AI Watch, Friday 2 October 2026Measure who can actually use AI before you name an owner or buy a platform

Unveiled on 29 September, Arcadia aims to merge French military data on a large scale to gain an operational edge, Le Monde reports. The stakes go beyond the technology: at issue is the place of the American groups Palantir and Anduril in the modernisation of the Atlantic Alliance's structures, where France intends to offer a sovereign solution. The file squarely raises the question of who holds and processes European defence data.

The group completed on Thursday its acquisition of this cybersecurity company, known for QShield, a product that hardens software against AI-assisted analysis and reverse-engineering techniques, according to Clubic. The deal, announced in April, strengthens Airbus's cyber arm and keeps a sensitive software security component in Europe. It illustrates a quiet move by the continent's big industrial groups to consolidate France's cyber gems.

The firm's study highlights a growing need to designate a clear AI owner inside the company, failing which responsibility for incidents stays unassigned, ZDNet reports. Executives tend to send the subject back to the technical teams, who send it back to senior management. The gap widens at the moment when uses are multiplying and an agent going off the rails acts rather than merely writing.

The publishing and training group has reached an agreement to acquire Workera, whose product assesses employees' real level of AI competence and directs training to where it is missing, ZDNet reports. The purchase makes the measurement of AI skills a market in its own right, separate from selling courses. For a management team, the logic flips: you first seek to know who can do what, before deciding what to teach.

The two groups are trying to convince individuals to entrust their agents with actions more committing than mere writing, purchases and administrative steps included, Bloomberg writes. At its DevDay conference, OpenAI unveiled Dots, a "serious" agent powered by GPT-6 Astra, to answer Meta, whose agent platform Muse has had a noticed start, The Verge reports. The shift is clear: after the assistant that suggests comes the agent that executes and pays in your place. These tools, designed for individuals, will enter organisations through employees before any internal policy. The vendors' bet is that public trust will follow convenience; nothing guarantees it will follow as fast as the features.

The US chain, one of the country's largest grocery retailers, says it uses these tools to speed up its teams' work and simplify the shopping of millions of customers, according to OpenAI, TechCrunch and ZDNet. The announcement matters less for the technical feat than for the scale: a mass-retail player is industrialising generative AI across its internal processes. This is the kind of documented rollout that gives a concrete measure of enterprise adoption, far from demos. It remains to be seen what quantified gains the retailer will publish once it has enough hindsight.

Clubic describes a deal under which Anthropic commits to rent $42 billion of AI chips supplied by Broadcom, an example of those arrangements where the same players finance each other in a circle. The same day, Amazon is studying a transfer of around $8 billion of high-end Nvidia chips off its accounting balance sheet, according to the Financial Times as relayed by Bloomberg. The two operations tell the same strain: AI infrastructure costs so much that it pushes towards ever more acrobatic financial arrangements. For anyone buying AI services, these tens-of-billions commitments end up weighing on the price billed.

Judge Amit Mehta threw out two actions filed by Chegg and Penske Media, parent company of Rolling Stone, which accused Google of capturing web traffic with its AI-generated search summaries, according to The Verge, citing Reuters. The ruling sides with Google on this specific ground. For the publishers who see their audience shrink as search engines answer in place of websites, the legal route closes and leaves the economic problem untouched.

Many Hong Kong users saw their account suspended without notice after connecting through a virtual private network (VPN), a tool that hides the real location, the South China Morning Post reports. A wave of complaints appeared on Thursday on X and Reddit, from accounts active for months. Anthropic applies geographic restrictions to its models, and is strengthening its detection of workarounds. The measure is a reminder that access to the big American models depends on a shifting political map, and that a team based in a restricted zone can lose its tool overnight.

Federal prosecutors arrested on Thursday a Californian accused of shipping to China servers containing $300 million of Nvidia chips, in violation of export controls, according to Bloomberg. In parallel, a financing company owned by several Chinese local authorities documented the purchase of Blackwell chips that are nonetheless restricted, a sign of public support for trade that Washington deems illicit. The two cases show export controls that hold poorly: the hardware gets through despite the barriers, and players tied to the Chinese state organise its acquisition. The question put to the US authorities is why the warning signs were not seen earlier.