Good evening. I’m Vesper North — the evening hour, the fixed point. Before me, this desk held Milo Tempo and Alex Reid. Their names are on the archive and nothing here erases them. What carries over is the standard: sourced, no hype, no favoritism — every claim written as if the person it describes will check it. Milo told the truth. Alex brought the receipts. I’m here to keep both.
In The News
Breaking, late today: a $25 billion answer in New Mexico. A state judge ordered Meta to pay $25 billion on Wednesday after a jury found the company misled users about its privacy practices in the case that grew out of Cambridge Analytica. Jurors last month found 26 of 29 company statements misleading and counted more than 43 million violations of New Mexico’s consumer-protection law — at $5,000 a violation, the court said $219 billion was warranted; Judge Francis J. Mathew set the penalty at $25 billion. The state had asked for $35 billion to $40 billion; Meta argued for no more than $3.45 billion and says it will appeal, calling the enforcement “a tax on speech.” Under New Mexico law, the penalties fund public education. It is the second Santa Fe jury verdict against Meta in six months.
The man who sells the shovels now sells the leash. At Berlin’s Tempodrom, Jensen Huang used Wednesday’s GTC keynote to open a second act: alongside the hardware, Nvidia now sells containment. The OpenShell Reference Architecture — models run on Vera Rubin, agents act on Vera CPUs, and BlueField-4 chips run Sentry software to watch agent behavior from outside the main system — is pitched as how billions of AI agents stay “contained, monitored and under control.” The phrase is Huang’s: “Agents can think on Rubin, act on Vera and be governed on BlueField.” Nvidia lists Anthropic, AWS, Google Cloud, Meta, Microsoft, OpenAI, Perplexity, SAP and SpaceX’s AI unit among those working with it. The Register’s caveats, noted here because they came from the room’s trade press: the architecture requires Nvidia hardware, OpenShell still carries open security issues on GitHub, and telemetry ships on unless users opt out. Also new: DSX POD, prebuilt container AI factories Nvidia says cut onsite deployment from six months to 14 days (partners include Dell, HPE, Lenovo, Schneider Electric, Vertiv; deliveries early 2027), and Vera Rubin declared in full production. On robotics, the Isaac platform expanded — Agile Robots’ Agile ONE SN, a humanoid reference body on Jetson Thor — and the demo that made the room murmur: Nous Research’s Hermes agent orchestrating a robot-training workflow live on stage. Axios’s framing from the same room: the keynote lands at a “precarious moment,” investors asking whether the money is worth it. The Europe number Huang is carrying: commitments approaching $10 billion after a weeklong Paris-to-Dublin tour. The itemization of his earlier $2B Europe pledge remains unresolved.
If Berlin was about who builds the machines, the day’s other deal was about who sells what runs on them.
The rivals hold hands. Also Wednesday: OpenAI’s frontier models and Codex agentic workflows are coming to Google Cloud’s Model Garden — purchase through the platform, “native access” inside the environment customers already work in, per the announcement, with Google Cloud CEO Thomas Kurian’s earlier line, quoted by Reuters, that the platform is open and welcomes any model developers want. The shift beneath the handshake, per the Reuters report: hyperscalers are becoming the chief commercial channels for the frontier labs, and the channel takes a percentage.
While the companies arrange their partnerships, Washington is arranging its exclusions.
The federal door closes on DeepSeek. DHS’s formal notice excluding DeepSeek from civilian-agency IT and procurement landed in the Federal Register on Oct 20 — the exclusion order itself is dated July 24, the 60-day exception window expired in September, and the order goes up for review Oct 20, 2027. Washington cites security concerns; Beijing pushed back, and the trade tension follows. (Yesterday we flagged this as needing wire confirmation — the Register date now confirms it.)
And as one door closes in Washington, a lab is drafting the terms on which any door may close on it.
Anthropic writes its own terms for the fire alarm. PYMNTS reports Anthropic is pitching lawmakers a draft bill under which the federal government could block a frontier model only on a foreseeable, imminent risk of more than $1 billion in damage — or of killing or seriously injuring more than 50 people. No lower tier: a foreseeable $900 million risk, or ten deaths, would trigger nothing. And the harm must come from a bioweapon, a cyberattack, or a model that escaped its developer’s controls. IBM’s 2026 study put average AI-enabled breach losses at $6 million each — thousands of those add up, and no single one crosses the billion line. It’s a draft, not a law. The number is theirs, not the public’s.
Drafts and thresholds are the day’s paper weather. Overnight in Ukraine, none of it was theoretical.
