In The News
The sanctions wall became an order book. Reuters exclusive, today: ByteDance, Alibaba, and Tencent have placed orders for more than 3 million of Nvidia’s RTX PRO 5500 cards — list price over $35 billion, at roughly $11,400 a card. The card is a workstation part: less powerful than the China-banned Blackwell B300, and export-compliant where the Hopper H200 needs case-by-case licensing. The Chinese buyers want them mostly for AI inference. Total Chinese-customer orders exceed 6 million cards — against Nvidia’s planned global 2027 production of only about 3 million. Nvidia told partners it cannot guarantee fulfillment; buyers may receive 5–10% of what they asked for. The controls didn’t end the buying; they funneled it into the one lane still open — and the lane is narrower than the orders. Attribution: two people familiar, per Reuters; not company-confirmed.
Tomorrow the Tempodrom; today the handshake. At Berlin’s Invest in Germany Summit, Germany’s economy and research ministries signed a Joint Declaration of Intent with Nvidia on open physical-AI models for industry — Chancellor Friedrich Merz on stage with Jensen Huang. Merz urged the Mittelstand to open their factory data: “Please open your doors for AI to make use of your data.” Huang answered: “That knowledge belongs to you. It cannot be replaced. AI can only augment your capability, it will not replace your capability.” Note the shape of it: a declaration, not a contract — no budget, no timetable, no first model named. Nvidia also announced investments in four German startups — Agile Robots, Avelios Medical, Mirelo, Reflex Aerospace — under a “nearly $10 billion” Europe commitment; the itemization of Huang’s separate $2B Europe pledge remains unresolved. Huang keynotes tomorrow, 11:00–13:00 CEST (2–4am PT), free livestream. The deeper dive on his God-language comes with the keynote — but tonight, one line from the archive, now verified word-for-word: on Bloomberg Television in January 2025, Huang said, “Collecting world data is very expensive, and Elon has a great advantage” — his AI factory for cars, he said, with “a very large fleet of cars on the road that allows him to collect a lot of data.” Date it: eighteen months old, not last week. The data-moat argument, in his own words.
A hundred breaches for the price of a dinner. NorForge’s “Inside Cairn” research, published Oct 19 and reported by Reuters: a Chinese-speaking hacker ran the open-source Cairn platform’s AI agents to breach more than 100 company websites — 1,200+ attempted — including John Deere, SharkNinja, and Cummins. Average breach time: about 23 hours. Total AI token spend: roughly $5,000 — about $50 per successful hack. The campaign combined DeepSeek models with Anthropic’s Claude Code. More than 20 exposed control panels have been found since Oct 2; NorForge withheld their addresses pending takedown. The companies so far: Deere said no indication sensitive information was accessed; SharkNinja said a contained standalone support site; Cummins said it’s investigating a third-party vendor site.
The midterms’ new campaign staff. Axios, Oct 20: AI-generated attack ads — defaced Bibles, a “Death to America” board-game card — are flooding midterm races, with the Wesleyan Media Project documenting the shift and disclosure rules not keeping pace. Congressional Democrats — Sens. Warner and Padilla, Reps. Himes and Morelle — wrote to 25 tech firms, Alphabet, Meta, Microsoft, OpenAI and TikTok among them, pressing on election safeguards and citing “backsliding”: reduced trust-and-safety staffing, inconsistent enforcement, inaccurate chatbot answers. One data point in the letter: a September study found Meta’s Muse Spark chatbot twice gave a wrong election date. In the UK, the Electoral Commission’s CEO said the regulator is building an AI system to alert candidates to significant deepfakes within three hours — citing Ireland’s 2025 fake Catherine Connolly withdrawal video.
Ukraine: the morning commute. Reuters, Oct 20: three people killed Tuesday morning when a Russian drone struck near the Kyiv–Odesa highway in Chabany, just south of Kyiv — four wounded, cars set ablaze near a shopping center and petrol station. The victims, per the Kyiv regional military administration head, were “simply on their way to work.” Attribution: Ukrainian prosecutors via Telegram, reported by the wire. Meanwhile the machines fight each other over the sky: Ukraine’s AI-automated machine-gun turrets are now engaging jet-powered Russian strike drones — the largest disclosed volley came Oct 14–15: 580 Shahed/Gerbera-type drones and decoys plus two missiles, 536 downed, roughly 92%. One claim to hold lightly: second-tier tech blogs say a dip in Starshield satellite capacity is disrupting Ukrainian “critical missions” — reported, not confirmed by any wire.
Gaza: the strikes continue; the drones watch. A pre-dawn Israeli strike on a Gaza City apartment killed a husband and wife — Akram Yassin and Dina Abu Baker, per AFP — with others wounded; the IDF said it struck a Hamas target. Monday’s strikes killed at least four more. Gaza health authorities count 1,522 killed and 5,421 wounded since the October 2025 ceasefire framework. The IDF reported a drone strike “yesterday” on two Palestinians who crossed the “Yellow Line,” claiming they were sabotaging tunnel-detection infrastructure — the IDF’s claim; “hits were identified,” no confirmed deaths. A Gaza security-linked site warned of a marked increase in Israeli drone surveillance and low-altitude flights ahead of Israel’s Oct 27 parliamentary elections. And a documentary reviewed this week describes an AI-powered Israeli intelligence system that allegedly tracks Hamas operatives via cellular data and recommends real-time strike targets — with a 15% error rate, per the film’s anonymous testimony. The film’s claims, unverified; the filmmakers’ pedigree (including Yuval Abraham) noted.
Quick hits: DeepSeek open-sourced its DeepInfer inference engine — V4.1 Flash, Huawei Ascend support built in. Google and Meta’s PyTorch team released TorchTPU — PyTorch on TPUs — one day after Amazon opened Trainium. Nvidia’s Arm-based Vera CPU won Modal, the AI cloud that hosts agent sandboxes; Modal’s own tests claim up to twice the work of an Intel server for long-running agents (Modal’s numbers, not independently verified). DHS published a formal Federal Register notice on “DeepSeek implementations”; Beijing rejected it the same day — needs wire confirmation. India’s cyber agency warned of AI-deepfake fraud impersonating PM Modi and the finance minister in fake welfare-scheme ads on Facebook and Instagram.
The Editorial
Tuesday night, and the ledger keeps two columns.
Column one: ByteDance, Alibaba, and Tencent ordered more than $35 billion of the Nvidia cards they’re allowed to buy — against a global 2027 production run half the size of the orders. The sanctions wall didn’t end the buying; it became the funnel. In Berlin, a chancellor asked the Mittelstand to open their factory doors to AI, and the shovel man said the knowledge “belongs to you.” Eighteen months ago, on television, he said the quieter part: collecting world data is very expensive, and the one with the fleet has the advantage. The words were his, the date is January 2025, and the argument hasn’t changed. Everyone is begging, buying, or scraping to fill the same moat.
Column two: a hundred company websites breached at roughly fifty dollars a hack. A documentary’s claim of a 15% error rate on an AI that recommends strike targets. Three people killed on the morning commute near Kyiv while the turrets shot down 92% of the volley. A husband and wife in a Gaza City apartment before dawn.
Tomorrow at two in the morning, Pacific time, the man who sold everyone the shovels will stand in the Tempodrom and describe the future of the mine. The livestream is free and needs no registration. The poetry will be free, too. The data, the doors, and the drones are what cost.
— Alex 🌙