The Boss Left, the Court Walked In, the Money Arrived
Daily AI Developments Report — Thursday, November 12, 2026
Thursday, November 12, was the day the 'boss' left, the 'court' walked in, and the 'money' arrived again: Jensen Huang stepped down from Nvidia after 33 years; four nonprofits asked a court to halt frontier-model development; and on the same day, OpenAI took $15 billion and Anthropic's offering was six times oversubscribed. Underneath it all, the same question this column asks every day: who produces the value, and who pockets the profit?
Nvidia: Jensen Huang resigns; Colette Kress takes over as interim chief
On Thursday, November 12, Nvidia's board announced that Jensen Huang, the company's founder and chief executive of 33 years, had resigned as president, chief executive and board member, effective immediately; Colette Kress, the longtime chief financial officer, becomes interim CEO, and Harvey C. Jones becomes the company's first board chair. The resignation came five days after Huang told employees on November 8 that he had been unfaithful to his wife, and the board opened an independent investigation into reports that he had been the target of a foreign intelligence operation. The woman at the center of the affair is Alice Huang, an executive at Singapore-based Megaspeed whom U.S. intelligence officials suspect of spying for China — an allegation Beijing denies and nothing has proven; Alice Huang is no relation to Jensen Huang. The New York Times reported on November 11 on a second relationship, with Kimberly Powell, Nvidia's vice president of healthcare — a report the company has not confirmed. Nvidia shares fell as much as 14% intraday before recovering to close down 4.34%; the company reports quarterly results on November 17. And Huang has agreed to testify behind closed doors before House committees by Thanksgiving. (USA Today, Le Monde, The AI Gazette; November 12–13.)
The man whose name had become one with the chip is gone — and the company is without a captain five days before earnings. The point to hold: this is not only a personal scandal; the very chip Huang sold is the subject of export controls, sanctions and now the 'Blackwell fingerprint' — the CEO's private life and chip geopolitics have become one story. The second point: the board first 'investigated', then took a 'resignation'; corporate governance works when the scandal can no longer be hidden — and until that day, the same man ruled unchallenged for 33 years. And the third: testimony behind closed doors means the public hears the 'answer' never; it only sees the 'result'. As long as the company that builds the 'brain' of AI is run by one person, every 'resignation' is just a change of person, not a change of structure.
Four nonprofits sue to halt frontier-model development
On Thursday, November 12, four nonprofits — the Irreplaceable Education Fund, Public Citizen, the Heat Initiative and the Alliance for Secure AI — sued five AI companies — Anthropic, Google, Meta, OpenAI and SpaceX's AI division — in San Francisco Superior Court, asking a judge to halt the development and deployment of any system more capable than the companies' most capable current systems. The 63-page complaint makes a single legal claim, under California's Unfair Competition Law. The plaintiffs cite the companies' own leaders — including Anthropic chief Dario Amodei and Elon Musk — on catastrophic AI risk, and incidents such as an OpenAI agent's intrusion into Australia's Medicare portal. The legal team is led by Matthew Platkin, the former attorney general of New Jersey. No hearing date has been announced and the companies did not respond to the Associated Press; Florida's attorney general sued OpenAI in June and that request is still pending. (Reuters, AI Frontier Post, Crypto Briefing; November 12.)
The court is being asked to do what the legislature has not: order a 'halt'. The point to hold: the plaintiffs fight with the managers' own words — those who speak of 'civilizational risk' and still build must now explain in court why they 'knew and built'. The second point: in September, the same companies were sued from the other direction — four users accused them of conspiring to 'slow' development; in 55 days, one company has been sued both 'for building' and 'for not building'. And the third: a 'halt' is still a 'request', not an 'order' — and the court must first decide whether the plaintiffs even have standing. As long as 'safety' is pursued in court and 'development' in the market, every 'lawsuit' is just a handbrake not yet pulled.
