SoftBank to Invest Up to €75 Billion in AI Data Centers in France
SoftBank's largest AI infrastructure commitment in Europe: up to 5 GW of capacity, with a first phase of 3.1 GW in Hauts-de-France by 2031.
On May 30, 2026, SoftBank announced an investment of up to €75 billion (roughly $87 billion) to build data centers dedicated to AI in France: according to TechCrunch, it is its largest infrastructure commitment of this kind in Europe. The plan aims to bring online up to 5 GW of new total capacity.
The first phase — 3.1 GW to be completed by 2031 — is concentrated in the Hauts-de-France region, with sites at Dunkerque (Loon-Plage), Bosquel, and Bouchain. SoftBank, which is both an investor in and a customer of OpenAI, is thus replicating in Europe the same scale logic already seen in the United States, where it announced a data center in Ohio powered by a 9.2 GW natural gas plant.
The French government welcomed the announcement as a confirmation of its own strategy: Economy Minister Roland Lescure called it "a testament to President Emmanuel Macron's ambition to position France as a leading destination across the entire AI value chain." The scale of the commitment shifts the AI competition onto the terrain of energy and territory: capacity of this size requires massive access to electricity and suitable sites — factors that in France rest largely on the nuclear fleet. Independent coverage from Data Center Dynamics reports the same figures on the size of the buildout.
Why it matters
- Entrepreneurs: AI compute capacity of this scale on European soil could strengthen the localized offering in Europe and the digital sovereignty narrative, provided it is actually available to European customers. At the same time, electricity demand of up to 5 GW adds to upward pressure on energy prices. Compliance and access variables in the investment plans remain to be monitored.
- ICT engineers / IT managers: New on-shore compute capacity in Hauts-de-France means potential access to AI compute with European latency and data residency. That is an advantage for workloads subject to regulatory constraints. However, actual availability over time should be assessed, given that the first phase arrives only by 2031, along with the energy and territorial dependency of infrastructure of these dimensions.
Meta Launches Paid 'Plus' Subscriptions for Instagram, Facebook, and WhatsApp
Meta is rolling out monthly paid tiers for its apps globally, priced between $2.99 and $3.99, and is preparing separate plans for AI, creators, and businesses under the Meta One brand.
Meta has officially launched paid subscriptions for Instagram, Facebook, and WhatsApp, placing a range of consumer features behind a paywall for the first time. According to TechCrunch, which broke the news on May 27, 2026, the plans are Instagram Plus and Facebook Plus at $3.99 per month and WhatsApp Plus at $2.99 per month, following several months of testing.
The tiers unlock exclusive features: on Instagram and Facebook, detailed Stories analytics, ephemeral posts extended beyond 24 hours, custom themes and reactions (the "Super heart" reaction), the ability to select multiple audience lists, the option to see who rewatches a Story, and the ability to post without appearing in your followers' feeds. As Engadget reports, WhatsApp Plus adds app themes, exclusive ringtones, enhanced stickers, and additional pinned chats.
Meta also announced a new umbrella brand, Meta One. Under this brand it will test AI plans — Meta One Plus at $7.99 and Meta One Premium at $19.99 — as well as plans for creators and businesses, with Meta One Essential at $14.99 and Meta One Advanced at $49.99. Meta Verified remains unchanged. On Daring Fireball, John Gruber notes that Meta, having to go through Apple's and Google's app stores, will pay in-app commissions. To avoid them and maintain a direct relationship with the subscriber, it may push users toward the web.
Why it matters
- End users: Personalization and Stories-control features that until now were free or new are becoming paid: those who don't subscribe stay on the basic experience, while the number of monthly micro-subscriptions to evaluate app by app keeps growing.
- Entrepreneurs: With the Meta One creator/business tiers (Essential at $14.99 and Advanced at $49.99) come a verified badge, a boost in search ranking, analytics, and scheduling tools: new recurring costs but also visibility levers for those who use social media as a commercial channel.
ChatGPhish: ChatGPT's Web Summaries Become a Phishing Surface
Researchers at Permiso Security have shown how a hostile page, once summarized by ChatGPT, can inject links, fake security alerts, and malicious QR codes into the assistant's 'trusted' interface.
On May 29, 2026, Permiso Security disclosed a prompt injection technique dubbed ChatGPhish, which exploits ChatGPT's web page summarization feature. When a user asks it to summarize a public page — a GitHub README, a documentation portal, a blog, or a SaaS dashboard — hostile instructions hidden in the content silently make their way into the model's response. The researchers describe it as a Cross Prompt Injection Attack (XPIA), an evolution of earlier demonstrations against Microsoft Copilot.
The technical crux is chatgpt.com's renderer implicitly trusting links and images in Markdown format coming from the page it just summarized: the assistant retrieves them and displays them as active, clickable elements within its own interface. From here, the report describes four attack vectors:
- phishing links presented with no attribution of their origin;
- fake "account security" alerts that inherit the visual credibility of the assistant;
- QR codes generated automatically from attacker-controlled S3 buckets, which bypass URL checks because the destination is revealed only on scanning;
- passive tracking images that transmit the victim's IP, User-Agent, and timing data.
