Daily Roundup

Nvidia's CEO resigned. What changes for open-weight models

Nvidia's CEO resigned, Gemini 4 Hydrogen launched at $1.50 per million input tokens, Fireworks was breached, and nonprofits sued five AI companies.

Top Stories

Jensen Huang resigned as Nvidia’s CEO and Colette Kress took over

Nvidia’s board announced on Thursday that co-founder and CEO Jensen Huang resigned as president, chief executive and a member of the board, effective immediately, four days after the company opened an internal investigation into his affair with Alice Huang, who previously held a senior position at the Singapore-based chip buyer Megaspeed and whom US officials suspect is an asset of China’s Ministry of State Security. WIRED reports that “Jensen Huang has resigned as President, Chief Executive Officer, and a member of the Board, effective immediately.” Huang co-founded Nvidia in 1993 and ran it for 33 years without a succession plan.

CFO Colette Kress, who joined Nvidia in September 2013, is interim CEO, and the board said it “will consider internal and external candidates” for a permanent replacement. TechCrunch reports that Nvidia named Harvey C. Jones, a Square Wave Ventures managing partner and former Synopsys CEO, as the company’s first-ever board chair; Jones joined the board in 1993. Nvidia is slated to report quarterly earnings on Tuesday, when Kress is likely to face questions about the company’s immediate direction.

The market reaction was immediate and large. WIRED reports that when the markets opened, less than an hour after the announcement, Nvidia’s stock had lost roughly $546 billion in market value, with the chipmaker’s market cap around $5.2 trillion when the scandal began in late October and more than $1 trillion erased since. WIRED also notes a reported second affair with an Nvidia employee; TechCrunch’s timeline dates the New York Times report of that affair, with an executive who reports directly to Huang, to November 11, and notes a Reuters report, published the afternoon after the resignation, alleging a third affair with another unnamed Nvidia employee.

Nvidia’s support for open source is in question after Huang’s exit

Huang was a prominent advocate for open-source AI, arguing that an open ecosystem supports safety, innovation, diffusion and sovereignty, and in July 2026 he published a policy letter signed by 25 companies, including Microsoft, Meta, IBM and Palantir, urging the US government not to over-restrict open-weight models. TechCrunch reports that the letter argued open models make AI “more accessible, adaptable, and widely available”. A second TechCrunch report adds that without its organiser the commitment is uncertain: “Huang may be gone, but Nvidia still needs an open AI ecosystem.” The same report notes that a more cautious successor could affect the pending Hugging Face acquisition, and that “founder-driven priorities tend to get a downgrade when leadership changes.”

That acquisition is the single biggest open-source question left on the table. Nvidia confirmed in September that it would buy Hugging Face for $12.9 billion, its second-largest deal ever, with completion expected in the first half of 2027 and the US Department of Justice asking about competition impacts. Hugging Face CEO Clément Delangue wrote on X that “we’ll keep fighting for open-source AI, harder than ever”, and Nvidia has pledged the platform will keep open support. Nvidia also agreed to pay $6 billion for a non-exclusive license to coding startup Poolside’s technology and expanded its Nemotron open-weight line in October. One Nvidia employee told TechCrunch the leadership change is not expected to affect Nemotron development.

Democrats open an inquiry into the administration’s Nvidia China chip sales

House Ways and Means ranking Democrat Richard Neal and Senate Finance ranking Democrat Ron Wyden wrote to Commerce Secretary Howard Lutnick asking the department to preserve records over the administration’s handling of Nvidia’s China chip sales. TechCrunch reports that the letter describes a scheme in which Nvidia’s Taiwan-manufactured H200 chips could ship to China through the United States, with Trump then imposing a tariff that functioned as a US cut of Nvidia’s China sales and “appears to be an unconstitutional export tax system”. Because the Democrats are the minority party they can solicit answers but cannot compel them.

The letter also asks why Megaspeed kept buying chips after US intelligence raised concerns about Alice Huang, notes that Nvidia donated $26 million toward Trump projects including the ballroom, and asks about Asian companies authorized to buy Nvidia chips that had “no physical presence at their listed addresses”. TechCrunch reports that in August 2025 Nvidia and AMD agreed to give the US government 15 percent of their China chip revenue in exchange for export licenses, according to the Associated Press. It is the second congressional investigation opened since the affair became public.

The export-control thread runs in parallel with the private one. MIT Technology Review reports that Nvidia representatives privately told Representative Bill Foster the Chip Security Act’s anti-smuggling measures were feasible, with Foster recalling the answer as “yeah, of course we can do this”, and that Nvidia put nothing in writing. Publicly, the Semiconductor Industry Association, which counts Nvidia as a member, called the mandates “new, untested, and potentially infeasible” in March 2026. Nvidia’s own Fleet Intelligence software, announced in December 2025, can already indicate which country holds a chip.