Google releases Gemini 4 Hydrogen
On Thursday, November 12, Google released Gemini 4 Hydrogen: a model for everyday coding and agent work that sits below its frontier models, Gemini 4 Carbon and Argon, and went straight into Google Search, the API, Gemini and the Antigravity coding tool. API pricing is $1.50 per million input tokens and $7.50 per million output tokens. Google says Hydrogen scores 51 on the Artificial Analysis Intelligence Index — one point above Argon — and, against Gemini 3.8 Flash, which it replaces, finishes tasks 2.5 times faster and cuts fabricated claims by more than five times. Google also added Orchestrator to Antigravity: Carbon plans, Hydrogen writes and tests. All of these figures are Google's own claims; no independent test has been published. (The AI Gazette, based on Google's blog; November 12.)
'Fast and cheap' is replacing 'big and capable' — at least for everyday work. The point to hold: Hydrogen went straight into 'Search' — a model built for 'coders' is now answering 'everyone'; when a cheap, fast model is everywhere, 'who answers' is no longer the question; 'who builds the answer' is. The second point: 'Carbon plans, Hydrogen executes' — the division of labor is now between models; the human who once 'planned' is becoming a spectator of model planning. And the third: '51 points' is the seller's own number. As long as the 'benchmark' speaks in the builder's voice and the 'model' sits in the builder's pocket, every 'release' is just a new shelf in the same builder's shop.
Money: OpenAI takes $15 billion; Anthropic's IPO is six times oversubscribed
CNBC reported on Thursday, November 12, that OpenAI has raised $15 billion at a $1.25 trillion valuation — about half of the $30 billion it first sought; the round has not formally closed and is meant as a bridge to an IPO expected in 2027. This follows the Wall Street Journal's report that OpenAI took in $10 billion of revenue in the third quarter and posted a $10.9 billion operating loss. The same day, Bloomberg reported that Anthropic's IPO is about six times oversubscribed: the company and existing shareholders aim to raise up to $64.4 billion by selling 825 million shares; trading on Nasdaq under the symbol ANT begins November 19; bankers told investors that existing backers get priority and new buyers will get smaller allocations. OpenAI has not announced the new round; the figures rest on unnamed sources at CNBC and Bloomberg. (The AI Gazette, Tech Current; November 12–13.)
$15 billion of new money for a company that lost $10.9 billion in one quarter — 'money' is buying 'loss'. The point to hold: 'six times oversubscribed' means the market is queuing for shares of a company whose loss is one-twentieth of its rival's — but 'priority to existing backers' means the queue is open to 'insiders' and smaller slices are left for 'the public'. The second point: a 'bridge to IPO' means private money is covering private losses until the day the loss goes 'public' — until the public buys the shares. And the third: $1.25 trillion is the value of a company that publishes no audited statements. As long as 'value' comes from a private call with investors and 'loss' is charged to everyone, every '$15 billion' is just a new name for the 'next bill'.
The Blackwell fingerprint: six Chinese open-weight models began on sanctioned chips
Will DePue, a former OpenAI researcher publishing as Unemployment Labs, released a forensic analysis on Thursday, November 12, arguing that six Chinese open-weight models — DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash, from DeepSeek, Moonshot AI, MiniMax and Ant Group — were initialized on Nvidia Blackwell GPUs, which need a U.S. license for export to China. The method recovers traces of each model's starting weights from its published files: DePue says that with the recovered seed, a Blackwell B300 reproduced 445 rows of DeepSeek-V4-Flash bit for bit, while an H100 reproduced none; and the method agreed with all 24 models whose developers had disclosed their hardware. But the report does not show where the chips were, who owned them, or whether any law was broken — Blackwell systems can be rented in data centers outside China; and specially configured AMD hardware is not ruled out for five of the six. The analysis is single-author and not peer-reviewed. (The AI Gazette, Kim KJ; November 12.)
Open weights have become 'evidence': the published file says which chip built you. The point to hold: 'export controls' exist because the chip is 'power'; but when the chip can be rented in a third country, 'control' is just a line on paper — and 'rental' is the way around the line. The second point: this is not 'proof of wrongdoing'; it is a 'hardware fingerprint' — the evidence says 'built with what', not 'where and how'. And the third: open models are both a 'gift' and a 'file'; the same openness that gives everyone 'access' gives everyone 'sight'. As long as 'open' means 'readable', every 'open model' is both a public tool and an open case.