The disclosure timeline starts with a Bugcrowd report on April 29, 2026 ("Untrusted Markdown Rendering Leads to XSS, Phishing, and Data Exfiltration"), followed by revised submissions on May 1 and May 7. According to Cybersecurity News, OpenAI responded that the initial report was not reproducible and classified the revised version as a duplicate of an already known issue. No specific fix was detailed. The Hacker News coverage confirms the dynamic and recommends treating links and alerts in summaries as potentially attacker-controlled.
Why it matters
- End users: The attack targets the very interface the user considers trustworthy: malicious links, alerts, and QR codes appear inside the assistant's response with no signals about their origin. The practical defense is not to automatically trust what ChatGPT shows after summarizing a page, to verify the real destination of links, and to be wary of alerts that push you to act in a hurry.
- ICT engineers / IT managers: ChatGPhish shows that the output of an AI assistant summarizing external content must be treated as untrusted, like any unverified input. The countermeasures: limit the permissions of AI browsers, require human confirmation on links, avoid summarizing user-generated content, and monitor anomalous requests to unknown domains. It is also a concrete case of XPIA to bring into corporate security policies.
Mistral AI expands from generative models to industry and infrastructure: Le Chat becomes Vibe
The French lab connects generative AI to industrial simulation and design, builds a network of European data centers, and renames Le Chat to Vibe, making a decisive push into the enterprise market.
Mistral AI has announced an expansion that takes the French lab beyond generative models, toward industry and infrastructure. The centerpiece is an industrial AI technology born from the acquisition of Emmi AI, which connects generative models to simulation and design software. This allows manufacturers to accelerate design and development: they use AI for rapid predictions and reserve conventional simulations for verification only. The technology is already being tested at Airbus, BMW, and ASML — the latter applying it to support its service technicians and to software development, significantly reducing the time required for certain analyses.
On the infrastructure front, Mistral is building a European network of data centers. A training facility near Paris is already operational; a second facility in Les Ulis, in Essonne, dedicated to inference, is expected in the third quarter of 2026. The company attributes the move to "rapidly growing" demand for European AI capacity and to the need for greater control over data processing.
The Le Chat assistant is also being renamed Vibe and extended into an organizational platform: managing email and documents, scheduling, and integration with Microsoft, Google, and collaboration tools. In parallel, Mistral has consolidated functions previously spread across separate models — image analysis, coding, reasoning — into a single multimodal generation designed for technical and industrial applications. The overall message is clear: a repositioning toward companies and governments that prefer a European alternative to U.S. AI providers (Techzine).
Why it matters
- Entrepreneurs: For business leaders, especially in manufacturing, a European and sovereign alternative to the U.S. giants is opening up. AI connected to simulation and design (tested by Airbus, BMW, and ASML) and data centers located on European soil reduce dependence on foreign technology and offer greater control over data. This is an increasingly relevant factor for compliance and operational continuity.
AutoTTS: an Agent Discovers on Its Own How to Scale Inference-Time Compute, −69.5% Tokens at the Same Accuracy
A research framework turns the design of test-time scaling strategies into an automated search driven by a coding agent: it finds a controller that cuts tokens by ~69.5% compared to Self-Consistency@64 while keeping the same accuracy.
A group of thirteen researchers from University of Maryland, University of Virginia, Washington University in St. Louis, UNC, Google and Meta has presented AutoTTS, described in the paper "LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling". The idea: instead of hand-designing the heuristics of test-time scaling (TTS) — how much compute to allocate at inference for multi-step reasoning — the task is handed off to a coding agent that tackles it as an algorithmic search.
The agent proposes controllers defined via code, evaluates them against a "frozen" replay environment (reasoning trajectories already stored), receives fine-grained feedback on the execution traces, and refines the solution. A beta parameterization collapses several hyperparameters into a single scalar, making the search space tractable. The key point: during the evaluation of the controllers there is no call to the base LLM, because the replay uses cached segments. The proposal and refinement of the controllers, on the other hand, are carried out by a coding agent such as Claude Code.
The controller that emerged — the Confidence Momentum Controller — unifies the decisions of width (when to branch) and depth (how far to extend the branches), with trend-based stopping and cautious abandonment of branches. The results come from Qwen3 models at four scales (0.6B, 1.7B, 4B, 8B). The search starts from AIME24: ~69.5% tokens saved compared to Self-Consistency@64 (at β≈0.5), with unchanged accuracy on the held-out benchmarks AIME25 and HMMT25. A single discovery run cost about $39.9 and 160 minutes of wall-clock time. Code and data are public (AutoTTS repository, project page); independent coverage comes from VentureBeat.
Why it matters
- Frontier research: It is a concrete case of "LLMs improving LLMs": an agent automates the discovery of inference strategies, outperforming hand-designed heuristics and shifting TTS from a craft art to an algorithmic search problem with feedback on replayed traces.
- LLM builders / devs: A ~69.5% cut in tokens at the same accuracy directly affects the inference costs of multi-step reasoners; with public code and data and a discovery costing ~$40 and ~160 minutes, it is a replicable approach for optimizing your own compute budget.