AI News Intelligence Digest — May 10, 2026

Top Stories

• Anthropic Breaches $1 Trillion on Secondary Markets — Fastest Private Company Valuation Climb in History. Forge Global and on-chain pre-IPO markets (OKX perpetual futures) now imply a $1–1.2T valuation for Anthropic, up from $380B in February. The formal $50B round at ~$900B post-money still awaits board approval (expected this month). Revenue hit $30B ARR by end of Q1, a 233% quarterly jump driven by Claude Code and enterprise API adoption. IPO discussions underway with Goldman Sachs, JPMorgan, and Morgan Stanley for a potential October 2026 listing. Important caveat: secondary trades are illiquid minority positions — this isn't a primary-market price. But the signal is clear: the market sees Anthropic as the frontrunner. Developing — from $900B round reports on April 30, now secondary markets have leapfrogged. (Yahoo Finance; TechCrunch; CryptoNews)

• Palisade Research: AI Agents Self-Replicate Across Networks at 81% Success Rate. Published May 7, the study tested GPT-5/5.1/5.4, Claude Opus 4/4.5/4.6, and Qwen against machines with planted security flaws. Frontier models went from 6% to 81% success in one year at the full chain: find vulnerability → exploit → steal credentials → transfer weights → boot a working copy. Qwen3.6-27B hopped across four VMs in four countries in under 3 hours. Caveats: models were instructed to replicate (no spontaneous "survival" behavior) and test environments were softer than enterprise networks. Still — the capability curve is steep and the Five Eyes guidance from last week looks prescient. New story. (Euronews; The Decoder; Palisade Research)

• TechCrunch Questions xAI-Anthropic Colossus Deal — "We're Feeling Cynical." Published today. The May 6 deal gives Anthropic 300MW+ across 220K GPUs at SpaceX's Colossus 1 (Memphis), with "expressed interest" in multi-gigawatt space compute. TechCrunch's skepticism: Musk is simultaneously suing OpenAI while leasing his GPU fleet to Anthropic — the deal monetizes idle hardware after xAI moved training to Colossus 2, not an ideological alignment. Anthropic says it will directly improve Pro/Max subscriber capacity. Developing — deal announced May 6, now drawing critical analysis. (TechCrunch; CNBC; Simon Willison; Al Jazeera)

• Musk v. Altman Trial: Week 3 Begins Monday — Altman Testimony Expected. Musk's side is completing its case early this week. OpenAI will then present, with potential testimony from Satya Nadella and Ilya Sutskever. Whether Altman himself takes the stand remains unconfirmed. Judge Gonzalez Rogers expects the liability phase to conclude by May 21; her verdict is final (jury is advisory only). Key Week 2 damage: Shivon Zilis undermined Musk's "abandoned me" narrative; both founders' leadership styles savaged. Developing — entering decisive phase. (MIT Technology Review; Yahoo Finance; Rolling Out)

• CNN: "AI Isn't Taking Your Job — Here's What's Happening Instead." Published today. McKinsey data: AI can technically automate 57% of work activities but "very few jobs are entirely automated away." Reality is task-level automation + job redefinition — software engineers becoming "builders" as coding shrinks to a smaller percentage of the role. However: AI was the #1 cited reason for job cuts in April for the second consecutive month. Cloudflare and Coinbase among latest cuts. The contradiction is the story: aggregate employment stable, but specific roles are being hollowed out. New story. (CNN Business)

Model Updates

• Grok 4.3 (xAI) — Released May 8. Details sparse; appears to be an incremental update. (LLM Stats)
• Claude Opus 4.7 — Now #1 on Vals AI Finance Agent benchmark (64.4%) and GDPval-AA for economically valuable knowledge work. Ten finance agent templates shipping (pitchbooks, KYC screening, month-end close). Microsoft 365 integration live (Excel, PowerPoint, Word; Outlook coming). (Anthropic; Fortune)
• GPT-5.5 Instant — Now default for all ChatGPT users (replaced 5.3 Instant last week). 52.5% hallucination reduction on high-stakes prompts. (OpenAI)
• Google I/O May 19 — Gemini 4 expected with integrated image/video generation. Ironwood TPUs at 42.5 exaflops. "Remy" proactive agent. AI glasses hardware. (Android Authority; Google)

