sector's large employers are already recomposing their jobs around the tools.

Osmo claims to have given AI a human-level sense of smell

American start-up Osmo says it has built a "foundation model" of smell, a general-purpose model trained once then reusable, whose artificial nose matches that of a human. In a laboratory near Newark airport, some thirty professional "sniffers" describe each scent ("hazelnut", "peach", "marine") and cross these notes with a chemical analysis to feed the training.

The goal is to map every smell on the planet digitally. The intended applications range from perfumery to early disease detection.

China and Asia

Tencent sextuples its investment to catch up in the AI race

Tencent more than sextupled its capital spending in the third quarter, above $12 billion, to close its gap in the Chinese AI race. Adjusted net profit fell 6% year on year, to 66.5 billion yuan ($10 billion), in line with the analyst consensus; revenue rose 8%, to 207.3 billion yuan.

The group calls the American chip restrictions "regrettable", warning they could slow its model training in the short term. The priority given to computing power weighs on margins and cash.

Washington could require companies to declare their use of Chinese AI

The White House is considering requiring American companies to state publicly whether they use Chinese AI models. According to Reuters, the plan is under discussion between the presidency's science policy office and the Treasury Department, with no guarantee at this stage that it becomes a rule.

The measure would extend the policy of distrust towards DeepSeek, Qwen and the other Chinese open models. It would force companies to track precisely the models they run.

SourceNext.ink

Indian start-up Sarvam resells access to DeepSeek

Sarvam, presented as a spearhead of India's technological independence, has started selling access to DeepSeek and other Chinese models, stirring an online controversy. The company is in parallel pursuing the development of a sovereign Indian model.

The case shows the tension in emerging ecosystems, torn between an ambition for sovereignty and dependence on the cheapest open models.

SourceBloomberg
Research

Research and papers

For technical teams

OpenAI measures how much its agents speed up its own research

The lab publishes an assessment of what coding agents contribute to its internal work and describes a "jagged frontier": the gains are clear on some tasks, uneven on others. The underlying question is recursive self-improvement, when AI already helps build the next AI.

The exercise matters above all because it comes from the lab itself, putting a figure on a phenomenon often described without data.

SourceOpenAI
The rest of the news

In brief

  • More than three in four French people now use generative AI, according to a Havas Market study; 77% of them also use it to prepare their purchases, which raises the question of how objective the answers are, as advertising begins to slip in.

  • Scammers broadcast a fake Tesla livestream on YouTube in which an AI-generated Elon Musk promised to double the cryptocurrencies sent in; according to Coinbase, the platform blocked $4.7 million in transfers from its customers before they left.

    SourceClubic
The Masteria read

AI has shortened the gap between finding a flaw and exploiting it

Anthropic and Google cut the patch window from 90 to 30 days: their models find flaws in hours. An organisation's security will rest on its reaction time, not its tooling.

Anthropic has cut from 90 to 30 days the time it gives a publisher to fix a serious flaw before making it public. Three days earlier, Project Zero, Google's elite team, announced the same shift. The reason is identical on both sides: their own models find vulnerabilities faster than human teams can patch them.

The 90-day rule dated from 2014. For twelve years, it set the tempo of security work worldwide: a publisher had three months to deliver a fix. That window has just been cut by two thirds, under a concrete pressure.

Google estimates that autonomous attacks, driven end to end by an AI, are already hitting French websites. The same week, the maker of Chrome rolled out three security updates in eight days.

What is at stake is speed. For years, an organisation could live with a patch backlog of several weeks without trouble: exploiting a known flaw took time, skill and a determined attacker. AI erases all three frictions. It reads a security bulletin, writes the exploit (the code that attacks the flaw) and launches it while your team is still opening its ticket.

For a French IT manager, the consequence can be dated. The window between "a flaw is published" and "it is attacked" is now measured in days. A patch plan set on a monthly cycle becomes a risk. The question to ask on Monday fits in two measures: how many hours separate the alert from the deployed fix, and who, by name, watches the bulletins.

Anthropic set this new tempo on the very day it went public at $1.5 trillion. The company that sells the tool able to find the flaws also sets the clock everyone will have to follow.

An organisation's security will be measured by its reaction time, more than by the thickness of its walls. In 2027, the budget line that counts holds in a single figure: the number of hours separating an alert from its fix.

The Masteria editorial team.

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

174 articles read, 14 kept

Selected on the morning of 19 November from 38 news feeds. Our method (in French)

Today's sourcesClubicNext.ink01netSiècle DigitalTechCrunchMIT Technology ReviewSouth China Morning PostBloombergOpenAI

Next edition: Friday 20 November.

What the 19 November edition means for your teams

Anthropic sets its IPO at $83 a share and is worth $1.5 trillion. 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.