Is Jensen Huang’s job on the line? Yale’s Jeffrey Sonnenfeld estimates it’s “about a 50-50”
Nvidia’s board announced Sunday that it’s launching an independent investigation after CEO Jensen Huang told staff he was unfaithful to his wife. But this isn’t a case of just another CEO having an extramarital affair. The romantic interest is Megaspeed executive Alice Huang, whom U.S. officials believe is a Chinese spy who was feeding information back to Beijing. In short, the stakes reach well beyond Nvidia, given how many tech and financial partners depend on its place at the center of the AI economy. “I’m responsible for an entire industry,” Huang recently told Bloomberg in a sit-down.
A spokesperson for the Chinese Embassy in Washington dismissed the allegations in late October as “groundless smears and slanders,” and no one has been charged. But U.S. intelligence officials reportedly briefed President Donald Trump about their concerns a year ago, and unnamed senior officials inside the Trump administration told the Times that they doubt Huang could have remained unaware all that time of Alice Huang’s intelligence-gathering role for China. When Trump was first briefed, the Times reported, officials didn’t believe Huang knew, and it isn’t clear whether their assessment has changed. The board, meanwhile, says it sees no sign that national security or company secrets were compromised.
The question is whether Huang can survive this. To think through what could come next, this editor spoke Sunday with famed professor Jeffrey Sonnenfeld, senior associate dean for leadership studies at the Yale School of Management and founder of the Chief Executive Leadership Institute, which has hosted Huang at its CEO summits. Sonnenfeld likes Huang, he said, and he called Huang’s career salvageable but put the odds at “about a 50-50.” He also had a lot more food for thought.
First, the damage. Until now, Huang has been “revered and trusted by everybody in technology,” Sonnenfeld noted. But he called this episode “remarkably poor judgment.” Even if Huang was merely naive, he said, “any employee in Nvidia other than him would be fired over this,” and the recklessness alone would mean “an irretrievable loss of confidence.”
Sonnenfeld suggested that Huang’s saving grace could be what social psychologists call idiosyncrasy credits, if he can prove he was being duped. The term comes from Edwin Hollander, a social psychologist who proposed in a 1958 paper that people earn latitude in a group by showing competence and following its rules. Huang has been earning latitude since he co-founded Nvidia in 1993. But all that goodwill can run out when a screw-up looks self-serving or touches the group’s core norms. Here that group includes Nvidia’s board, its roughly 42,000 employees who are bound by security rules, and, in a looser sense, the U.S. government, given it restricts its exports.
We also talked about precedents, except, as it turns out, there really are none for what’s unfolding right now. Executives who survive personal scandals sometimes pick up where they left off at another company, Sonnenfeld observed. Former HP CEO Mark Hurd is one example. Between 2005 and 2010, he drove HP to the top of the tech industry by revenue, then was pressured to resign after he filed inaccurate expense reports that hid a personal relationship with a contractor. Oracle hired him about a month later. Huang has no equivalent because Nvidia is his company
As for the bench at Nvidia, that doesn’t really exist either. While there is typically a hierarchy inside companies of Nvidia’s scale and importance, Huang has kept things as flat as possible, with about 60 direct reports by his own account. That structure makes it hard to figure out who would succeed him (though his children, both of whom work at Nvidia, are sometimes the subject of succession speculation). That won’t stop the worldwide guessing game that’s sure to follow. And boards in crisis often reach for a “messianic solution” from outside, Sonnenfeld said, floating former Google CEO Eric Schmidt as one idea.
Nvidia’s role at the center of a massive and complex circular financing web is what makes this particular predicament so tricky and, undoubtedly, terrifying for a wide number of constituents. Nvidia isn’t just the biggest provider of chips driving the AI boom but also largely the bank behind the buildout. In addition to stakes in hundreds of smaller startups, Nvidia has invested in OpenAI, Anthropic, xAI, CoreWeave and Nebius. In August, it also signed preliminary agreements with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR aimed at mobilizing more than $500 billion over time, and Huang has said Nvidia could backstop up to a quarter of it. The move was widely interpreted as a way to spread risk. Still, a leadership crisis at the center is how a stumble becomes a collapse.
So what’s Nvidia’s board likely doing right now? Sonnenfeld noted that Huang’s constant, public advocacy for selling advanced chips to China had already made people anxious. Anthropic’s Dario Amodei, whose company Nvidia has backed, had a lot to say about selling chips to China at Davos back in January, likening it to “selling nuclear weapons to North Korea.”
Now the board’s independent members — meaning everyone but Huang — will have to establish what technology reached China if any, and by what routes, including by better understanding Nvidia’s relationship with Megaspeed. The Singapore cloud company, which grew out of a Chinese gaming business and became Nvidia’s largest buyer in Southeast Asia, bought roughly $2 billion of its chips, the New York Times reported in October 2025, raising broader questions about whether it was helping China circumvent U.S. export restrictions by re-exporting the chips to China or renting them to Chinese customers remotely. (An Nvidia spokesperson told The Hill after the Times report that the allegations were “thoroughly investigated and debunked last year.”)
In the meantime, Huang will keep running the company, Nvidia said Sunday. That’ll be a tall order, even for a CEO who already spends much of his time traveling the world to promote Nvidia. For starters, three House chairmen want a closed-door briefing from Huang on November 18. Nvidia is also reporting its newest earnings on November 17. Sonnenfeld expects that even if Huang keeps the title, he’ll have to hand off some of the day-to-day decision-making.
Only one thing is certain right now. Huang doesn’t own the kind of super-voting stock that would cement his power at Nvidia. Sonnenfeld often argues for dual-class structures that empower founders, and he says Huang’s success without one has been the best counterargument to that case. But if Huang had one, Sonnenfeld continued, removing him would be harder.
