AI Watch, Tuesday 8 September 2026Mistral AI raises €3 billion at a €21 billion valuation in a round led by Samsung; the labor-law scholar Jean-Emmanuel Ray reminds employers that the AI Act requires them to inventory their high-risk AI systems and inform staff by the end of 2027, while BPCE stands up its AI Factory and Bercy renegotiates the French state's IT contracts; China targets 9,800 eflops of compute and Huawei launches its first stacked-circuit chip; Revolut completes the first AI-initiated payment in France

In his column for Le Monde, the labor-law scholar Jean-Emmanuel Ray advises human resources departments not to wait: by the end of 2027, they will have to inventory and clarify the artificial intelligence systems used in their departments, then inform employees. He revisits the workplace side of the AI Act, the European regulation on AI, whose obligations are drawing closer. The message is to take the inventory now, before the deadline turns it into an emergency.

The banking group BPCE will have an AI Factory (an internal unit tasked with industrializing AI uses) this year, overseen by Yves Tyrode, chief AI, digital and payments officer, reports ZDNet. Its mission is to simplify the spread of AI across the bank's various business lines, from the adviser to the back office. The announcement illustrates French banks moving from scattered experiments to a dedicated organization.

David Amiel, minister for the budget and the civil service, urges ministries to reopen the contracts signed with the French state's IT suppliers, reports Clubic. The argument: thanks to generative AI, these suppliers produce more in less time, and Bercy wants that saving to benefit public finances too. The move raises a question that reaches beyond the public sector, that of how productivity gains are shared between the client who pays and the supplier who automates.

GPT-6 Astra, OpenAI's new model presented as a step toward AGI (artificial general intelligence), costs up to $33 for a single query, and it is not the priciest option in OpenAI's catalog, reports Next.ink. Since its release, several users have found that it burns through their usage quotas at speed. The figure gives a sense of the real cost of frontier reasoning: a single question put to the best model can be worth the price of a whole monthly subscription. For an organization working out an inference budget (the cost of running a model once it is trained), the price gap between a routine model and Astra changes the equation before any talk of value produced.

Anthropic has decided not to acquire the AI startup Decart AI, an acquisition valued at around $6 billion, according to sources close to the matter cited by Bloomberg. The report rests on anonymous accounts and has not been officially confirmed by either company. The retreat comes in a sector where labs are multiplying tie-ups to capture teams and technology. It suggests that Anthropic prefers, at this stage, to preserve its cash rather than pay a steep premium to grow through acquisition.

Reaching the point where AI performs most tasks better than a human used to pass as a measurable milestone; the heads of the major labs now present it as a fuzzy goal, one whose definition slips away as you get closer, reports Bloomberg. The shift in vocabulary matters: as long as AGI stayed a quantifiable threshold, you could argue over a date; now an elastic notion, it mainly serves to keep the promise alive. For the professional reader, the lesson is to judge these systems on what they do today in a specific job, not on the horizon attributed to them.

China's Ministry of Industry and Information Technology published on Monday a five-year plan setting a target of 9,800 eflops of intelligent computing capacity by 2030, a fourfold increase, reports the South China Morning Post. The plan calls for deploying compute clusters holding 100,000 accelerators each and an investment of 3.8 trillion yuan, about $532 billion. The effort targets the infrastructure that underpins the country's AI ambitions, at a time when US restrictions limit its access to the most advanced chips. Beijing answers with volume and state planning where access to frontier components is closed to it. The announced sum, more than $500 billion over five years, exceeds what most states devote to all of their public research.

Huawei unveiled on Monday in Guangzhou the Kirin 9050 Pro, described by Richard Yu Chengdong, head of its consumer business, as its most powerful Kirin chip to date, reports the South China Morning Post. The component introduces the LogicFolding architecture, which stacks logic circuits vertically to gain performance without relying on transistor miniaturization alone, out of reach for a China under sanctions. It makes real the Tau scaling law that Huawei had defended on paper the previous week. The stakes go beyond the smartphone, launched in a tri-fold version: proving that the country can design frontier components on its own despite the American block on equipment. A first mass-produced product still says nothing about large-scale reliability or manufacturing yields, two points where the case remains to be made.

Seoul wants to rise to become the third major global player in artificial intelligence, behind the United States and China, through massive investment and reforms, reports Le Monde. The ambition runs into internal resistance, over both the future of the labor market and the energy needs the country struggles to meet. A sign of that tension, the government plans to combine five state-owned electricity companies to prepare for the power demand of data centers and chip plants, reports Nikkei Asia. The Korean case shows that the AI race is played out as much on reactors and high-voltage lines as on models. A country that wants to train and run giant AI must first produce the electricity that feeds them.