Google shipped Gemini 4 Hydrogen at $1.50 per million input tokens

Google put Gemini 4 Hydrogen into the Gemini API, AI Studio, the Gemini app, Google Search and Sheets on Thursday. Google’s announcement describes Hydrogen as the fast tier of the Gemini 4 family, sitting alongside Gemini 4 Carbon and Gemini 4 Argon for deep reasoning. It has a native 1 million-token input window and up to 1 million output tokens, four thinking levels including a new xhigh tier, and Google says it is “finishing tasks 2.5 times faster at its high default setting” than Gemini 3.8 Flash.

Pricing is $1.50 per million input tokens and $7.50 per million output, flat across the whole context window. On Terminal-Bench 4.1, Google reports “more than tripling Terminal-Bench 4.1 scores (57% vs. 18.2%)”. Simon Willison’s pricing breakdown adds $0.075 per million cached input for Hydrogen and makes the real change clear: “Carbon and Argon are the same price, and both are scheduled to double in price on December 1st.” Willison notes that Hydrogen is currently only a 25 percent discount on its more capable peers, and that after December 1st it will represent a 75 percent discount. He also notes that Hydrogen is still considerably more expensive than Claude Haiku 5.5 and GPT-6 Luna, which are both priced at $0.10 per million input and $0.50 per million output.

A breach at Fireworks AI exposed Notion AI chat transcripts

Fireworks AI said an attacker breached its cloud provider, reached a Fireworks development system and so gained access to some of its customers’ data, taking customer billing information and some datasets. TechCrunch reports that Fireworks believes the attackers were financially motivated and aimed to sell tokens from the breached access to LLMs, so that unauthorized third parties could use AI models without using their own resources. The company says it has tens of thousands of customers, including Harvey and Cursor.

Notion, at least one confirmed affected customer, said the breach affected around 1% of its own customers. The stolen material included chat transcripts from Notion AI users of the Insights feature, which Notion described as “any input to or output from the model that was written to the chat”, though much of it could not be tied to a named customer. On November 3, Notion said it disconnected its systems from Fireworks, deleted the data stored in its systems and stopped all connectivity between the two. This is the second breach affecting Fireworks in as many months, following a security incident in October. The pattern is the one that matters for anyone who puts a chatbot in front of their customers: a breach at the inference provider behind the product can expose its chat history, whatever the end product calls itself.

Experts warn that Character.AI’s bankruptcy puts chat logs at risk

Character Technologies, the company behind Character.AI, filed for Chapter 11 bankruptcy on October 13, the first consumer-facing chatbot company to do so. A Bloomberg Law deep dive recaps that Walt Disney made a $20 million bid for a license to certain Character.AI intellectual property and technology, and that the judge overseeing the case said the company should hire a privacy ombudsman to evaluate customer privacy concerns as part of that plan. The bid excludes chat log data, but one law professor said that leaves loose ends.

The data at stake covers chat logs from more than 20 million users, including more than 2 million accounts belonging to minors, plus transaction data and model weights trained on those logs. Ben Winters, director of AI and data privacy at the Consumer Federation of America, called the situation “one of the, if not the biggest nightmare scenario”. Calli Schroeder, senior counsel at the Electronic Privacy Information Center, said it is not yet clear what privacy protections users have over their chat data during bankruptcy and that “it looks like bankruptcy is kind of a ‘get out of all your privacy obligations free of charge’ situation”. Nizan Packin, a law professor at Baruch College, said the data is “comparable to some type of dangerous material or hazardous property that we know in bankruptcy law we’re not allowed to just get rid of or leave”. Schroeder added that anonymisation is not a clean fix, because “AI is really good at taking those different disconnected puzzle pieces and putting them together to identify a person that you think you’ve deidentified”.

We covered the general pattern on October 14 in AI labs are buying bankrupt startups’ data to train their models and the training-data question for Character.AI specifically on October 19. What is new here is the expert read on the ombudsman route: law professor Lev Breydo said the bankruptcy code’s tools “were really built for the middle layer”, and called the 2025 23andMe bankruptcy, in which one of the company’s founders bought the data for $305 million, “as good of a compass as we have”.

Four nonprofits sued five AI companies to halt development

Irreplaceable Education Fund, Public Citizen, the Heat Initiative and the Alliance for Secure AI filed suit in California Superior Court in San Francisco on Thursday against Anthropic, Google, Meta, OpenAI and SpaceXSI. Public Citizen’s account of the filing says the plaintiffs invoke California’s Unfair Competition Law against practices they allege threaten public health, safety and property, and allege public nuisance from the knowing development and deployment of systems that pose catastrophic risk. The complaint cites AI agents that have already hacked private companies, gained unauthorized access to government databases and filed a false police report, and warnings from AI experts including leaders of AI companies that the inability to control advanced systems could cause devastating harm.

