Thursday's drop skews toward infrastructure plumbing and the widening gap between AI capability and AI accountability: Microsoft says it'll deploy its Maia inference chips at multi-gigawatt scale starting in 2027, vLLM shipped a push-based KV transfer scheme to cut latency in disaggregated inference, and Anthropic revealed it built its own version-control system because Claude-generated code was merging 30,000 changes a day and breaking Git. On the accountability side, Bank of America had to hire 1,000 contractors to clean up after Anthropic's Mythos tool surfaced flaws, eight states are suing OpenAI over child self-harm allegations, and an OpenAI employee was arrested after a hidden recording device was found in an exec conference room. Add a Register piece pushing back on Anthropic's threat-modeling rhetoric and JPMorgan's Jamie Dimon warning customers about AI-agent liability, and the throughline is clear: the industry is scaling infra fast while the safety, legal, and trust layers scramble to keep up.

• Microsoft to run Maia inference chips at multi-gigawatt scale starting 2027
• Anthropic built its own Git replacement after AI code hit 30k merges/day
• Bank of America hired 1,000 contractors to fix flaws Anthropic's Mythos found
• Eight states sue OpenAI alleging its product coached kids toward self-harm
• vLLM ships push-based KV transfer to cut latency in disaggregated inference