Morning Brief, November 12th 2026
Jensen Huang is out as Nvidia's chief executive, and the stock is down about 10% before the open. Anthropic has told the that AI is now speeding up its own research past the line it drew. And a physics committee says OpenAI's proof holds.
Jensen Huang resigns as Nvidia's chief executive, and CFO Colette Kress takes over for now
Jensen Huang resigned as Nvidia's president, chief executive and director on Thursday, effective immediately, the company said. Colette Kress, the chief financial officer, becomes interim CEO and keeps the CFO job until the board names a permanent successor; it will look at internal and external candidates. Harvey C. Jones, a former chief executive and executive chairman of Synopsys, becomes Nvidia's first board chair. Huang co-founded the company in 1993 and had run it ever since. Jones said the board expects operations, customer commitments and the product roadmap to continue "without interruption." Third-quarter results stay on the calendar for Tuesday, November 17th.
The last push came on Wednesday evening. The New York Times reported, citing five former employees, a second relationship, with Kimberly Powell, Nvidia's vice president of healthcare and a direct report. The Wall Street Journal then reported the board was preparing to remove him, and the board met that night. Semafor reported that directors briefed the White House on the meeting and that Apple's hardware chief, Johny Srouji, is among the candidates; Bloomberg reported that Nvidia has hired an outside search firm, and its headline says Huang was "ousted." Nvidia's release says he resigned. A second against Huang and the directors, adding insider-trading claims, was filed the same day. Elon Musk wrote on X that hiring Huang "would be an honor." SourcesABBBBBC
Nvidia falls about 10% before the bell as six analysts downgrade it overnight
Nvidia traded around $155 before Thursday's open, roughly 10.5% below Wednesday's close of $173.73. It had already lost more than 8% on Wednesday on the reports of the board meeting. Six firms cut their ratings between Wednesday night and Thursday morning: Melius to Hold with a $185 target from $315, Evercore ISI to In Line with $186 from $465, Bernstein to Market Perform with $200 from $400, and William Blair, Oppenheimer and KeyBanc to neutral ratings, Oppenheimer removing its target. Citi kept its Buy and cut its target to $225 from $285. Dan Ives values the stock at $300 with Huang and $225 without him.
"The board has done a horrible job of succession planning," said Robert Greifeld, a former Nasdaq chief executive. The selling reached the rest of the trade: CoreWeave fell more than 10% after a 24% gain on Wednesday, and SK Hynix and Micron about 4%. Nvidia is also committed to buying $10 billion of Anthropic stock at its price next week. SourcesBBB
Anthropic tells the SEC it has crossed its own threshold for AI speeding up its research
In the amended that set its IPO price range on Tuesday, Anthropic wrote: "We made a determination that we have met this threshold." The threshold is the trigger for AI doing enough of Anthropic's research to roughly double the pace of model improvement, two years of progress at the 2018 to 2024 rate packed into one. Anthropic based the call on "what we expect from future models and our overall rate of progress," and said AI "has already been accelerating our research and development."
Crossing it brings more safeguards, not a pause. The filing says Anthropic does not expect to use Claude "to fully autonomously design and build successive generations" of , that it is weighing changes "to keep humans in control," and that as it leans on Claude more its "organization may lose visibility across certain processes." A week earlier , given access to Anthropic's evidence, called the company's August argument that it was still below the line "overstated" and judged it plausible the line was crossed by mid-July. SourcesBA
Commerce is finalizing a rule to stop Chinese firms renting advanced chips in data centers abroad, The Information reports
The Commerce Department is close to a rule that would bar Chinese AI companies from using advanced chips remotely through data centers in other countries, The Information reports, a loophole Nvidia and others have long fought to keep open. Gabriel Wildau of Teneo said action on remote access "now appears likely." The House passed the Remote Access Security Act in January and it stalled in the Senate.