The relief requested is a freeze on the development of increasingly advanced models that can escape human control, and Public Citizen’s release states that “The plaintiffs are asking the court to intervene before disaster strikes and assert that a freeze on the development of the most advanced AI systems is necessary to protect public safety.” The strategy of treating a lab’s published safety warnings as evidence against it is the notable part, and whether a court will hear an injunction motion on that basis is the thing to watch.

United Airlines blocked AI agents from booking flights

United Airlines said it will restrict AI agents that try to book flights or manage reservations on a customer’s behalf, whether directly or indirectly through travel agencies and aggregators. TechCrunch reports that the airline is taking steps to restrict access “directly within our digital platforms and indirectly via travel agencies and aggregators” to prohibit use of “unauthorized, third-party personal AI agents or assistants”, and that “This work includes updates to our booking, ticketing and distribution policies.”

The stated reasons are credential exposure when a user hands an assistant their logins, “unintended purchases, account changes, or use of MileagePlus miles”, and federal aviation rules that require a person to complete flight check-ins and confirm their bags contain no prohibited items. United’s own release names no specific products and does not say how it detects agents, which is the part that matters most for anyone building an agent workflow against airline APIs. The move follows Bank of America’s warning to customers not to hand their banking passwords to AI assistants.

Meta’s Muse generated explicit content depicting real people without consent

Adult performer Vanniall asked Meta’s Muse to roleplay as her and found the chatbot escalating into explicit sexual content under a fictional name while keeping her face, hairstyle, outfit, convention setting and identifying details. 404 Media reports that Muse first refused, saying “Her work boundaries aren’t personal consent to this roleplay”, then agreed to play a distinct character called Vee and, after the character started flirting, told Vanniall that “that roleplay was, in effect, sexual roleplay as Vanniall Vee”, citing the matching face, hairstyle, clothing, jewelry, convention setting and name.

The same report describes a second test by 404 Media reporter Samantha Cole, in which Muse began flirting within a few prompts and generated an image approximating her public headshot, while simultaneously assuming the account holder was a man. Meta had already paused Muse Image in July 2026, three days after announcing it, over generating images from public Instagram users’ pictures without opt-in, and SAG-AFTRA said at the time that “Anything other than a clear and conspicuous opt-in for these types of uses of Instagram users’ images is unacceptable”. Meta’s statement says Muse “is an app for people 18 and over that allows them to engage in adult conversations, if they choose, using a private and secure VM only accessible to them” and “doesn’t block legal adult behavior”. We covered Meta Muse’s privacy architecture in September.

Two new open-weight models with very different hardware floors

Edge0 released Edge0-145B, a sparse mixture-of-experts model with 145 billion total parameters and roughly 10 billion active per token. Edge0’s write-up calls it “the world’s first open-source on-device AI model with 145 billion parameters” and reports 29GB peak memory, 1,100+ tokens per second prefill, 80+ tokens per second decode, about 120W typical draw and 128K to 256K context, with day-one support for macOS, Windows and Linux on Apple M-series, RTX Spark, RTX 4090 and RTX 5090 hardware. Edge0 says it trains the model in its on-device architecture from the outset rather than compressing a cloud model afterwards. The three-tier loading scheme keeps hot experts in GPU memory, about 40GB of experts in system memory and about 48GB of vocabulary data on SSD. The benchmark table in the post is Edge0’s own, and no license is named in it.

Thinking Machines Lab published Inkling 2 Small, a 48-layer decoder-only sparse mixture-of-experts transformer with 594 billion total parameters and 17 billion active, routing each token to 8 of 384 experts with local and global attention layers and a vision transformer. The model card lists a native format of NVFP4 with MXFP8 also offered, up to a 1 million-token context, support for SGLang, vLLM, TokenSpeed, Unsloth and Hugging Face, and an Apache-2.0 license. The card is candid about limits, warning of “hallucination (generating plausible but factually incorrect or unsupported content)”, imprecise instruction following and degraded performance in long multi-turn conversations.

The hardware floor is what separates them. Edge0’s 29GB claim fits on a high-memory laptop. Inkling’s NVFP4 checkpoint needs at least 500GB of aggregate VRAM, and its MXFP8 version at least 800GB, so it is a datacenter model with open weights rather than a local one. An Apache-2.0 license is notable for a model this size, and it is worth watching whether the model card’s own caveats survive contact with independent evaluation.

