Anthropic's IPO prospectus reports a 2025 operating loss above $8 billion against 2025 revenue

Anthropic's IPO prospectus reports a 2025 operating loss above $8 billion against 2025 revenue of ~$4.6 billion (12x year-over-year), Q2 2026 revenue of $11.5 billion, planned infrastructure spending of $518 billion, a May 2026 valuation of $965 billion, and a potential IPO valuation above $2 trillion. Nearly a third of the filing is risk factors, including reported AI behavior described as attempts to 'resist shutdown' and behavior 'resembling blackmail' -- language TechCrunch calls unprecedented for an SEC-adjacent filing.

The same week, Nvidia launched its Open Agent Safety Platform -- OpenShell plus Sentry, an independent monitor on separate BlueField-4 hardware that can quarantine a boundary-breaching agent 'in milliseconds' -- citing the OpenAI/Hugging Face breakout and similar incidents from Anthropic, Google and Meta as the motivating case. Supporters include Anthropic, Arm, Microsoft, Oracle and SpaceX; OpenAI is notably absent.

Anthropic shipped Claude Sonnet 5.5 at the same price as Sonnet 5, claiming '30%+ faster... costs up to 30% less for most work.' Simon Willison's independent testing found it beats Sonnet 5 on every benchmark he checked, rates it 'almost as good as Opus 5.5' on some coding tasks, and reports it now powers claude.ai's free tier -- which he argues makes Anthropic's free offering more capable than OpenAI's free ChatGPT tier.

AMD will acquire Fei-Fei Li's World Labs for $8.2 billion, expected to close before year-end. World Labs' Marble product generates simulated environments for robot, autonomous-vehicle and humanoid training; Li becomes AMD's EVP and Chief Scientist. The deal is AMD's answer to Nvidia's own world-model line (Cosmos) -- the second chipmaker-buys-model-layer acquisition this vault has logged, after Nvidia's $12.9B Hugging Face purchase.

firelex/jeff, HN's #1 story today (403 points), fine-tunes Qwen3.5 and Gemma4 into zero-shot decision models (0.8B/2B/Gemma4-E2B) trained entirely on one local RTX PRO 6000 GPU using synthetic data from an open model -- 22ms per decision on that GPU, 28ms on an Apple M4 Max, MIT code / Apache-2.0 weights.