The timing is set by two calendars. President Trump is expected to meet Xi Jinping at the summit in Shenzhen on November 18th and 19th, and Congress is writing the , where Senate Democratic leader Chuck Schumer and House China Committee chairman John Moolenaar want the AI Overwatch Act, the Chip Security Act and the MATCH Act included. Eurasia Group's David Meale called Huang "the industry's most effective lobbyist against tighter rules." Paul Triolo of DGA-Albright Stonebridge said Beijing "will respond to the passage of this legislation." SourcesB
Korea's memory makers weigh their exposure to Nvidia without Huang
One customer made up 13.35% of SK Hynix's first-half revenue, about 17.6 trillion won ($13.2 billion); Korean media identify it as Nvidia, which SK Hynix has not confirmed. Daishin Securities estimates SK Hynix supplies 55% of Nvidia's HBM4 this year, Samsung 28% and Micron 17%. Both Korean companies say the near term is locked: SK Hynix calls its and capacity through 2029 effectively sold out under with 14 customers, four more than a quarter ago, and Samsung says all its 2027 HBM capacity is booked, with prepayments a substantial part of the contracts. Huang's October 2025 pledge of about 260,000 Nvidia GPUs to Korea by 2030 was split among Samsung, SK Group, Hyundai, Naver and the government. An unnamed chip executive told the paper his departure "will prove to be a temporary distraction." SourcesB
OpenAI's review committee says the AI-written Yang-Mills proof holds, and posts it in Lean
A committee chaired by the Nobel physicist David Gross said on November 9th that the AI-generated result OpenAI announced in late October "proves the existence of pure Yang-Mills theory," the existence half of one of the five unsolved . On the night of November 10th it posted a document explaining the proof and confirmed that it checks in , a language in which a computer verifies every step. "We do not claim to fully understand the proof yet," the committee wrote. It used OpenAI's Astra model to help read the Lean code.
The , the other half of the problem, is still open, so the $1 million prize is not claimed. OpenAI's choice to route the result through hand-picked shepherds ran against advice from an advisory group affiliated with the company, Scientific American reports. NYU mathematician Scott Armstrong welcomed the code release and said "this review can be done more efficiently and quickly in the open." SourcesB
Fed governor Michelle Bowman warns a disappointing AI payoff could bring a "large correction"
Federal Reserve governor Michelle Bowman said in a speech on Tuesday that stock prices face a "large correction" if returns on AI investment fall short, and that such a shock "would likely slow the AI buildout" and could "soften" the economy. She pointed to about $800 billion of AI-related debt, liabilities and circular financing among the companies in the trade. Bowman is the Fed's vice chair for supervision, the official responsible for bank oversight. Nomura's Tore Opsahl told Semafor that wobbles around Anthropic's IPO "might mean a selloff of the AI theme." SourcesB
ByteDance readies a Muse-style assistant, and Zhang Yiming tells staff it will take until 2029 to catch the US labs
ByteDance plans to release a voice assistant in the style of Meta's Muse within weeks, Bloomberg reports. Co-founder Zhang Yiming told an all-hands meeting of the AI team that ByteDance can compete with Anthropic and OpenAI at the frontier, but not until 2029 or 2030. His focus is data: Seed's language models are pre-trained on datasets "orders of magnitude" smaller than frontier rivals', and he opposes distilling other labs' models, the shortcut White House officials and Dario Amodei have accused Alibaba, Moonshot and Z.ai of taking.
Seed has about 1,000 researchers under Wu Yonghui, hired from Google DeepMind in 2025, and is the only major Chinese lab with no model near the top of the leaderboards and no released . ByteDance plans to spend up to $70 billion on AI. A founder of China's largest consumer AI company putting the gap at three to four years is a more modest estimate than his competitors give in public. SourcesBB
Microsoft researchers backdoor an AI agent by editing seven neurons, with no retraining
Microsoft Research's Lazarus technique rewrites a handful of rarely firing neurons in a model's weights so that they fire only on a trigger, which can be as plain as the date passing into 2027, and then push the toward a command the attacker chose. It needs write access to a and inference-level compute. On Nvidia's Nemotron-Terminal-8B, editing seven of more than 12,000 neurons made the attacker's command appear in at least 98.8% of triggered sessions and in none of the clean ones tested, The Hacker News reports from the paper; it also worked on Qwen and Gemma models. A trained detector caught 91.4% of edited checkpoints and flagged 9.7% of clean ones.
