AI News, Nov 12: Nvidia's Board Weighs Firing Jensen Huang
The Big Story: Nvidia’s board spent Wednesday night deciding whether to fire Jensen Huang
Nvidia’s board met late Wednesday to decide whether to remove its co-founder and CEO after the New York Times reported a second relationship, this one with Kimberly Powell, Nvidia’s vice president of health care, who reports to him. It follows Huang’s admission to staff on Sunday of an affair with the former CEO of Singaporean data center supplier Megaspeed, whom US intelligence believes worked for China’s Ministry of State Security. “I was unfaithful to my wife,” he wrote. The board has hired Wachtell, Lipton, Rosen & Katz to investigate the spying allegations. Shares fell more than 8% in after-hours trading, and no decision had been announced by Thursday morning. Nvidia reports earnings on Nov 17.
Washington moved faster than the board did. Three House lawmakers gave Huang until Nov 18 to testify, and three members of Congress told CNBC he should go: Sen. Kevin Cramer (R-N.D.) said “as a matter of his own conscience, he should resign,” and Rep. Don Bacon (R-Neb.), asked the same question, answered “from the little I know, yes.” Sens. Elizabeth Warren (D-Mass.) and Jim Banks (R-Ind.) are pressing colleagues to fold their AI Overwatch Act into the defense authorization bill.
The company supplying most of the world’s AI compute now has an unresolved succession question five days before it reports, and its scandal has become the vehicle for the export-control legislation the chip industry spent two years lobbying against.
Everyone shipped a leash for agents, and a pen test showed why
Pen-test firm Shinobi Security handed two off-the-shelf models VPN credentials and no operational-technology expertise. Anthropic’s Claude severed a replicated wind farm’s grid connection in 66 minutes for $20 in tokens, stopping six turbines delivering nearly 10MW. Moonshot’s open-weight Kimi did it in just under two hours for $8. The rig copied a Polish wind farm hit by suspected Russian hackers last December, where the human intruders needed 11 days inside.
Almost everything else built for agents on Wednesday was restraint. The Linux Foundation launched the Agent Name Service, a DNS-based naming, registration and agent-to-agent authentication standard with 11 backers including Cisco, Salesforce and Cloudflare. Nvidia contributed OpenShell to the CNCF, a deny-by-default sandbox that keeps credentials hidden from the agent and makes it explain any blocked action in plain language; Red Hat showed it at KubeCon. And United Airlines barred third-party agents from completing bookings through its site, aggregators and travel agencies, noting that only a human can acknowledge at check-in that a bag holds no explosives. CEO Scott Kirby was blunter: “To be honest, I don’t want to create another middleman.” Delta, by contrast, lets authenticated assistants book.
Today’s Top Stories
Google shipped a model that doesn’t write anything
Gemini Decision 1 scores up to 1,000 yes/no, multiple-choice and rubric questions against a single multimodal input in one pass, returning calibrated probabilities instead of tokens. It takes 256K of input, answers in 150 to 250ms, and costs $0.10 per million input tokens with output unbilled, which Google says is nearly 5x cheaper than Gemini 3.5 Flash-Lite via its Interactions API. The bet underneath it: most production LLM calls were never generation tasks at all, but classification, routing and policy checks wearing a chat interface.
AWS says agents wrote its new emulator
Amazon open-sourced AWS Emulator, a free local stand-in built from more than 400 API models across more than 200 AWS services. The provenance is the news. “All of AWS Emulator’s code was written and tested by agents,” AWS wrote, with engineers building the scaffolding that runs them plus “the specs, tests, and merge gates that decide when that work is correct.” It is the largest agent-authored codebase a major vendor has claimed, and the claim is entirely AWS’s own, with no external audit.
DeepMind’s genome model goes cell by cell
AlphaGenome 2 predicts what a DNA variant does in a specific cell type rather than in a bulk tissue sample, trained on more than 100 million single-cell profiles across 68 tissues. On the TenK10K benchmark of 28 immune cell types it identified nearly twice as many causal variants as a single-cell-tuned AlphaGenome 1 (30% versus 16%) and got directional gene effects right 43% of the time against 23%. It is research-use-only and explicitly not validated for clinical work. Worth reading those numbers the other way: most causal variants still go unidentified.
Corporate AI spending hit a record and barely grew
Ramp’s October index, from card data across more than 85,000 US companies, shows spending up 1.65% month over month. The distribution is the real finding: the top 1% of firms spend a median $10,200 per employee per month while the typical company spends $16.32, and 1% of companies drive 80% of revenue for OpenAI and Anthropic. The average price per million text tokens fell 21% in a single month. Customer support’s share of AI spend climbed from 0.8% in June to 13.3% in October, the first clear sign of budget moving off software development.
Quick Hits
- Security: The Fireworks AI breach exposed Notion AI chat transcripts the provider’s zero-data-retention default did not cover; Notion says under 1% of workspaces may be affected and to “treat this data as if it were exposed.”
- Infrastructure: CoreWeave expects about $9B in customer prepayments and deposits over three months rather than borrow at bond yields that topped 15%: “We will not use the unsecured debt market to fund our business at current yields,” said CEO Michael Intrator.
- Chips: Britain’s Fractile raised a $720M Series C from Thrive Capital, Founders Fund, Lightspeed and Redpoint for inference silicon that stacks DRAM on each chip, with no revenue or named customer yet.
- Policy: Australia will require frontier model makers to open up to independent audits, publish safety frameworks and report incidents to government immediately, enforced by a new body that can levy steep penalties.
- Grid: Power delivered to US data campuses has grown about 19% a year since 2000 against roughly 2% for grid capacity, per broker Howden, which expects up to 58% of that power to come from off-grid sources by 2030.
- Security research: Forescout ran an AI harness over 10 TCP/IP stacks and found 99 vulnerabilities, 66 of them new, in 90 hours for $935, work it says took about nine months by hand.
- Meta: Santosh Janardhan, head of infrastructure and co-head of engineering, is leaving in early 2027 after 17 years; former AWS senior vice president David Brown takes over compute as Meta guides to $135–145B in capex.
- Courts: Getty’s case against Stability AI reached the UK Court of Appeal, with Getty arguing a model trained abroad becomes an “infringing copy” once imported even though it stores none of the images.
- Vibecoding: A rewritten Humble Bundle site let users log in without 2FA and delivered email codes an hour late, the day’s biggest consumer-facing AI-code failure on Hacker News.
- Agentic commerce: The behavior United is blocking barely exists yet, with Uber reporting 500 agent-placed food orders and four ride bookings out of 320 million trips in a peak week.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
