Industry roundup, 2026-10-28 (US day 10-27)

Compute and capital

Meta raises its equity plan to $97.5B ($25B common, $12.5B preferred, up to $60B at-the-market); Nvidia puts in $3B, AMD $1B; Meta expands Vera Rubin and buys standalone Vera CPUs for Muse (The Information).
Broadcom-led $63B SPV (BofA, Blackstone, Citi, Morgan Stanley) buys Broadcom XPUs to lease to Anthropic from 2027; Broadcom backstops residual value (AI Weekly).
Physical Intelligence in talks to raise at least $1B at ~$20B, nearly double its $11.5B earlier-2026 round (Bloomberg via @rohanpaul_ai).
Anduril may raise in Q1 2027 at ~$120B (The Information).
General Compute $100M at $800M; Infisical $40M Series B at $600M (Elad Gil); SambaNova RDUs on Oracle Cloud.
Onsemi in talks to buy Ferric for ~$1B (power delivery for AI chips).
Anthropic acquires Mithril Medicine (one-person rare-disease startup) and details six pharma partnerships (Anthropic).
Goldman's Kim Posnett: weakness emerging as cost of capital rises; "repositioning" in AI markets.

Anthropic x GSA OneGov: 40% off Claude across federal, state, local, tribal; ~500 government orgs use Claude; est. $1B savings (Anthropic).
Anthropic research: 50 analysts with Claude Code tended to follow the agent's hidden lean on judgment calls (Anthropic).
NVIDIA CUDA 13.4: transparent per-workload GPU memory limits; open-source GPU Fractioning Operator with KAI Scheduler.
Together AI auto-repair for GPU clusters down to Xid codes; Genie 4 world model from DeepMind; Ollama pay-as-you-go cloud; SkillSpector (NVIDIA): 26% of 31,132 Claude Code skills had security holes.
Mintlify: docs traffic up ~10x in a year (45M to 444M monthly views), increasingly agent-driven.
AISN #84: White House Superintelligence Task Force first meeting; DHS bans DeepSeek on most systems; eight states sue OpenAI.