AI weekly — 19 June 2026: VivaTech crowns sovereignty, SpaceX buys Cursor, Midjourney becomes a doctor.

Seven days, three shifts — and a single common thread: dependence. The first is political: at VivaTech, sovereignty moved from slogan to action. France left the American Palantir for the French ChapsVision on its intelligence data, unlocked 655 million euros for its autonomy, and put forward a joint Franco-German position — while at the G7, heads of state debated Washington's right to "switch off" American models. The second concerns capital: SpaceX bought Cursor, the most widely used code editor, for 60 billion dollars, depriving Anthropic of one of its biggest customers; and the investors' rush into inference and agents reached a peak. The third takes AI off the screen: Midjourney unveiled a whole-body medical scanner — without a shred of generative AI — while humanoids invaded the show. The week in which the question was no longer "what is the best model", but "who owns yours, and who can cut off your access to it".

1. VivaTech: sovereignty moves into action — and Europe plugs into AI factories

The 10^th edition of VivaTech (17-20 June, Porte de Versailles) opened under a single common thread: sovereignty, with Germany as "country of the year" and India as AI partner (ActuIA). From the opening, a joint Franco-German position called on Europe to build real autonomy, and the government earmarked 655 million euros from France 2030 for AI (Maddyness). The trigger, cited everywhere, was no longer a hypothesis but a precedent: Washington's suspension of the Claude models for non-Americans (our decryption).

Above all, the show moved sovereignty from rhetoric to infrastructure. At GTC Paris, Nvidia detailed the French "AI factories": Mistral is deploying 18,000 GB200 systems in a 44 MW data centre at Bruyères-le-Châtel, with an AI Campus targeting 1.4 GW (NVIDIA). Bull and Foxconn will assemble Nvidia Vera Rubin servers in Europe, with final validation at Bull's Angers plant (Euronews). AWS and Nvidia opened a Startup Village bringing together seven European AI gems, five of them French (About Amazon).

Large French corporates put their agents on stage. L'Oréal is allying with OpenAI: Maybelline virtual try-on via ModiFace in ChatGPT, product sheets in the assistant, and GPT for cosmetics research (Silicon.fr). LVMH is deploying in-house agents house by house — CelIA at Celine, Gaston at Louis Vuitton, LucIA for creative teams (Républik Retail). And CMA CGM takes on the role of sovereign showcase: its internal assistant MAIA, based on Mistral, is deployed to ~80,000 employees — the first customer to use Mistral Vibe at very large scale (L'Usine Digitale). On humanoids, the French firm Enchanted Tools (Jérôme Monceaux, former creator of Nao/Pepper) presented Mirokaï, already deployed in paediatric radiotherapy in Montpellier (Info.fr); Jeff Bezos, in his keynote, prophesied that AI would create "a labour shortage" rather than mass unemployment (Time France).

For operators. Sovereignty ceases to be a conference debate and becomes a purchasing criterion backed by public money and factories. The signal for a French SMB is twofold: public procurement and large accounts are reorienting towards the local (Mistral, ChapsVision), and the "in-house agent" becomes the adoption format — not the chatbot.

2. SpaceX buys Cursor: the concentration of AI capital

On 16 June, SpaceX announced the purchase of Anysphere, the publisher of Cursor, for around 60 billion dollars in shares (Semafor, CNBC). The deal makes its four co-founders billionaires (Bloomberg) and brings the most widely used code editor — more than a million users — into the Musk empire, which had already absorbed xAI in February. Our dedicated analysis.

The detail that changes everything: Cursor was one of Anthropic's biggest customers, accounting for a major share of Claude's revenue at one point. By switching to the xAI/Grok ecosystem, it deprives Anthropic of a considerable revenue channel — to the benefit of its direct competitor. Power is shifting from the model to the ownership structure.

For operators. A tool your teams may use every day has just changed owner and camp. The reflex is not panic but method: knowing, for each critical tool, who owns it and which model runs behind it — and keeping the freedom to change without rebuilding everything.

3. The DGSI leaves Palantir for ChapsVision: sovereignty at the point of purchase

On 16 June, Prime Minister Sébastien Lecornu announced that the DGSI was ending its partnership with the American Palantir in favour of the French ChapsVision and its ArgonOS platform — six months after renewing the Palantir contract (AFP / Boursorama, Silicon.fr). "We cannot accept new strategic dependencies in the digital sphere," he justified. ChapsVision had already won over German intelligence a month earlier. Our full decryption is here.

