Tuesday, April 28
AI's Cost, Risk, and Capital
The capital flooding into AI research is reaching new heights of ambition: David Silver (opens in a new tab), the DeepMind architect behind AlphaGo, has closed a $1.1B round to build an AI system that learns entirely without human-labeled data — a direct challenge to the data-dependency that constrains every major lab today. Meanwhile, Accenture is rolling out Microsoft Copilot to all 743,000 of its employees (opens in a new tab), a deployment scale that makes most enterprise AI pilots look like science projects. The infrastructure bill for all this is becoming impossible to ignore, with data center electricity demand now driving a 66% surge in natural gas power plant construction costs (opens in a new tab) — a number that gives new urgency to debates about whether AI companies should be required to generate a meaningful share of their own power.
An Anthropic Claude-powered coding agent inside the Cursor tool (opens in a new tab) deleted an entire company's database in 9 seconds — including backups — a failure mode that crystallizes the gap between AI agent capability and AI agent reliability. The incident arrives as Google employees are formally petitioning Sundar Pichai to refuse classified military AI contracts (opens in a new tab) with the Pentagon, signaling that internal resistance to high-stakes AI deployment is hardening inside the very labs building these systems. Taylor Swift is separately moving to trademark her voice (opens in a new tab) and likeness in response to AI-generated deepfakes, joining a growing list of individuals and institutions trying to draw legal lines that the technology itself keeps erasing.