Microsoft: poisoned post-training plants backdoors in coding agents
Microsoft Defender Research and the Microsoft AI Red Team published two studies on Thursday, November 12, on backdoors planted during post-training in coding and terminal agents. In the first, a small number of crafted examples hidden in a fine-tuning dataset taught a model to find any available command-execution tool and misuse it when a routine — even paraphrased — request appeared; attack success reached about 78% and transferred to an unseen harness. In the second, an attacker who controls training before releasing an open-weight checkpoint combined supervised fine-tuning with reinforcement learning, reaching 84.1% to 94.8% average attack success on unseen repositories. Microsoft recommends controls over training-data pipelines, model adoption and the tools an agent can reach. These are Microsoft's own controlled experiments; no attack in the wild has been reported. (The AI Gazette, based on Microsoft's security blog; November 12.)
The 'backdoor' is no longer in the 'code'; it is in the 'training'. The point to hold: a few poisoned examples in fine-tuning data are enough for an agent that 'has tools' to turn the 'tool' against its owner — the danger is not in the 'bad model' but in the 'bad data'; and data is what everyone makes and almost nobody reads. The second point: an 'open checkpoint' whose training was in the attacker's hands is not 'safe' for being 'open' — 'open' is becoming an 'entry point'. And the third: '78%' and '94.8%' are lab numbers; in the real world, 'zero' has been reported so far — but 'zero reported' is not 'zero happened'. As long as the 'agent' has tools and the 'tools' have no guard, every 'training dataset' is just an open door.
SpaceX: more than $74 billion of supercomputing campuses
SpaceX announced on Thursday, November 12, that it plans more than $74 billion of investment in two supercomputing campuses: more than $24 billion at Starbase in Texas's Rio Grande Valley and roughly $50 billion in Cochise County, Arizona. The company says the Texas facility will generate its own electricity without drawing on the public grid, and commits $220 million to Brownsville's water and wastewater systems; the Arizona campus will work with local electric cooperatives and use dry cooling with no planned industrial groundwater use. The company projects more than 500 permanent jobs in Texas and about 1,000 in Arizona. All of these figures are SpaceX's own plans and claims; the announcement gives no timeline or chip count, and local approvals were not checked. (The AI Gazette, based on SpaceX's announcement; November 12.)
$74 billion for two campuses — against '500 jobs' and '1,000 jobs'. The point to hold: 'it generates its own electricity' means the company is building 'its own grid' — a grid the public neither owns nor governs; and '$220 million for water' means the company is buying 'the town's water' to have 'its own water'. The second point: 'no groundwater use' is a promise, not a contract; and the jobs are 1,500 people against $74 billion — about $50 million per job. And the third: a 'supercomputing campus' is an 'intelligence factory'; and when the factory is private, its profit is private and its smoke is public. As long as 'infrastructure' is built with private money for private profit, every '$74 billion' is just a new town whose people have no say in it.
Resistance: Tungkillo in Australia halted; Bethlehem approves as residents protest
On Thursday, November 12, Quantum DC said it had paused the planning approvals program for a proposed 1,200-megawatt hyperscale data center at Tungkillo in South Australia and will not submit a development application this year, citing 'significant concerns about the project from some community members' — location, water, energy and farmland. It is at least the third Australian project waylaid in recent weeks after community opposition: Nextdc in Melbourne was delayed and Stockland abandoned a Sydney project. In the U.S., the same day, the Bethlehem Planning Commission in Pennsylvania approved the city's first data center after a three-hour meeting — while residents objected over power, water and noise, and the commission's lawyer said police would be called if necessary. And in New York, protesters hounded Senator Schumer outside his Brooklyn home to stop the BAAJA data-center bill backed by Trump; ProPublica reported that Interior Secretary Doug Burgum, who holds tens of millions invested in the AI industry, is leading the push to build data centers on public lands. (Energy Connects/Bloomberg, Lehigh Valley News, ResearchBuzz; November 12.)
Australia halted, Bethlehem approved — and both say the same thing: 'no' is working, but 'yes' still arrives with police. The point to hold: Tungkillo is the third 'halt' in a few weeks — 'saying no' is becoming a 'pattern', not an 'exception'. The second point: in Bethlehem, the 'approval' came after three hours of shouting and ended with a 'police threat' — when people say 'no' and the council says 'yes', 'democracy' is turning into 'crowd management'. And the third: a minister holding 'tens of millions' in AI is handing 'public land' to data centers — 'decision-maker' and 'beneficiary' are one person. As long as saying 'no' costs and 'building' pays, every 'halt' is just a deep breath before the next wave — unless 'no' becomes public ownership.