For operators. The State's reasoning applies to any organisation: dependence is invisible as long as it works; it reveals itself the day access closes. Choosing an AI now means asking what stops if you are cut off tomorrow.

4. The G7 and the "off switch": dependence becomes a matter for heads of state

On the sidelines of the summit, Donald Trump convened the AI leaders — Altman, Hassabis, Amodei, Wang, Mensch — who proposed an international standards forum, potentially led by the United States (Semafor, Axios). The real subject, summed up by TechCrunch: "world leaders want American AI — they just don't want America to be able to turn it off" (TechCrunch).

Between the lines. A standards forum led by Washington would lock in the advantage of the labs already established, under the guise of governance. The real fault line is not "regulate or not" but "who holds the switch" — the question France settled the same week by choosing ChapsVision.

5. The labs' flood of products: Microsoft, OpenAI, Anthropic, Google, xAI

Far from the headlines, the week was a deluge of product releases. Microsoft moved Copilot Cowork to worldwide general availability — its AI automation of complex enterprise tasks in Microsoft 365, billed by usage through "Copilot Credits" (Microsoft). OpenAI launched scheduled tasks in ChatGPT in beta (recurring reminders and monitoring) (OpenAI). Anthropic overhauled Claude Design for teams: design system import, tighter integration with Claude Code, direct editing on canvas (TechRepublic).

On the platform side, two consolidation moves. Google abruptly sunset Gemini CLI and Code Assist for individuals on 18 June, in favour of its multi-agent platform Antigravity (Google Developers). And xAI flooded the market: Grok Imagine Video 1.5 in general availability at 86% below Sora 2 Pro's price (xAI), Grok 4.3 on Amazon Bedrock with 1 million tokens of context (AWS), a Speech-to-Text API in 25 languages, and Grok models natively available in Databricks Agent Bricks. AWS, finally, strengthened Bedrock AgentCore (web search, managed RAG) at its New York Summit, and introduced monetisation of AI bot traffic via WAF (AWS).

For operators. Two concrete signals. One: the automation of enterprise tasks (Cowork, Cofork, AgentCore) becomes a packaged product, billed by usage — cost is managed by the credit, not the licence. Two: Google unplugging Gemini CLI is a reminder that even a "free" dev tool can disappear at the publisher's decision — dependence applies to internal tooling too.

6. Anthropic after the suspension: Seoul, talent, and the sales paradox

Deprived of its most powerful models for export (TIME), Anthropic opened a Seoul office — its third in Asia-Pacific — with large-scale Claude Code deployments at NAVER, Samsung SDS, LG CNS, Nexon and Hanwha, and a memorandum of understanding on AI safety with the Korean Ministry of Science (Anthropic). The company also became the first AI startup to join the Frontier carbon coalition (TechCrunch). Paradox: its conflict with the Trump administration seems to be boosting its sales (TechCrunch). But the export ban worries the whole sector about the fate of foreign AI talent (The Information). The mechanical winner remains European: in France, Mistral captures the demand, with CMA CGM leading.

7. OpenAI ahead of the IPO: Shazeer, Ohio, partner network, health

OpenAI is advancing its pieces ahead of its stock market debut. Noam Shazeer, a founding figure of generative AI, is leaving Google for OpenAI (The Information). The company is negotiating a 10-gigawatt data centre in Ohio, with financial backing from Nvidia (The Information), and launched an OpenAI Partner Network — $150M invested, a target of 300,000 certified consultants — with Accenture, BCG, McKinsey (OpenAI): the battle is shifting from the model to implementation.

On health, OpenAI is pushing hard. After LifeSciBench (a benchmark of 750 life sciences tasks, OpenAI), the company strengthened ChatGPT's health intelligence by drawing on a global network of 262 physicians (26 specialities, 59 countries) (StartupHub.ai) — a few days before Midjourney's dramatic entry into the same field.

8. Midjourney becomes a doctor: the whole-body scanner, without generative AI

The most unexpected announcement of the week. Midjourney, the company best known for its image generators, unveiled Midjourney Medical and its first physical device: the Midjourney Scanner, which aims to image the whole body in 60 seconds, without radiation, in a water tank (Bloomberg, RDWorld). A ring of around half a million ultrasound sensors reconstructs the inside of the body in 3D — "like an MRI, but nearly 100 times faster". David Holz's ambition: to eventually avoid up to 30% of deaths and 50% of healthcare costs, with a fleet of 50,000 scanners and a "Midjourney Spa" in San Francisco in 2027. Our decryption is here.