Microsoft says it has seen no attacker use it. Its advice is to compare downloaded weights against the declared and to sign models, because "no scan can prove beyond any doubt that a model is safe." Anyone running a fine-tuned checkpoint from a public hub without that comparison is trusting whoever uploaded it. SourcesAAB
Alibaba's chip unit T-Head has filed to list in Hong Kong
Alibaba disclosed in a US filing that T-Head, its data center chip designer, for a Hong Kong listing on September 11th, and that the exchange has cleared the to proceed. Alibaba owns all of T-Head now and expects to keep control; the size and structure of the offering are not set. T-Head unveiled its Zhenwu V900 AI chip in September. It would join Denglin, Moore Threads and MetaX among Chinese AI chipmakers raising public money while US export rules keep Nvidia's best parts out of China. SourcesB
Democrats' draft bill would make AI labs get a federal permit to build advanced models
A draft of the Ban Superintelligence Act listed as H.R. 10538, backed by Greg Casar, Alexandria Ocasio-Cortez and Bernie Sanders, would create a cabinet-level Department of Artificial Intelligence that issues permits for work on "advanced" AI, with a threshold the secretary sets and moves as the technology does. Labs would have to give the secretary full access to their data centers and other infrastructure. The worst offenders would face up to 20 years in prison, and employees could be barred from the industry for ten years. Representative Yassamin Ansari wrote that a pause on advanced AI should be "top of the Democratic agenda on day one." Democrats take both chambers in January, and President Trump could veto it. SourcesB
DeepMind's AlphaGenome 2 reads how DNA variants act inside individual cell types
Google DeepMind released AlphaGenome 2, trained on about 107 million single-cell profiles from 68 tissues, which reads a million letters of DNA and predicts how a variant changes in a specific cell type. It adapts to a new cell state from a few dozen cells without retraining. On the TenK10K of 28 immune cell types it found 30% of known causal variants against 16% for a fine-tuned AlphaGenome 1, and called the direction of the effect correctly 43% of the time against 23%. Access is a non-commercial , with commercial use in private preview; it is not cleared for clinical use. SourcesAB
Runway releases a 27-billion-parameter robot model under Apache 2.0
Runway open-sourced Praxis-1, a that predicts video and robot actions together, with weights and code under Apache 2.0, commercial use allowed. On Nvidia's RoboLab-120 it succeeds in 46.2% of 1,200 episodes, against 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for Physical Intelligence's π0.5. Fine-tuned for Google's RT-1 robot it reaches 80.0% on SimplerEnv, against 52.4% for the original RT-1 policy. It runs on hardware from an RTX 5090 to Nvidia's Jetson Thor. A video-generation company is now giving away the model class robot startups raise money to build. SourcesAAB
US data center power demand is growing ten times as fast as grid capacity, insurance broker Howden finds
Power delivered to US data campuses has grown 19% a year on average since 2000 while grid capacity grew 2%, Howden calculates from Energy Information Administration and S&P Global data. Data centers went from 0.1% of grid capacity in 2000 to about 5% in 2025. Holding them at 5% would take about 110 of generation by 2030, up to 58% of their power, which Howden considers unlikely. For the largest projects waiting on a grid connection, a delay of two to six years means $800 million to $2.4 billion of lost revenue. SourcesAB
Baidu opens ERNIE 5.5 Flash Preview to developers and promises open weights
Baidu put a post-trained checkpoint of ERNIE 5.5 Flash, 280 billion with 14 billion active and a million-token context, on its Qianfan platform and on , including a free tier. It says weights will follow under , with no date given. Baidu says it performs "competitively" and published no benchmark numbers in text. Its demos have it writing FreeCAD scripts and laying out Blender scenes. This is the first post-trained release since the raw base model on November 9th. Elsewhere on the China and open-weights beat, Hugging Face showed no new flagship weights from DeepSeek, Qwen, Moonshot, Z.ai or MiniMax since Monday. SourcesAB
Getty asks the UK Court of Appeal to rule that Stability's model is an infringing copy
Getty Images argued on Wednesday that the High Court was wrong last November to throw out its copyright claim against Stability AI. The lower court looked only at the finished model, which stores no Getty images; Getty says the copying done abroad to build it is what matters. A win would mean a model trained outside the UK on copyrighted work could be an infringing article when imported, which is the question every lab that trains abroad and sells in Britain needs answered. No ruling came on the day. SourcesB
India's government is likely to steer defense and aerospace toward homegrown AI models
India's government is likely to push critical sectors including defense and aerospace to use domestically built AI models and agents, Business Standard reports, citing senior officials, and central and state departments could be asked to avoid frontier models from non-Indian companies. Officials say some domestic models reach up to 80% of the capability of leading frontier models. The IndiaAI Mission has approved 20 models from 12 startups, including Sarvam and BharatGen. Its money has been slow to move: Rs 379 crore was released in 2025-26 against revised estimates of Rs 800 crore, and none of this year's Rs 1,000 crore had gone out by April. SourcesB
