LWIA #347: Nvidia spy probe, Anthropic's $1.4T IPO, near-recursive self-improvement warning

Last Week in AI · rss · 2026-11-11

Last Week in AI #347 recaps several weeks of major AI news:

Nvidia spy scandal

Nvidia's board opened an independent investigation after reports that China's Ministry of State Security recruited Alice Huang (a former Megaspeed executive, no relation to the CEO) to target Jensen Huang. Huang wrote to staff admitting an affair ("I was unfaithful to my wife"); the board said it has "no reason to believe" national security or company secrets were compromised and retains confidence in his leadership.
Export-control holes exposed: the NYT reported New Matrix, an approved H200 recipient, occupies an empty office tied to a Chinese PE firm blacklisted since 2024. The Trump administration relaxed Biden-era export limits after Nvidia agreed to hand the government 15% of sales.
Other firms brace for threats: WSJ reports OpenAI, Anthropic and Google run counterintelligence-style training on bribery, eavesdropping, honey pots and seduction. An OpenAI employee was arrested for allegedly planting a bug in a conference room.

Anthropic IPO pricing

Range $70–78/share, 825M Class A shares, raising $57.75–64.35B; up to $1.404T fully diluted valuation; Nasdaq ticker ANT around Nov 19, up to 20% retail; net proceeds est. $84.6–93.7B.
Financials: $81B run-rate revenue by end of October (+$10B vs September), Q3 revenue $17B, nine-month revenue $33.2B; nine-month net loss $104B (mostly non-cash fair-value charges on convertibles); compute/infra commitments $900.9B vs $132B cash. All seven co-founders cash out $250M each, retaining control via Class F shares with 50.1% of votes.
Pricing lands below the earlier $2T ambition; after the Nasdaq's 5%+ single-session drop (worst in 19 months), investors saw even $1.5T as a stretch. CoreWeave and neocloud stocks sold off; Ramp data shows token spending grew just 1.5% across Aug–Sep (after 42% in July). OpenAI raised $11B of an expected $20B round at a $1.25T valuation.

Anthropic warns it may be nearing recursive self-improvement

Revised S-1/A risk factors say the company "may imminently meet" its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI" — and concedes it has previously identified accelerations only after they occurred. No current model has been designated as exceeding the threshold, and Anthropic doesn't expect Claude to design successive frontier models safely without meaningful human supervision.
Same day, it released Mythos 5.2 and Fable 5.2, cutting Fable input pricing from $10 to $8/M tokens with 30% faster responses — its first release since embedding METR as assessment group.

More from Companies & People