OpenAI freezes part of Astra over cyber risk
AI Watch, Monday 10 August 2026

As the heatwave forces EDF to shut down the Golfech reactor, for lack of water cool enough to cool it, ZDNet sets the €75 billion of announced data centre investment (the server warehouses that run AI) against the grid's real capacity to power it. Thermal stress on the lines and generation points makes AI's energy bill less theoretical than a press release. For a public decision-maker or an industrial site manager, where to place a data centre is now inseparable from how to power it and how exposed it is to the climate.

The GDPR lets an organisation give other duties to its data protection officer (the DPO, tasked with ensuring personal data is used properly), provided those duties do not put it in the position of judge in its own cause. The CNIL (France's data protection authority) publishes a grid for identifying these situations: a DPO who would define the very processing they are meant to oversee, for instance an IT or marketing director holding the role on the side. For a legal team or an IT department, the note gives a concrete test before handing out the DPO hat, at a moment when AI projects multiply the sensitive processing to watch.

Several Journal du Net analyses make the same point: general-purpose AI stalls on the precision of a real profession, and an agent only becomes useful once it takes on the vocabulary, the constraints and the liability specific to a trade. A baker, a lawyer, an engineering firm, three trades with nothing in common share the same blind spot before a general model that ignores their implicit rules. The reasoning echoes another column on "brand voice", where a company that does not encode its way of expressing itself watches its content become interchangeable. For a team equipping a trade, value shifts towards this work of making things explicit, a subject we develop in today's read.

On Monday 10 August, Meta launched Muse Glimmer, an open-weight model (which you can download and run yourself), multimodal and agentic, light enough to run on a personal computer without going through a data centre. The announcement extends the open strategy shown last week with Muse Code, the group's low-cost coding agent. By making a capable model downloadable and runnable locally, Meta revives the debate in the United States over open access to an ever more powerful technology that some would rather keep under control. For a team concerned about the confidentiality of its data, a model that runs on its own machines answers a constraint that online services do not lift.

The Taiwanese foundry TSMC, which manufactures the chips designed by Nvidia, AMD and Apple, reported a 45% rise in monthly sales, a sign of steady appetite for AI hardware despite market jitters. The same day, Sony and TSMC are in talks over a joint $6.4 billion investment in an image sensor plant in Japan, while Intel raises $15 billion in shares to fund its growth in data centres. Spending on AI infrastructure keeps feeding the whole silicon chain, from foundries to memory makers. For anyone tracking chip costs, these figures say that the hardware constraint, not demand, remains the limiting factor.

Anthropic, the maker of Claude, has formed a strategic partnership with the Australian asset manager Macquarie and Singapore's sovereign fund GIC to build data centres dedicated to its models. The deal shows the labs' race to lock in, ahead of time, the compute their growth depends on, funding it with outside capital rather than from their own cash alone. The structure follows other AI-linked debt deals, where players such as BlackRock arrange financing worth several billion. For a customer of these models, the concrete stake is continuity of service: a supplier that secures its compute is a supplier less exposed to shortages.

Wired documents a turn: platforms acknowledge that the public does not want to consume content churned out by AI, the "slop", and a growing number of sites and apps are adopting tools and rules to detect, label or ban these outputs. The shift touches social networks, marketplaces and editorial sites, all trying to preserve their readers' trust. The turn answers a saturation: when automatic content becomes indistinct and pervasive, the human signal regains value. For a brand or a newsroom, the stake shifts towards proof of human involvement and towards an explicit policy on the use of AI.

China's AgiBot, based in Shanghai, has overtaken its domestic rival Unitree to become the world's top humanoid robot vendor in the first half of 2026, according to the firm Smart Analytics Global. AgiBot took 44% of the global market by shipping about 8,400 humanoid robots between January and June, carried by the boom in "physical AI", machines that embody models in the real world. Both companies are preparing to go public. China's dominance of this nascent segment is being settled while the West is still at the demo stage, and it gives Beijing an industrial lead on a market set to weigh heavily. For a European company considering service robotics, what is available and at what price will increasingly be decided in Shanghai.

China's most advanced AI models are still trained on Nvidia chips, according to sources at several large developers cited by the South China Morning Post, with the prohibitive cost of switching to local semiconductors holding back Beijing's self-sufficiency goal. Domestic hardware is improving, but changing chip architecture is a considerable engineering bottleneck: rewriting the software stacks, revalidating the models, absorbing the performance losses. "Training models on Nvidia chips remains the norm for now among Chinese developers," one source sums up. US export controls therefore matter less than the accumulated software dependence. The hardware decoupling that has been announced will take years, not quarters.

The rally in memory makers' shares is running out of steam, according to analyses relayed by the South China Morning Post, with the slowing rise in prices raising the question of how long the cycle will last, even as AI-driven demand stays strong. The price rises that produced record profits in recent quarters would signal a late-cycle phase, just as Chinese producers prepare to add capacity. For a company budgeting for server hardware, the signal is a warning not to extrapolate today's memory shortage over several years.