The escalating AI infrastructure arms race took a sharp turn as Meta entered talks to lease computing power to Anthropic in a potential $10 billion deal. This arrangement pits two fierce AI competitors in an unusual symbiotic relationship, with Meta renting out its vast GPU clusters to the very company building models that rival its own Llama line. It also signals that even Anthropic, despite recent compute deals with SpaceX, cannot secure enough hardware to support its ambitions. Across the Pacific, TSMC committed another $100 billion to its Arizona expansion, bringing total U.S. investment to $265 billion. The scale here is staggering, a direct bet that geopolitical risk in Taiwan is now the single most important variable in the global chip supply chain. Meanwhile, the human cost of treating users as beta testers continues to mount. A streamer lost his 25-year-old Xbox and OneDrive account when Microsoft nuked it following a compromise, only restoring access after public pressure. The initial assumption was that the account belonged to a hacker, not a victim. This is the logical endpoint of automated trust and safety systems. A new essay captures the mood of the developer class perfectly, arguing that the human-in-the-loop is tired. The burden of reviewing AI-generated code, catching hallucinations, and justifying every output to a skeptical manager is creating a cognitive tax that undermines the productivity gains these tools are supposed to deliver. A fresh report from the AISI tracked the cyber capabilities of open-weight models versus closed ones, finding that the frontier remains firmly closed. Open models are catching up, but not in the areas that matter most for high-stakes offensive operations. This gap justifies the current regulatory push to gatekeep model weights, though it also suggests the debate over open-source safety is increasingly moot. The frontier is already locked behind corporate APIs and export controls.