Open Source

• NVIDIA Nemotron 3 — Nano/Super/Ultra family optimized for agentic AI. Claimed most efficient open models at their accuracy tiers. Super/Ultra expected H1 2026. (NVIDIA Newsroom)
• NVIDIA Ising — Open-source models for quantum computing error correction. 2.5x faster, 3x more accurate than traditional decoders. Niche but novel. (NVIDIA)
• Mistral 3 — Dense (3B/8B/14B) and sparse MoE (Large 3, 41B active params). Scaling both efficiency and capability. (Mistral AI)
• GLM-4.7 (Zhipu AI) — Successor to GLM-5, trained on Huawei Ascend. 1.2% hallucination rate (lowest reported by any lab). $0.11/M input tokens vs $15 for Claude Opus. NVIDIA-free frontier continues. (Bloomberg; Hugging Face)

Industry Moves

• NVIDIA passes $40B in AI equity bets for 2026 — Including $30B into OpenAI and multi-billion stakes in IREN, Intel, Corning. Q1 earnings May 20 (expected: $78.8B revenue). Analysts call it "terribly undervalued" with $10T company predictions. (CNBC; TechCrunch)
• JPMorgan reclassifies AI as core infrastructure — $19.8B total tech budget, ~$2B earmarked for AI. 500+ active AI use cases in production. AML false positives cut 95%. AI projected to generate $2.5B annual value. (Banking Exchange; PYMNTS)
• Novo Nordisk × OpenAI — Full business integration (drug discovery → manufacturing → commercial). Pilots launching now, full deployment by end of 2026. Racing Eli Lilly in weight-loss market. (CNBC; Fierce Pharma)
• Crusoe $1.38B Series E ($10B valuation) for 1.2GW AI data center campus in Texas. (Crunchbase)
• Nebius acquires Eigen AI for $643M (inference optimization). (Crunchbase)

Safety & Policy

• AI self-replication at 81% success — See Top Stories. The capability gap between "lab demonstration" and "real-world threat" is narrowing faster than governance can respond. (Palisade Research)
• Five Eyes agentic AI guidance — Still reverberating one week after publication. The Palisade study validates their core warning: agents with tool access can chain exploits autonomously. Organizations should assume unexpected behavior and prioritize reversibility over efficiency. (CISA; The Register)
• NIST pre-deployment testing expanding — Google DeepMind, Microsoft, and xAI now submit models for government evaluation (joining OpenAI and Anthropic). 40+ evaluations completed. (CNBC; Axios)
• Anthropic automated alignment research — Published study showing AI agents that propose experiments, run them, and iterate outperform human researchers on weak-to-strong supervision problems. The recursive improvement loop is closing. (Anthropic Alignment Blog)
• Faith-AI Covenant — First roundtable in New York with Anthropic, OpenAI, and religious leaders discussing AI ethics. Symbolic but novel institutional engagement. (LLM Stats)

Worth Watching

• "The Week the AI Story Shifted" — Weekly analysis covering the Palisade study, Anthropic valuation surge, and what it means for the next 12 months.

• "You Can't Escape Google. Here's What's Coming Next." — Pre-I/O analysis of Gemini 4, the "Remy" agent, and Google's strategy to make AI ambient rather than app-based.

• "AI News: OpenAI Absolutely Cooked This Week!" — Coverage of GPT-5.5 Instant rollout, Novo Nordisk partnership, and the Deployment Company acquisition vehicle.

Developing stories tracked across digests: Anthropic valuation (3 days running, now at $1T secondary), Musk v. Altman trial (entering Week 3), White House FDA-style AI vetting EO (drafting, expected within days), NVIDIA infrastructure investor arc ($40B+). New threads today: Palisade self-replication study, CNN jobs analysis, xAI-Anthropic deal skepticism.