ai-industry · 2026-10-15 · Tier 3

Industry roundup, 2026-10-15: Anthropic's Q3, Cohere buys AISLE, agent tooling ships

Industry roundup, 2026-10-15: Anthropic's Q3, Cohere buys AISLE, agent tooling ships

Date: 2026-10-15 · Sources: The Information (RSS), lab watcher (Cohere, NVIDIA, OpenAI, Anthropic, Microsoft HF), X Following feed, Gmail (Runpod, Concordia), YouTube Raw: raw/rss/2026-10-1[45]-the-information-*.md, raw/labs/2026-10-15-143721-labs.md, raw/gmail/2026-10-15-newsletters.md, X feed captures (private)

Funding, valuations and deals

  • Anthropic Q3 revenue $16.8B, +45% QoQ (Q2 was >140% per The Information, ~2.4x per FT), 3.6x all of 2025; adjusted profitable for a second quarter (11% adjusted operating margin); $1.70 revenue per $1 of compute; rival-lab distillation had been 5-7% of revenue until cut back in Q3; bankers have floated an IPO raising >$100B at ~$2T (The Information).
  • Anthropic licenses Opaque Systems (confidential computing: AI jobs inside hardware-sealed enclaves with cryptographic run attestations) and hires CEO Aaron Fulkerson and CTO Rishabh Poddar. Opaque was valued at $300M after a $24M Series B (X).
  • Cohere to acquire AISLE (autonomous vulnerability discovery; 500+ CVEs in year one, all 12 in OpenSSL's January release, 18 of 24 in curl's latest) after Reliant AI and Aleph Alpha (Cohere).
  • Anduril employee tender at ~$90B valuation, $94.80 per share, capped ~$1.3B (The Information).
  • Harbor ($10M seed, covered 10-14) details its policy: six named perils, from hallucination to IP infringement (X).
  • Sovereign funds approached Perplexity to run its agent product entirely on hardware they own (X, Bloomberg via @rohanpaul_ai).
  • Sweden startups raised $2.8B in 1H 2026, projected $5B for the year; Nordic AI clouds plan ~1 GW by end-2027; the Sferical One consortium (Ericsson, AstraZeneca, Saab, SEB, Wallenberg) runs 1,152 GB300 GPUs (NVIDIA).
  • Morgan Stanley CEO Ted Pick expects another decade of major AI investment with bottlenecks and overbuild periods (The Information).
  • People: OpenAI compute-strategy VP Sachin Katti leaves 11-15 to build something new (The Information).

Products and models

  • Meta Muse Wonder (Max) enters Agent Arena at #12 (+6.4%), $1.13 median per task (2.1x Muse Spark 1.3), within 1 pt of GPT-5.6 Sol at 25% lower cost (Arena).
  • Claude Code: dynamic workflows (script-held orchestration of many subagents), a session sidebar sorted by "needs you" with recap cards, and agent step tracing (latency, tokens, tool calls per step, OpenTelemetry-style export) in public beta; Max and Team plans now include monthly API credits ($100 / $200 / up to $500 pooled).
  • Microsoft UserRepGen-4B (MIT): a Qwen3.5-4B LoRA that writes task and behavior descriptions of a user from a conversation to condition user simulators (HF).
  • Grok Voice 3.0 claims the top conversational voice model; Apple Intelligence launches in mainland China in December (The Information).
  • Runpod Global Volumes on Serverless (one weights copy served across data centers; object storage, no locking).
  • OpenAI case studies: SentinelOne (malware analysis weeks to hours with Daybreak) and Block security; OpenAI says over half of EU enterprise AI use now involves agents, small-business users up 8x.

Policy and people

  • Jensen Huang: "If the product is not safe, just don't release it," and he would not sell GPUs to irresponsible labs. Anthropic's international chief says open-weight models carry more safety risk by definition (Bloomberg via X).

Related: TSMC and neocloud credit · Monitorability and governance