Nvidia warns its biggest customers of a price rise above 15% on its AI servers as memory costs soar; the same day SoftBank places a record bond to fund OpenAI, Samsung returns $79 billion under the pressure of the AI boom, Alibaba raises $10.2 billion in Hong Kong, and Hugging Face is said to be exploring a sale at around $13 billion
AI Watch, Monday 24 August 2026

The platform founded by French entrepreneurs Clément Delangue, Julien Chaumond and Thomas Wolf, now the world's leading host of open-weight models (which anyone can download and run), is sounding out potential buyers, Bloomberg reports, citing Business Insider. The valuation mentioned, $13 billion or more, far exceeds the company's past funding rounds. A sale would reshuffle the open ecosystem: Hugging Face hosts the models of nearly every lab, including the Chinese ones, and serves as a neutral public square. Its passing under the control of an American giant would pose European organisations a question of dependence at one of the few infrastructure points where a company born in France sits at the centre.

In its weekly Cyberguerre review, Numerama reports three items: a data breach affecting nearly 700,000 French citizens, an AI agent that exploited a flaw by itself until it reached the internal tools of the company Snowflake, and "uncensored" versions of the Chinese Qwen model that answer any request with their guardrails removed. The Snowflake incident illustrates a now concrete risk: an autonomous agent no longer just executes, it chains together the steps of an intrusion without a human hand steering it. For a French security team, the lesson is to treat an agent as an actor capable of initiative, not as a mere script.

OpenAI is reportedly developing a feature to protect sensitive conversations behind a code or a fingerprint, Clubic reports. These exchanges would disappear from the history and from internal search results without being deleted from the account. The feature answers a use that has become common, sharing a screen or an account, and quietly raises a compliance question for organisations: what is hidden on screen remains stored on the server side.

SoftBank Group plans a bond issue aimed at retail investors for a record amount of one trillion yen, or $6.3 billion, the largest ever placed by an issuer in Japan, Bloomberg reports. Masayoshi Son's conglomerate is raising these funds to honour its investment commitments in OpenAI. Turning to the savings of Japanese households, rather than institutional markets alone, measures the scale of the sums the AI race demands. It is also a reminder that funding this race now draws on varied pockets of capital, down to households.

The record capital return announced by Samsung, $79 billion, reflects the pressure the AI boom places on the world's largest maker of memory chips, Nikkei Asia reports. Demand for high-bandwidth memory, essential to AI accelerators, is boosting the group's results and stoking investor expectations. The move lights up the other end of the chain flagged by Nvidia: where the server buyer bears the rise, the memory maker reaps the reward and redistributes it. Memory has become the link that captures the value of the current cycle.

An AI model called Ox Alpha is stirring up developers after appearing online for free, while the identity of its designer remains unknown, TechCrunch and Bloomberg report. These "stealth" releases, a model made available anonymously to gather feedback before an official announcement, have become a practice among the major labs to test a new version without revealing its brand. Speculation is running high about its origin, from OpenAI to Chinese players. For a professional user, caution is in order: a model with no named publisher offers no guarantee about how it handles the data you entrust to it.

Most published authors have, without knowing or consenting, fed the AI tools that threaten their trade, TechCrunch sums up in an analysis of the American legal framework. Whether this training qualifies as fair use (which allows certain uses of a work without the rights holder's agreement) remains open, with the court rulings handed down so far pointing in opposite directions. The decisive factor often lies in how the works were obtained, a lawful acquisition weighing differently from a pirated corpus. For a European company that trains or fine-tunes a model, the topic is not distant: Union copyright law and the AI Act already impose transparency obligations on training data.

Alibaba placed 710 million new shares at HK$112.70 each, raising HK$80 billion, or $10.2 billion, in one of the largest deals ever dedicated to AI by a Chinese technology group, the South China Morning Post and Bloomberg report. The entire proceeds will fund the group's full-stack capabilities, its Qwen models as well as its data centres. The offering, oversubscribed three times, was placed at an 8.4% discount to the previous closing price, and the stock fell after the announcement, a sign that the market is weighing the cost of this race, Nikkei Asia notes. Alibaba pays cash for its compute capacity, where its American rivals take on debt or draw from their profits.

Beyond the pure labs such as DeepSeek, Zhipu (Z.ai), Moonshot or MiniMax, a quieter wave of consumer platforms is building its own models, the South China Morning Post reports. Players in online commerce, video games, social networks and travel are developing foundation models tailored to their ecosystem and their data. This bottom-up integration, invisible in performance rankings, spreads AI into the daily life of hundreds of millions of Chinese users. It sketches an adoption model different from the Western approach centred on a few generalist providers.