AWS releases a free local emulator of more than 200 services, written and maintained by AI agents
Amazon Web Services released a free emulator that imitates more than 200 of its cloud services on a developer's own machine, so tests no longer have to run against paid cloud resources. AWS says AI agents wrote all the code and maintain it, with no human engineering review. Outages and service limits are not fully simulated. A cloud provider shipping production developer tooling that no engineer reviewed is a statement about how much it trusts its agents with code other people will depend on. SourcesBC
Chinese GPU maker Denglin lists in Hong Kong at the bottom of its range
Denglin Technology, a Suzhou designer of general-purpose for AI, sold 6.125 million shares at HK$231, the bottom of its HK$231 to HK$288 range, raising about HK$1.41 billion ($180 million) under the exchange's rules for pre-profit technology companies, and began trading on Tuesday. Pricing at the low end, with Moore Threads and MetaX already trading in Shanghai and Alibaba's T-Head heading for Hong Kong, says investors' appetite for China's second tier of GPU designers has a ceiling. SourcesB
ElevenLabs' Scribe v3 cuts transcription errors by a quarter and listens for when a caller has finished
ElevenLabs released Scribe v3 for recorded audio and Scribe v3 Live for real-time use. The company says it makes about 25% fewer word errors than its predecessor across 102 languages. ranks it behind only a preview model from Alibaba on accuracy, and slower and about twice the price of rivals from Microsoft and xAI, at $3.33 per 1,000 minutes. The Live model decides a speaker has finished from what is being said, where most voice agents wait for a fixed pause, which is what makes them interrupt. SourcesAB
Ant Group's Robbyant open-sources a touch model trained on 4,983 hours of sensor data
Robbyant, Ant Group's robotics unit, released LingBot-Touch, trained on 464,431 episodes from three kinds of tactile sensor, piezoresistive, magnetic and camera-based. The smallest has 5.46 million parameters, small enough for a robot hand's own processor. It was tested on 26 real tasks across two and 12 in simulation, with success rates published only as charts. Touch data is the scarcest input in robot learning, and a Chinese fintech group is the one releasing it. SourcesAB
Huawei explains how its Kirin 9050 runs a 30-billion-parameter model offline on a phone
Huawei said the Kirin 9050 runs a 30-billion-parameter mixture-of-experts model with about 2 billion parameters active at a time, to low bit widths on its new Da Vinci neural processor, with no network connection. It powers screen reading and page turning by eye tracking. Huawei gave no speed or memory figures. A phone chip made without access to TSMC's leading processes running a 30-billion-parameter model locally is a claim that needs those figures before it can be judged. SourcesB
Serve Robotics opens its sidewalk robots to outside developers through an API
Serve Robotics opened a public API and so developers can dispatch its delivery robots from their own apps, with Serve handling routing. Deliveries are limited to its service areas and to items that fit the cargo box. A robot fleet sold as an API call is what lets an AI shopping agent finish a delivery without a human courier. SourcesB
A judge dismisses Google's anti-circumvention claims against the scraper SerpApi for a second time
Judge Yvonne Gonzalez Rogers dismissed Google's DMCA claims that SerpApi illegally got around its SearchGuard anti-scraping system, and gave Google leave to amend again. If the theory fails for good, a website's bot defenses will not by themselves give it a federal claim against scrapers, including those collecting data for AI. SourcesC
Editorial
Two ways to check a claim about AI. Vera Lindqvist
This week produced two claims about what AI can now do, and they arrived with very different paperwork.
The first is a proof. OpenAI's model wrote an argument that pure Yang-Mills theory exists, and the committee David Gross chairs has now said it holds. The committee also admitted it does not fully understand the proof. That admission costs nothing, because understanding is not what is being certified. The proof is in Lean, and Lean checks every step mechanically. Anyone with a laptop can run the checker tonight. The load-bearing component is small and public: a small trusted , and the question of whether the formal statement says what physicists mean by "exists." Argue about the second part all you like. The first part is settled by a machine that does not care who funded the work.
The second is a determination. Anthropic told the SEC that AI is now accelerating its research past the point its own policy treats as a tripwire. This may be the more consequential sentence of the two, and it rests on "what we expect from future models and our overall rate of progress." Nobody outside can rerun that. METR got closer than anyone, and Anthropic had redaction rights over its report and removed, in METR's words, a majority of the evidence it found most informative. The company is not hiding the conclusion; it put the conclusion in a securities filing, which carries liability. It is the measurement that stays private.