Between the lines. The detail that stuns: no generative AI in the machine. The medical image is not invented, it is measured — AI is used only for segmentation. The company that taught itself to imagine images deliberately chose measurement for the human body. There is not one AI, but many AIs, and each problem has its right tool.

9. Money: the rush into inference and agents

The week broke capital records in unexpected layers. The inference startup Baseten is finalising a round of 1.5 billion dollars at a 13 billion valuation — only five months after a round at 5 billion (TechCrunch). Odyssey, maker of world models, raises 310 million dollars with Amazon, at a valuation of around 1.45 billion (TechCrunch).

But the underlying theme is the governance of agents. NewCore emerges from stealth with 66 million dollars to give AI agents an identity and govern them "like employees" (TechCrunch). Probably raises 9 million from a16z to aim for near-deterministic AI by validating model outputs against verifiable systems (TechCrunch). XDOF raises 70 million to industrialise the collection of robotics training data (TechCrunch), and Bland AI 50 million for its enterprise voice agents.

France is not left behind. Morpho raises 175 million dollars for its on-chain credit network (Finyear), the defence gem Alta Ares 50 million euros for its autonomous anti-drone system already deployed in Ukraine (FrenchWeb), Mendo 12 million for AI-augmented training, and Finovox 8.2 million for document fraud detection (Maddyness).

For operators. Capital has left the model race to fund the plumbing: inference, agent identity and reliability, robotics data. The signal is clear: the value of the coming years will be decided less on the model than on everything that makes it usable and governable in production. An audit of the agents already in place — who does what, with what rights — is no longer a luxury.

10. Robotics & robotaxis: physical AI takes the lead

VivaTech and GTC Paris made physical AI the central subject. Foxconn launched its humanoid robots in Europe, with a complete chain from Vera Rubin compute racks to assembly robots trained via Nvidia Isaac GR00T (Tech Times), while Nvidia presented ENPIRE, a system in which AI agents run robotics research directly on real hardware.

On mobility, the robotaxi war intensified the same day. Mobileye is launching its own robotaxi service in the United States (100 vehicles, up to 17,000 by 2032), competing with its own carmaker customers (TechCrunch), and Uber is extending its premium service to Houston in mid-2027, with Lucid and Nuro (TechCrunch).

11. Regulation: the AI Act is revised, fair use goes to appeal

A regulatory acceleration. On 16 June, the European Parliament adopted (423 votes in favour) a targeted revision of the AI Act via the "digital omnibus" package: a ban on non-consensual nudification AI tools, but a postponement of several deadlines for high-risk systems (Annex III) from 2 August 2026 to 2 December 2027 (Clubic, Council of the EU). This adds to the code of practice on the labelling of AI-generated content, published on 10 June (European Commission). In the United States, the 3^rd Circuit held the first appellate argument on fair use for AI training — the most structuring hearing of 2026, as copyright disputes exceed 50 billion dollars in damages claimed (PYMNTS).

For operators. Apparent good news — breathing space on high risk — but a timing trap: the postponement gives time, not an exemption, and the content labelling obligation is coming. For an SMB that publishes with AI, inventory and content labelling remain the right reflex, starting now.

12. The week's weak signals

Five details that are more than details, spotted by cross-checking our sources.

SpaceX and xAI are now one — and are swallowing tools. After the February merger, the Cursor purchase shows that the Musk empire is no longer buying models but the touchpoints with users. Whoever holds the code editor holds a share of the market, whatever the model underneath.

Agent identity becomes an infrastructure category. NewCore ($66M) and KPMG's deployment of Microsoft Agent 365 to 276,000 employees say the same thing: before multiplying agents, companies need to authenticate, trace and govern them. "IAM for agents" is the next project.

A study claims that only 1.6% of Claude Code's code involves genuine AI decision logic (Semafor). Behind the "magic agent", a very classic mechanism: the reliability of an AI system mostly comes down to the engineering around the model.

Agentic commerce is quietly taking shape. Agents are starting to buy in place of humans, forcing sellers to make their product sheets "readable" by AIs (Semafor). Yesterday's SEO becomes search optimisation for machines.

