SoftBank commits up to 75 billion euros to French AI data centers

SoftBank committed up to 75 billion euros to build AI data centers in France with up to 5 gigawatts of capacity, while GitHub Copilot’s shift to token-based billing on June 1 has developers reporting 10-25x cost increases over the previous flat-rate subscription. Anthropic published a detailed breakdown of its containment architecture across Claude products, revealing three major vulnerability categories discovered through real security incidents.

Infrastructure #

SoftBank says it will invest up to 75 billion euros to build French data centers #

TechCrunch

SoftBank announced its largest AI infrastructure investment in Europe, committing up to 75 billion euros to build data centers across three sites in the Hauts-de-France region (Dunkirk, Bosquel, and Bouchain), targeting 3.1 gigawatts by 2031 with a total goal of 5 gigawatts. The investment aligns with President Macron’s agenda to position France as an AI infrastructure hub and contrasts with growing opposition to data center expansion in the US, where SoftBank is simultaneously pursuing a 9.2-gigawatt natural gas-powered project in Ohio. The scale – roughly $87 billion – signals that sovereign AI compute capacity is becoming a geopolitical asset, with European governments actively competing for hyperscale commitments.

I put a datacenter GPU in my gaming PC for 200 pounds #

blog.tymscar.com / Lobsters AI

A developer installed a secondhand Tesla V100 SXM2 (16GB HBM2) for approximately 150 pounds alongside an RTX 4080, creating 32GB of total VRAM capable of running Qwen3.6-27B at 32 tokens per second with a 128K context window. The setup required a bare PCB adapter for the SXM2 socket and creative fan wiring to manage 82-decibel cooling noise. The practical takeaway: datacenter GPUs from the previous generation are now cheap enough that individual developers can run competitive inference locally – the V100 cost less than two months of some cloud API bills.

Developer Tools #

GitHub Copilot’s new token-based billing spurs consternation among developers #

TechCrunch

GitHub Copilot is switching from flat-rate subscriptions to token-based usage billing effective June 1, with developers reporting projected cost increases from $29 to $750 or $50 to $3,000 monthly depending on usage patterns. Critics argue Microsoft encouraged heavy consumption under the subsidized flat rate and is now shifting the true cost to users, while defenders counter that extreme bills reflect inefficient “vibe coding” workflows rather than reasonable usage. The pricing shift forces a reckoning across the industry: flat-rate AI developer tools were loss leaders, and consumption-based pricing reveals the actual cost of AI-assisted development that organizations were not tracking.

Funding & Business #

OpenRouter raises $113M Series B #

OpenRouter / Hacker News (420 points)

OpenRouter raised $113 million in Series B funding led by CapitalG (Alphabet’s independent growth fund), with participation from NVentures, ServiceNow Ventures, MongoDB Ventures, Snowflake Ventures, and Databricks Ventures. The company’s weekly token volume grew from 5 trillion to 25 trillion over six months, now serving over 8 million developers across 400+ models with plans to process over one quadrillion tokens this year. OpenRouter has expanded beyond text to support image, audio, speech, transcription, embedding, and video models, positioning itself as the routing and abstraction layer between applications and the increasingly fragmented model provider landscape.

Security #

How we contain Claude across products #

Anthropic / Simon Willison

Anthropic published a detailed engineering post documenting three distinct containment architectures: ephemeral gVisor containers for Claude.ai, OS-level sandboxing (Seatbelt/bubblewrap) with human-in-the-loop approval for Claude Code, and sealed VMs with platform hypervisors for Claude Cowork. The post disclosed that users approve roughly 93% of Claude Code permission prompts, prompting the addition of an OS-level sandbox that reduced prompt volume by 84%. Three vulnerability categories emerged from real incidents: pre-trust execution (config parsing before user consent), user-directed injection (a phished employee completed credential exfiltration 24 of 25 times), and allowlist exploitation (using approved API endpoints for data exfiltration). The level of detail – including specific failure rates and attack vectors – sets a transparency standard that other AI companies have not matched.

Regulatory & Policy #

California AI bills clear crossover deadline #

Transparency Coalition

Nearly all of California’s 30 AI-related bills passed their chamber of origin ahead of the May 29 crossover deadline, advancing to the second house for consideration before the July 2 summer adjournment. The bills span chatbot disclosure requirements, AI in professional licensing, and broader governance frameworks. Combined with Illinois passing its AI safety testing law earlier this week and Connecticut’s bipartisan SB5, the state-level regulatory momentum is accelerating rapidly – three major states advancing substantive AI legislation in a single week suggests that the federal vacuum is being filled from below, creating a patchwork that frontier labs will need to navigate.

Other #

Meta is reportedly developing an AI pendant #

TechCrunch

Meta is developing an AI-powered pendant based on technology from its late-2025 acquisition of Limitless, a wearable AI startup. The device would be worn as a necklace or clipped to clothing and can record conversations, extending Meta’s hardware strategy alongside its AI glasses and a new “Wearables for Work” business subscription. Reality Labs lost $4 billion in Q1 2026, and Meta is betting that ambient AI capture devices will succeed where previous AI wearables (including Humane’s AI Pin) have struggled with privacy concerns and limited utility.

Threads to Watch #

The true cost of AI-assisted development is becoming visible. GitHub Copilot’s shift to token-based billing, with developers reporting 10-25x cost increases, ends the era of subsidized flat-rate AI coding tools and forces organizations to confront per-token costs they were not tracking. OpenRouter’s $113M raise to scale model routing – essentially infrastructure for managing AI consumption efficiently – suggests that cost optimization across model providers is becoming a significant business in its own right.

Transparency about AI system security architecture is deepening. Anthropic’s detailed documentation of containment failures – including specific vulnerability categories, approval rates, and attack success rates from real incidents – goes further than any prior disclosure by a frontier lab. This level of transparency may be driven partly by regulatory pressure: California’s 30 AI bills clearing crossover and Illinois’s safety testing law create compliance obligations that reward companies who can demonstrate how their systems are actually secured, not just that they have security policies.

Sources Unavailable Today #

These sources could not be fetched today. Links point to their homepages so you can check them directly.

• Stanford HAI — scrape: content_truncated