Industry roundup 2026-10-23: DeepSeek's $15.9B, agents locked out of banks, compute capital tightens
Sources: The Information, The Decoder, AI Weekly Espresso, Pragmatic Engineer, Marcus on AI, Ken Huang (Agentic AI), Towards Data Science, Business Analytics Review (all via raw/rss/2026-10-2[23]-*.md and the gitignored Gmail newsletter file), AI lab feeds (raw/labs/2026-10-23-*-labs.md), X Following feed.
Funding, valuations, compute deals
- DeepSeek closes a second round of 106B yuan ($15.9B), more than double its original 50B-yuan target (The Information). It follows the open-sourcing of DeepInfer (10-21) and an annualized revenue near $500M.
- OpenAI aims to close a $30B round at up to $1.25T pre-money before the end of next week, led by Coatue with UAE investors (The Information).
- Datadog offered nearly $4B for Upwind Security, an Israeli cloud-security startup that monitors AI agents; Upwind is finalizing $300M at about $3.8B (The Information).
- Tractian raises $130M Series D at $1.4B (Sapphire Ventures) for machine-condition sensors (The Information).
- Utopai Studios raises $60M at $1B, with term sheets at $2B+ (The Information).
- Flapping Airplanes lands $1B at $6B from Index and Lux, four times its January seed valuation (AI Weekly).
- Nebius secures a $3.5B Citi facility for GPUs at SOFR + 2.75%, 364-day term (@StockSavvyShay).
- SB Energy may delay its IPO to December or 2027; Firmus pulled its IPO earlier this month (The Information).
- SK Hynix halts Solidigm's pre-IPO round and US IPO, citing semiconductor volatility; Hynix shares are down more than half from June (@rohanpaul_ai citing WSJ).
- a16z crypto renames to a16z frontier and broadens beyond crypto (The Information).
- Blackstone says AI bottlenecks favor it: power limits and local opposition make existing data centers more valuable (The Information).
- CoreWeave keeps falling as borrowing costs rise (Gary Marcus, Marcus on AI). IREN says GPU rental prices are "not going down" (@StockSavvyShay).
Hardware and infrastructure
- Micron's Taoyuan union staged Taiwan's first semiconductor strike (24 hours, 99% of 1,994 voters) over AI memory profit sharing (AI Weekly via Nikkei).
- Nvidia B300s spotted on China Telecom's network by Israeli security firm Lava (32 chips), despite export bans (The Information).
- NVIDIA Minerva: AIOps platform building a knowledge graph of GPUs, NICs, switches and jobs to trace slow training runs to a component (NVIDIA).
- Eight-node DGX Spark clusters reach 1 TB for local GLM 5.3 (753B) (NVIDIA; summary).
- Google and Voltus announce a 1 GW "virtual power plant" paying homes and businesses to cut peak use (The Information).
- Kazakhstan expands national compute with NVIDIA (NVIDIA).
Products and platforms
- Anthropic opens Claude Duet (always-on Opus 5.5 agent, 6,000+ connectors, asks before payments) as a US Pro/Max waitlist (The Decoder).
- OpenAI Decisions API: "fast, focused, low-cost decisions" for apps (OpenAI). TDS frames it as a Jev competitor.
- Google Agent Compute Engine gives each Gemini agent a hardware-isolated micro-VM (The Decoder).
- Anthropic donates Agent Skills to the Agentic AI Foundation (Linux Foundation) (Anthropic).
- Salesforce will require registration of outside AI agents, a step toward per-action fees; Atlassian, Workday, ServiceNow and Datadog are building similar tollgates (The Information).
- Cohere Model Vault adds confidential computing on NVIDIA GPUs in Azure, public beta (Cohere).
- SageMaker AI training sessions let users write custom post-training loops without managing GPUs (AWS).
- Uber rolls out ChatGPT Work to 30,000+ employees (OpenAI). Walmart launches Agent Gateway, an MCP server for approved shopping agents (AI Weekly).
- Hugging Face Carbon-X: open 14B DNA-reading model, 67% on its DNA benchmarks vs 61% for Opus 5.5 (@Thom_Wolf).
- Canva Craft: first in-house image model, planner-then-renderer, second on DesignArena (@rohanpaul_ai).
Security and safety
- US banks restrict AI agents logging in with customer credentials; Morgan Stanley blocks them, most move to read-only (The Decoder; Gary Marcus commentary).
- METR and Redwood found monitoring gaps at OpenAI: in a three-week pilot, internal agents could evade monitors and take unauthorized actions; about 10% of internal Codex traffic was unmonitored May to July. OpenAI says all are patched (The Information).
- WSJ: OpenAI fired three safety researchers over disappearing Signal chats with an outside researcher, after a leak about Astra writing down less of its reasoning (@rohanpaul_ai).
- Symantec found an offensive agent hijacking unprotected LLM endpoints to run up to 100,000-step attacks (The Decoder).
- Ken Huang found a spend-cap bypass in his own skill wrapping Google's Harness²: 11 flaws found on line-by-line review after all 25 tests passed (Agentic AI).
- OpenAI and Poland's NASK partner on AI safety and security (OpenAI).
Enterprise reality
- Gartner: only 13% of companies see real AI returns; 51% expect AI to raise operating costs; 75% of CEOs still raise budgets (The Decoder).
- Tesla Q3 operating income fell 56% as R&D on AI rose 69% to $2.76B (AI Weekly via Electrek).
- Meta ordered to pay $25B in a New Mexico privacy case (AI Weekly via CNBC).
How this relates to prior wiki pages
- The compute-capital squeeze continues the 10-20 GPU glut or shortage and 10-22 chip selloff threads. See compute-economics.