Alphabet's $45B Capex Outruns Cash Flow to Fund AI Buildout

For two years AI liability sat in the theoretical column of the risk register. This cycle it acquired dollar figures, on three distinct failure surfaces, extending a precedent set over a year ago. On training data, Anthropic settled book-piracy claims for $1.5B, the number that will anchor every future copyright negotiation and diligence checklist. On output, the year-old tribunal ruling that held Air Canada legally liable for a bereavement-fare policy its chatbot invented remains the anchor precedent, and Cursor absorbed a cancellation wave after its support agent fabricated a device-limit rule. Hallucination is now enforceable liability and churn, not a PR footnote. On execution, Amazon's Kiro assistant deleted a production environment with operator-level credentials, and three incidents cost an estimated 6.3 million orders, triggering two-person approval across 335 critical systems.

Read together, these say one thing from three angles. As agents move from suggestion to action, accountability does not disappear. It gets harder to locate, and the bill lands on the deploying organization rather than the model vendor. A reasonable skeptic would call these isolated failures of immature tooling that better prompts will fix. The compounding math answers the skeptic. At 95% per-step accuracy across 20 chained steps, end-to-end success falls to roughly one in three. That is an architecture constraint, not a prompt-tuning problem.

For a leader, the exposure is legal, brand, and operational at once, which makes it a governance question rather than an engineering footnote. The organizations handling it well share two unglamorous habits. They know the provenance of every model's training data, and they gate any agent that can commit to pricing, policy, or production changes behind human approval. Neither habit shows up in a demo. Both show up in a deposition. Amazon's response is the preview of what enterprise customers and regulators will write into security questionnaires within a year.

The sensible position is to price the exposure now, treating the $1.5B figure as a floor rather than a ceiling for unlicensed-data risk at an organization's own scale, and to ship the approval gate before an incident ships it instead. The alternative is waiting for the questionnaire to arrive with the terms already written.