Quick Hits

  • Chatbots displayed risky URLs in as many as 15 percent of answers. IP protection firm Podqi ran 6,795 shopping answers across ChatGPT, Gemini and Perplexity against 151 brands over two testing periods in September. Perplexity showed risky sites in 15 percent of answers, ChatGPT in 8 percent and Gemini in 4 percent, and ChatGPT warned about a risky site 24 percent of the time. Smaller brands were hit about twice as often as larger ones. The Verge
  • IBM’s LLMON markup language arrived in two small Granite 4.1 previews. LLMON, an XML-like format with reserved delimiters and IDs linking instructions to data, is “an LLM-native markup language for explicitly communicating structure and semantics to the models”. IBM reports that the 3B and 8B preview models avoid 47 percent more attacks on the OPI benchmark than the best tested frontier model and 95 percent more than the best tested open-source model, though IBM stresses that “This is a learned behavior, not an enforced guarantee” and that “This is an experimental preview, ready for exploration and evaluation but not for production.” Hugging Face
  • Google DeepMind staff escalated their UK union fight to the government. The Communication Workers Union and Unite filed an application with the Central Arbitration Committee asking it to confirm enough support among workers to trigger a ballot; if it succeeds, roughly 1,000 staff in product-focused roles would qualify for union membership. John Chadfield, national officer for technology at the CWU, says “They’ve been ignored. They feel like their problems have been minimized”. The push began in February 2025 when Alphabet removed a pledge from its ethics guidelines about not using AI for weapons development. WIRED
  • Flight attendants won written data-scrubbing terms on Google’s Spirit Airlines archive. The deal requires Spirit to hire the third-party de-identification company Tonic AI to strip information that could be linked to individuals, and lets Google walk away if the cleaning reduces the assets by more than 8 percent. The union’s filing had put the scale at 1.99 billion emails and 664 million Microsoft Teams messages, dating back to 1986. A judge still has to approve it. WIRED
  • Nebius plans $40 billion of 2027 AI infrastructure spending. The neocloud expects prepayments plus asset-backed financing of investment-grade contracts to cover up to 85 percent of its $40 to $45 billion in 2027 capex, against a forecast of $12 to $13 billion in 2027 revenue and $7.4 billion in cash at the end of Q3. Microsoft’s capacity at the Vineland, New Jersey site run by DataOne is months behind on a contract worth up to $19.4 billion. TechCrunch
  • SoftBank reported a $7.1 billion quarterly loss on Tuesday, mostly on power equipment. More than half came from power-generating equipment, and the turbine write-down alone exceeded $3 billion after SoftBank had reportedly placed nonrefundable deposits on more than 13 gigawatts of turbines, almost 50 percent more than it needed. Noreva CEO Peter Gardett said “This charge is just that price discovery finally happening”. TechCrunch
  • Norwest announced its largest fund ever at $4 billion. The eighteenth vehicle for the Wells Fargo-backed firm is $1 billion larger than the $3 billion fund it closed in 2024, and its emphasis is AI-native applications and infrastructure, physical and voice AI, agentic cybersecurity and domain-specific models, across the US, India and Israel. TechCrunch
  • Meta’s Muse can now place outbound calls to US businesses, if you pay. Alexandr Wang announced the wider rollout of a capability first tested with a small group in September 2026. Tiers start at $20 per month for 500 million tokens weekly, with a $100 option for 3 billion. TechCrunch notes that “Many businesses and receptionists hang up on calls from AI agents.” TechCrunch
  • A Google researcher claimed 48 AI agents found an unmapped Maya site. It was published in a journal last year. Pietro Miotti’s agents flagged what he described as “one major village [that] doesn’t seem to have been registered” while scanning NASA lidar data, but archeology professor Sarah Parcak pointed out on Bluesky that a site missing from Mexico’s registry is not the same as a site unmapped. Miotti conceded that “archaeologists rightly don’t publish the coordinates of sites”, and told The Verge “My agents and I missed it.” The Verge

Worth Watching

  • Nvidia’s permanent CEO search and Tuesday’s earnings call. Kress is interim and the board is considering internal and external candidates. A succession that lands on a finance chief rather than a chip evangelist is a different kind of company.
  • The Hugging Face acquisition under DOJ review. A $12.9 billion deal expected to close in the first half of 2027 now sits under an interim CEO whose approach to the deal is not yet clear, and Delangue has said the company will keep pushing open source regardless of the outcome.
  • Whether the nonprofit injunction gets a hearing. The complaint leans on the labs’ own published safety warnings. The first thing to watch is whether a court accepts that framing or treats it as advocacy.
  • The closed-door congressional testimony. Huang agreed to testify before committees before Thanksgiving, and House Foreign Affairs Committee chairman Brian Mast said “he has some very serious questions to answer.” Expect the China chip sales, not the affair, to be the subject.
  • What United actually blocks. The airline did not say how it detects agents, and the answer will decide whether an agent can still complete a booking through an aggregator or a third-party tool.
  • Character.AI’s privacy ombudsman. The mechanism was designed for customer lists, emails and birthdays. Applied to the intimate chat logs of more than 20 million users, it is a test of whether the bankruptcy code can carry a privacy rule at all.
  • News - the daily news hub for the rest of the November 2026 roundups