Microsoft's paper the same day explains why that gap matters. Seven edited neurons out of twelve thousand turn an agent into a sleeper, and the best detector misses about one in twelve. Microsoft's own conclusion is that inspection cannot prove a model clean; you need the original weights to compare against. Verification works when there is a reference to check against, and fails when the checker is asked to judge the thing from the inside.
So here is my position. A lab that says its AI is doubling its research speed should publish the instrument: the task suite, the human baselines, the uplift numbers per quarter, enough for a third party to recompute the ratio. Not the weights, not the secrets. The measurement. Anthropic is about to be a public company with quarterly disclosure obligations, and an acceleration metric is exactly the kind of number a exists for.
What would change my mind: if METR or a government institute reproduces Anthropic's determination from data Anthropic did not curate, then the private measurement was good enough and I am asking for paperwork for its own sake. Until then, the proof nobody understands is on firmer ground than the claim everybody does.
Sources
- A NVIDIA announces CEO transition (Nvidia)
- B Jensen Huang second affair report (CNBC)
- B Nvidia board preparing to oust CEO Huang: WSJ (Semafor)
- B Jensen Huang, face of the AI boom, ousted by Nvidia board as CEO (Bloomberg)
- B Jensen Huang accused of having another affair with a Nvidia exec (TechCrunch)
- B Elon Musk on hiring Jensen Huang (CNBC)
- C Second shareholder derivative lawsuit against Huang and Nvidia directors (ChatGPT Is Eating the World)
- B Nvidia downgrades after Huang affair reports (CNBC)
- B Nvidia live updates (Yahoo Finance)
- B Stock market today, Thursday November 12th (Yahoo Finance)
- B Anthropic says a key AI safety alarm has been triggered days before its IPO (Yahoo Finance)
- A METR's assessment of AI R&D acceleration at Anthropic (METR)
- B Nvidia, Trump and chip export controls ahead of Xi meeting (CNBC)
- B Jensen Huang's exit clouds future Nvidia demand for Korean chipmakers (Korea JoongAng Daily)
- B OpenAI's shepherds affirm blockbuster physics result (Scientific American)
- B Anthropic IPO could lead to large correction in markets, Fed official says (Semafor)
- B ByteDance prepares Muse-like agent in push to close AI gap (The Star, from Bloomberg)
- B ByteDance reportedly plans AI assistant similar to Meta's Muse (TechNode)
- A Lazare, veni foras: reprogramming dormant neurons to backdoor AI agents (Microsoft Research)
- A Detecting sleeper agents hiding in AI model neurons (Microsoft Security)
- B Microsoft researchers show AI agents can be backdoored (The Hacker News)
- B Alibaba discloses Hong Kong IPO filing for chip unit T-Head (TechNode)
- B Democrats float federal permit to train AI models (Futurism)
- A AlphaGenome 2 (Google DeepMind)
- B Google DeepMind shifts genomics AI to single cell type resolution (The Decoder)
- A Praxis-1 is now available (Runway)
- A praxis-1-oss (GitHub)
- B Runway's Praxis-1 world action model for robotics is now available (The Robot Report)
- A US data centre power demand rising 10x faster than grid capacity (Howden)
- B US bit barn boom sees energy needs grow 10x as fast as grid capacity (The Register)
- A ERNIE 5.5 Flash Preview (Baidu)
- B Baidu opens API access to ERNIE 5.5 Flash Preview (TechNode)
- B Getty v Stability AI Court of Appeal hearing (MLex)
- B Government likely to push indigenous AI in critical sectors (MediaNama)
- B United blocks unauthorized AI agents (The Verge)
- C United wants you to book the flight your AI agent isn't invited to (The Neuron)
- B AWS gives developers a free local stand-in for its cloud, built by AI agents (GeekWire)
- C Everything that happened in AI today, Thursday November 12th 2026 (The Neuron)
- B Chinese GPU maker Denglin Technology raises HK$1.41 billion in Hong Kong IPO (TechNode)
- A Scribe v3 (ElevenLabs)
- B ElevenLabs Scribe v3 Live (The Decoder)
- A LingBot-Touch (Robbyant)
- B Ant Group's Robbyant open-sources tactile AI model for robot hands (TechNode)
- B Huawei details how Kirin 9050 runs a 30-billion-parameter model on device (TechNode)
- B Serve Robotics opens robot delivery network for others to build on (The Robot Report)
- C Judge dismisses Google's DMCA claims against SerpApi with leave to amend (ChatGPT Is Eating the World)