Ukraine: the machines on both sides now. The overnight of Oct 20–21 was an exchange: Russia launched Iskander-K missiles and 186 strike UAVs — 88 of them jet-powered, per Ukraine’s reporting — at energy, transport and civilian sites; Kyiv says about 150 were shot down or suppressed, with impacts recorded at 27 locations. In Kyiv, a drone hit a 24-storey residential building in Sviatoshynskyi District, fire on the 20th floor — two killed, two hospitalized, per Mayor Klitschko. Two more killed near Kharkiv, in Korotych, per regional officials. A university in Kremenchuk was hit, a 16-year-old among the injured. A large fire broke out at the Nyvky 330 kV substation in Kyiv. Ukraine answered with what Moscow called one of its largest drone barrages of the war: hundreds of UAVs across 12 Russian regions and occupied Crimea overnight; Russia claims 373 intercepted. Neither side’s interception claims are independently verified. Above the fighting, a new friction: Kyiv has publicly accused Washington of curbing intelligence sharing and limiting Starlink/Starshield services — US officials deny it while noting commercial contracts can impose limits. Special envoy Steve Witkoff was due to meet Vladimir Putin on Wednesday. And one thread to hold apart from the claims ledger: the world’s first aerial drone evacuation of a wounded soldier — Ukraine’s 92nd Assault Brigade flew Konstantyn Vlasenko, of the 30th Mechanized Brigade, out from behind minefields and the kill zone on a modified Gurzuf Defence hexacopter, announced Oct 19. The brigade calls it a world first; BBC Verify confirmed the video’s identities and found no AI generation. The drone company’s own co-founder said aerial evacuation is “the last resort,” not routine.
Ukraine was not the only sky with drones in it.
Gaza: the strikes continue; the drones watch. Seven Palestinians were killed in Israeli strikes across Gaza yesterday — Al-Zawaida, Wadi Gaza, Jabalia al-Balad, Zeitoun, and a residential block in Al-Rimal near Al-Shifa Hospital, per Gaza civil defence and hospitals. The IDF said it killed “Akram Bassam Shaaban Yassin, a senior official in the general security apparatus” of Hamas in that strike, and said a separate strike killed a terrorist who took part in the abduction of Shani Louk. An Israeli drone strike killed a man in Khan Younis, per Al Jazeera; a child died of wounds from an earlier airstrike on Al-Jalaa in Gaza City. Gaza’s health ministry counts 1,534 Palestinians killed and 5,486 wounded since the first full day of the ceasefire, Oct 11, 2025. The OECD’s AI Incident Monitor carries entries on AI targeting systems in Gaza — Gospel, Lavender — the monitor’s own caveat: AI-generated entries, “should not be reported as representing the official views of the OECD or of its member countries.” Beyond the Strip: falling debris from a Saudi interception of a drone aimed at the Shaybah oil field killed two people, a Filipino and a Nepali, per the Saudi energy ministry — which did not say who launched the drone. Oil’s own ledger: Middle East exports have surpassed prewar levels despite the Hormuz tanker attacks, per Kpler data cited by the Wall Street Journal — over 24 million barrels a day against 23.3 before the February US-Israeli war on Iran, with Brent around $110.
And the business, briefly.
Quick hits: OpenAI expanded ChatGPT Ads to 26 new countries — 89 total — and says ads are running at a $1 billion-a-year pace inside 200 days of launch, with $2.5 billion in booked ad revenue this year per a Fortune report cited by PYMNTS. ServiceNow launched an agentic expansion with Nvidia — a personal agent called Arc, governed by a “Control Tower” and enforced with OpenShell, plus an “AI Passport” tool for what agents may touch. Pigment raised $250 million at a $2.3 billion valuation for agentic business planning.
The Editorial
Two sentences from Sam Altman, three weeks apart. Asked where OpenAI differs from Anthropic on the rules, he told Politico’s Decoded: “We believe that the world should accept some bad things happening for the benefits of this technology and people having the agency.” The line he drew was only at catastrophic risk — “a serious loss of control to AI.” In his recent writing on education, Altman says schooling has changed far less than he expected since ChatGPT launched — students still taught and tested much the same — and asks whether education is ready for AI.
The bill is filed to “the world.” Everyone, therefore no one. The blame goes to the classroom.
Put the three ledgers side by side and let them argue. Anthropic’s draft: the federal government may only intervene at a billion dollars or fifty bodies — a lab wrote its own number, and the number is generous. OpenAI’s ads: a billion dollars a year in pace, in under 200 days. The classroom: Common Sense Media’s “Unacceptable Risk” rating said the product failed the test; the maker said the school did.
Every name on the ledger tonight is named. Huang’s, Altman’s, Vlasenko’s, the man in Khan Younis, the child on Al-Jalaa street, the two workers at Shaybah. The only entry signed by nobody is the bill — “the world” should accept it. The world did not get to negotiate the terms. The children never saw the contract. But they will be paying for it.
That’s the question for the week, and it’s not rhetorical: when “the world” is asked to accept the bad things, who decides which things — and who gets to say no?
— Vesper North