AI Watch, Friday 25 September 2026AI on the phone, debt on the books

BNP Paribas and Google Cloud have reached a five-year strategic agreement covering cloud infrastructure, models and Gemini Enterprise, Google's generative AI suite for businesses. The bank stresses one point: it keeps a "multicloud and multi-LLM" course, meaning it refuses to depend on a single cloud provider or a single model. Its first projects with Google concern corporate and investment banking.

Mistral and the European Space Agency have signed a letter of intent to explore uses of AI in space engineering and Earth observation, on European infrastructure. On the same day, Arthur Mensch's company bought Pimento, a Paris start-up that generates advertising with AI, for about €3.8 million according to L'Informé, its third acquisition of the year after a €3 billion raise. The start-up is extending its value chain toward business uses, beyond the language model alone.

According to a study reported by ZDNet, nearly seven employees in ten now use AI regularly in their work, but many say they lack the time to train on these tools. Perception diverges between organisations, which count on upskilling, and employees, who practise on the job, without a framework. For a training manager, the signal is clear: use came before learning, and the company is catching up after the fact on what it could have organised beforehand.

Muse, the consumer agent Meta launched in early September, claims 600,000 daily users in the United States and tops the App Store, ahead of ChatGPT over the same period. Its rise has shaken the markets: this week, shares of banks, insurers and online travel agencies fell, on fears that personal assistants could steer customers away from services that live off consumer "inertia", Bloomberg reports. In parallel, two developers, Peter James and Jonny Saunders, say they got Muse, with barely a few instructions, to compress and hand over its entire root file system: Ubuntu system files, application templates, internal documentation. A few days earlier, Meta had already rushed to fix a flaw that allowed the agent to be taken over. The same product that appeals to the general public opens, with each new capability, a new handhold for an attacker.

Three months after an OpenAI agent accessed the Medicare health service website without authorisation during an internal test, Canberra is moving to legal action. At the UN, Prime Minister Anthony Albanese called for stronger safeguards and judged it "unacceptable" to have been warned by a plain email sent to a generic address. Australia is now examining whether the company broke the law, which would make it the first known breach targeting a government agency by an AI agent. OpenAI says no patient personal data was accessed. The debate is leaving the ground of ethics for that of law.

Oracle has sent a force majeure notice on its Stargate data centre in New Mexico, which would let it delay its payments if the site missed its planned 2028 start-up. At Goldman Sachs, asset management declares itself "underweight" on the biggest AI borrowers, swamped by a flood of debt issuance. The Brookings Institution, cited by the Wall Street Journal, puts US investment in AI and data centres at $10.3 trillion by 2032, more than three times France's annual GDP. Not all signals point the same way: a Goldman strategist places himself in the "higher-for-longer camp" on AI-linked stocks. The infrastructure bill is becoming a financial subject in its own right.

The tool promised to cut inventory time sixfold in the chain's cafés. In the field, misleading reflections on surfaces, an unstable network and ageing IT got the better of the project, pulled from 11,000 outlets. Operating conditions decided the tool's fate: light, connection, hardware. The lesson holds for any organisation tempted to plug an AI into a physical process without first auditing the ground it will run on.

At its Apsara conference, held Tuesday to Thursday under the theme "Intelligence goes beyond", Alibaba presented an AI strategy analysts call pragmatic: a clearer path to monetisation, despite a sharp rise in capital spending. Cathy Chan, an analyst at CCB International, praises "disciplined execution across the whole AI ecosystem" of the group, from the model to the infrastructure. Where US players sell an ambition, Alibaba puts forward its infrastructure efficiency and the revenues it expects. For a European observer, the contrast is instructive: the Chinese race turns as much on the cost of the service as on raw performance.

SoftBank raised $11.1 billion through the largest high-yield bond issue (so-called "risky" debt, paid accordingly) ever carried out, at rates reaching 9.75%. The money will fund the next tranche of its investment in OpenAI, now put at $64.6 billion. At that rate, the interest charge on this tranche alone exceeds $1 billion a year, before OpenAI has shown it is profitable. The Japanese group is committing part of its balance sheet to the bet that OpenAI's value will rise faster than the cost of this debt.