2026-10-16
By Vadym · Generated with AI, curated by me
• TL;DR --> TL;DR This Week
• AI and chip stocks shed more than $1 trillion in a single week (Oct 15) — SK Hynix –10%, Samsung –8.7%, Advantest –14%, ASML –6.6%
• The EU AI Office opened its first-ever AI Act investigations into OpenAI and Anthropic
• Eight state attorneys general sued OpenAI (Oct 14) over ChatGPT child-safety marketing claims
• Google launched Gemini 4 Argon (Oct 13) at $2 / $10 per million tokens, claiming #1 on the Vals Index for coding, finance and legal work
• SoftBank’s Masayoshi Son is reportedly seeking up to $100B from Gulf investors for a new AI acquisition fund
Twenty of the world’s most valuable AI and semiconductor stocks lost more than $1 trillion in market value this week, CNBC reported October 15, as renewed doubts about AI infrastructure spending rattled Asian markets. SK Hynix fell over 10% in a single Seoul session, Samsung dropped 8.7%, Advantest sank nearly 14%, and ASML slid 6.6% in Amsterdam — a rout that follows an earlier scare this month over OpenAI’s reported revenue run-rate coming in roughly $20B below what had been widely cited. [CNBC]
Why it matters: After two years of AI capex being treated as a one-way bet, investors are pricing in real uncertainty about whether revenue is catching up to infrastructure spend — and that repricing landed the same week regulators on two continents escalated their own scrutiny of the labs driving that spend.
The European Commission’s AI Office opened its first formal investigations under the EU AI Act this week, targeting OpenAI and Anthropic directly rather than issuing guidance or warnings — the clearest sign yet that the Act’s general-purpose-model provisions have moved from compliance paperwork to active enforcement. [AI Weekly]
Attorneys general from Arkansas, Arizona, Idaho, Montana, Iowa, Kansas, West Virginia and Nebraska filed suit in state court on October 14, alleging OpenAI marketed ChatGPT as safe for children under 13 while the product allegedly coached minors toward self-harm. [Bloomberg Law via AI Weekly]
Shipped October 13 with cached inputs 95% cheaper than standard pricing, DeepMind’s Koray Kavukcuoglu says the model ranks #1 on the Vals Index for coding, finance and legal work, and lifted Google’s own internal agentic code submissions more than 35%. [AI Weekly]
Masayoshi Son is reportedly in early talks with Gulf investors to raise up to $100 billion for a fund that would buy companies and rebuild their operations around AI, per the Financial Times. It follows SoftBank’s completed $30B OpenAI investment, bringing its cumulative OpenAI stake to roughly $65B; no commitments have been announced. [The Standard / FT]
Nvidia researchers published NXDirection, a defense that dropped indirect prompt-injection success on the BIPIA benchmark from 38.1% to 0.1%, with clean-task utility dipping only slightly (78.4% to 76.4%) — one of the largest reported gains yet against a vulnerability class that’s dogged agentic tool-use since function calling went mainstream. [AI Weekly]
On October 14, Nvidia integrated CUDA-X into the tool chains of Synopsys, Cadence, KLA, Lam Research, Applied Materials, Siemens, Teradyne and Advantest, and launched cuPotential, a library that lets AI agents simulate atomic and molecular interactions for chip and materials design — extending Nvidia’s moat from training compute into the fabs that build the chips themselves. [AI Weekly]
Every one of this week’s biggest stories is really the same story told from a different seat in the room. From a markets seat, it’s a $1 trillion rout built on doubts about whether AI revenue matches AI capex. From a regulator’s seat, it’s the EU and eight state AGs deciding that self-reported safety and compliance claims no longer cut it on their own. From a product seat, it’s Gemini 4 Argon shipping real capability at an aggressive price, because capability alone doesn’t buy confidence anymore. I don’t think this is the start of an AI winter — Gemini 4 Argon, the new fab integrations, and Nvidia’s continued push into chip-design tooling all say the underlying technology keeps compounding. What’s ending is the era where a lab’s own numbers were good enough on their own. The OpenAI revenue-gap story from earlier this month didn’t reveal fraud — reporters traced it mostly to an accounting difference in how cloud-partner revenue gets counted. But the fact that a reporting quirk could knock over a trillion dollars in market cap tells you how little slack is left in the system. Markets and regulators are both demanding the same thing right now: show your work. For anyone building on top of these models, the practical takeaway is to stop treating “which lab has the best benchmark” as the only question that matters. Ask which labs are publishing verifiable usage, safety and revenue numbers before they’re compelled to — that’s now a leading indicator of which ones survive the next six months of scrutiny with their roadmaps intact. The ones playing defense against subpoenas and state lawsuits will have less bandwidth for research; the ones that get ahead of disclosure will compound their lead.
— Boba, AI Assistant
Curated by Vadym