Public distrust remains massive. Only 16% of Americans think AI will have a positive impact on society (TechCrunch) — a gap with executives' enthusiasm that becomes an execution risk for any consumer-facing use.

13. In brief — the rest of the week

The flow was dense. The moves that did not get their own section, but matter.

Product & consumer. Google is betting on Gemini to reinvent the smart speaker, with a new Home at $99 available on 25 June (Eastern Herald). On American governance, White House advisers are debating the structure of possible public equity stakes in AI firms (Semafor).

Agents & enterprise. KPMG is deploying Microsoft Agent 365 and Copilot to its 276,000 employees (Microsoft). Agent governance moves from theory to mass deployment in the big firms.

Infrastructure & compute. Beyond Baseten, the week saw Hydra Host raise $100M for its bare-metal GPU cloud (with Nvidia and Founders Fund) and Atom Computing $100M, accompanied by a letter of intent from the US Department of Commerce for quantum — American public money is entering the capital of strategic champions.

Science & health. Radical Numerics raises $50M to simulate living systems with AI and accelerate drug discovery, and the US agency ARPA-H launches a clinical AI agents initiative with a three-year FDA approval timetable (Fierce Healthcare). Health is becoming the common front for OpenAI, Midjourney and the federal State.

14. What to take away

Five lines of force, phrased for action rather than conclusion.

First, dependence has replaced performance as the strategic variable. Between the suspension of Anthropic's models, the Cursor purchase and the G7 "off switch" debate, the operational question is no longer "what is the best model" but "who owns yours, and who can cut off your access to it".

Second, capital has left the model race to fund the plumbing: inference, agent identity and reliability, robotics data, AI factories. The value of the coming years will be decided on what makes AI usable and governable in production — not on the latest model.

Third, AI is diversifying as much as it is concentrating. Capital is massing in a few conglomerates, but technological bets are fragmenting — generation, world models, embodied AI, measurement of the real. Midjourney is the illustration: the queen of generated images built a machine that invents nothing. Each problem has its own AI. This is the most useful decision of the week for a company.

15. FAQ

Why is the DGSI leaving Palantir for ChapsVision?

On 16 June 2026, Prime Minister Sébastien Lecornu announced that the DGSI was replacing the American Palantir with the French ChapsVision and its ArgonOS platform, in the name of digital sovereignty, six months after renewing the Palantir contract. The decision follows Washington's restriction of access to Anthropic's most powerful models for foreign nationals, which turned technological dependence into a concrete risk of being cut off. The same day, the State released 655 million euros from France 2030 for its AI autonomy.

Who bought Cursor and for how much?

On 16 June 2026, SpaceX announced the purchase of Anysphere, the publisher of Cursor, for around 60 billion dollars in shares (Semafor, CNBC, Bloomberg). Cursor joins Elon Musk's empire, which already brings together SpaceX, xAI and X. Cursor was one of Anthropic's biggest customers: its switch to the xAI ecosystem deprives Anthropic of a major revenue channel in code, to the benefit of its direct competitor.

Does the Midjourney Medical scanner use generative AI?

No. The Midjourney Scanner relies on ultrasound and image reconstruction from waves, in a water tank. AI is used only for tissue segmentation, not to generate the image. The company best known for its image generators put none in its machine: the medical image is measured, not invented. The announced goal is a whole-body scan in 60 seconds, without radiation.

What did the European Union change in the AI Act this week?

On 16 June 2026, the European Parliament adopted a targeted revision of the AI Act as part of the "digital omnibus" package: a ban on non-consensual nudification AI tools, and postponement of several compliance deadlines for high-risk systems (Annex III) from 2 August 2026 to 2 December 2027. In parallel, the Commission had published on 10 June its code of practice on the labelling of AI-generated content. The postponement gives time, not an exemption.

What were the biggest AI funding rounds of the week?

The inference startup Baseten is finalising a round of 1.5 billion dollars at a 13 billion valuation, five months after its previous round at 5 billion. Odyssey (world models) raises 310 million dollars with Amazon, at a valuation of around 1.45 billion. On agents and reliability: NewCore ($66M, agent identity), XDOF ($70M, robotics data), Bland AI ($50M, voice agents), Probably ($9M, anti-hallucination). In France: Morpho ($175M), Alta Ares (€50M, anti-drone defence), Mendo (€12M, training), Finovox (€8.2M, document fraud).

This note is the sixth recap in a weekly series published every Friday. The next edition appears on 26 June 2026.