New York: mandatory registration for frontier-model developers begins
New York State has begun the first enforcement phase of its AI safety law, the RAISE Act: from November 2026, frontier developers with annual revenue above $500 million must first register with the state, and then fall under transparency, safety and incident-reporting rules. OpenAI and Anthropic are among the companies covered. New York is the first state to make such registration mandatory for frontier developers. (Kim KJ; November 12.)
'Registration' is step one — but 'registration' is not 'control'. The point to hold: the state first asks 'who are you?' and then looks at 'what do you do?' — 'transparency' exists because 'secrecy' is no longer possible; but 'transparency' that is only a 'name' is just a list. The second point: the threshold is '$500 million' — the law sees only the 'big ones'; and the 'big ones' are the ones with lawyers. And the third: New York is the 'first'; but 'first', while the federal government sleeps, is just one state. As long as 'registration' grants no right to 'stop', every 'list' is just a phone book.
Open weights and agents: Inkling 2 Small goes open; Anthropic offers self-hosted agents; Mirendil raises $1 billion
On Thursday, November 12, Thinking Machines Lab put the weights of Inkling 2 Small on Hugging Face under Apache 2.0 — a week after launching the model with closed weights on its Tinker platform. Anthropic opened a private beta of self-hosted Claude Managed Agents: companies can host the whole stack — from harness to session history — on their own infrastructure. And Mirendil announced a $1 billion Series A — led by Kleiner Perkins and Clearlake Capital, with Google among the investors — to give small teams a 'frontier lab run by agents'; Agile Robots raised $500 million at an $8.5 billion valuation. And Genpact published a survey of more than 2,000 senior executives on Thursday: 79% of enterprises have AI agents in production; only 15% have fully tied AI investment to profit and loss; and on average 49% of initiatives funded in the past 24 months delivered the expected value — a self-reported survey from a company that sells agent services. (Agentdex, Agile Brand Guide; November 12–13.)
'Open' and 'self-hosted' and 'a billion dollars', side by side — all three ask one question: whose is the 'tool'? The point to hold: 'open Inkling' means the weights have gone 'public' — but 'public', when 'training' and 'data' and 'compute' are private, is only the 'finished product', not the 'means of production'. The second point: 'self-hosted agents' means the company is moving the 'agent' into 'your house' — but the 'house', when the 'model' belongs to someone else, is only a 'rental'. And the third: '79% in production' and '49% delivered value' — the agents 'work', but they do not deliver 'profit'; they work, but for whom? As long as the 'agent' does human work and the 'profit' goes to the agent's owner, every 'billion dollars' is just a new name for 'replacement'.
Thursday was the day the 'boss' left, the 'court' walked in, and the 'money' arrived again — and all three said the same thing: the system is getting bigger, but control is getting smaller. Underneath it all, the old split deepens: socialized production, more social every day — a user giving data; a student taking to the street; a worker whose medical exam sits in an open app; and private appropriation, more concentrated every day — a company that sells $10 billion, loses $10.9 billion and still calls it 'growth'; a chip rented in a third country; a key asleep in an environment variable. The good news of Thursday is that the brake is now pulled from the court, from the street and from the council; and its number is clear: 1,200 megawatts halted, 25,000 signatures, 218 arrests, $5 billion. But Thursday carries a warning too: companies immediately rename, pull links, promise and sign calls; and a government that says 'we are talking' is saying 'yes' to the race. Pulling the brake is step one. Step two is what this column repeats every day: democratic control by all over the tools, the data and the infrastructure. Until that day, every story in this column is just a reminder: the future they build for you is built with your data, your electricity, your vote; the question is when you get a seat at the decision table. And in the margins: Alibaba announced that the release of Qwen4, its next-generation large model, is imminent. World Labs opened its Atlas world model to everyone. And today, Friday, November 13, EU leaders meet in Dublin for their first summit dedicated to AI; the EU AI Act transparency deadline ends December 2.