This episode may be remembered as a pivotal moment for how the tech industry guards its secrets. Sonnenfeld argues tech has been “way too lax” compared with defense contractors that face clearance regimes. The arrest last month of an OpenAI engineer points to the same problem. He was accused of hiding a recording device behind a TV in an executive conference room, and while OpenAI has said it has no indication that model weights, source code, or customer data were accessed, the case highlights how exposed tech companies are to people who get close.
Huang may well survive the investigation (Sonnenfeld hopes the board clears him of knowingly aiding anyone). The bigger test comes when Nvidia’s employees, customers and partners, and, indirectly, the millions of Americans with Nvidia in their 401(k) plans, have to decide whether an AI economy built largely on the judgment of Huang still makes sense.
https://techcrunch.com/2026/11/08/is-jensen-huangs-job-on-the-line-yales-jeffrey-sonnenfeld-estimates-its-about-a-50-50/Reposted. Will be removed on request. All rights belong to the original author.
More from this batch
Anthropic secures $3.4 billion generator deal for AI data centers, FT reports
Anthropic bought commitments for 1.7 gigawatts of heavy-duty gas-fired generators from OnePWR Solutions in a deal worth about $3.4 billion, …
SplatAnything(浙大 3DV):给前馈式 3D 高斯重建加「高斯间距归一化」,同一个模型能处理任意尺度的场景
改进了什么:前馈式 3DGS 模型一步就能从几张图重建 3D 场景,但场景大小和视角一变,预测出的高斯点分布就失准,泛化差。靠什么:作者提出 Gaussian Spacing Normalization,把场景级的高斯间距统计量规范到统一尺度再预测,相当于先消掉「这个场景有多大」…
网易有道 Confucius4-R2T2:不改模型结构,只用「最长稳定前缀」标签做后训练,让流式语音识别接近整句识别的准确率
改进了什么:流式语音识别为了低延迟要边听边出字,但已经吐出的字常常被后面的音频推翻重写,字幕一闪一闪,准确率也不如等整句说完再识别。靠什么:作者构造「最长稳定前缀」标签,只让模型学习输出那部分之后不会再变的文字,在现有模型上做后训练,不需要改架构。效果:论文称在低延迟下接近整句(…
Video-Coder:让智能体写代码搭 3D 场景再渲染成视频,用执行结果和视频质量做强化学习奖励
改进了什么:直接用扩散模型生成视频很难精确控制物体位置、运动和镜头。靠什么:Video-Coder 换了一条路,让模型写并执行代码来构建动态 3D 场景,再渲染出视频;训练分两步,先监督微调,再用「代码能否执行」和「视频是否符合要求」两类奖励做强化学习。效果:论文称可控性明显优于…
AgentBaton:智能体任务做到合适阶段就从强模型「交棒」给便宜模型,准确率不掉、API 成本大降
改进了什么:智能体全程用最强模型很贵,全程用小模型又容易失败。靠什么:作者发现任务前期规划交给强模型、后期执行交给高效模型,在合适的阶段交接就能保住准确率;他们训练了一个轻量检测器 AgentBaton,判断什么时候可以交棒。效果:论文称大幅节省 API 费用,并把准确率和成本的…
P²-MOPD:多教师在线蒸馏时,把各个专家模型的参数当先验,学生恢复更快、各项能力更均衡
改进了什么:把多个领域专家模型蒸馏成一个学生模型时,常见的多教师在线策略蒸馏(MOPD)收敛慢,而且容易偏向某几个教师。靠什么:P²-MOPD 在在线蒸馏目标下同时优化教师的混合系数和学生权重,让专家模型作为参数先验参与初始化和约束。效果:论文称比 MOPD 恢复更快、能力更均衡…
可灵团队 LongBanana:首个专家设计和校验的「长上下文多参考图合成」基准,考验智能体生图时能不能找对证据、拒绝干扰
这是个新基准,不是新模型:任务要求从大量分散的参考图里找到真正相关的视觉证据,排除干扰图,把人物、物体和关系对上,最后合成一张连贯的图。作者称这是首个由专家设计并逐题校验的此类基准,共 90 道精心构造的题目。它测的正是「智能体驱动的图像生成」里最容易出错的环节:多张参考图一多,…
韩国 ETRI 开源无代码 AI 开发工具 TANGO 2:网页里点选就能用行业数据训练 LLM,并部署到云、内网服务器或端侧
韩国电子通信研究院(ETRI)11 月 9 日宣布开源 TANGO,这是一个把模型训练、优化、部署串成一条流水线的无代码工具,目标用户是工厂、医院、造船、农业等现场的非 AI 专家。TANGO 1 做视觉 MLOps,TANGO 2 扩展到生成式 AI 和 LLMOps,能用行业…
花叔开源全球鸟类贴纸库:用 AI 为 AviList 2026 全部 11011 种鸟各画一张贴纸,带 43 种语言名称和结构化数据
作者「花叔」用 Claude Code、Codex 等工具批量生成并整理了全部 11011 种鸟的贴纸,附 43 种语言的鸟名和分类数据,配套网站 birdalbum.com 可以浏览。仓库说明写的是 CC BY 4.0(GitHub 未自动识别许可证),10 月 30 日创建,…
kubai-icons:给 Claude Code、Codex 等编码智能体用的动画图标技能,附 2600 多个风格统一的开源图标
这是一个 agent skill:装进编码智能体后,它能按同一套风格画出带悬停动效的图标,并提供 2600 多个现成图标,支持 React、Vue 和 SVG,配有浏览网站 icons.kubai.design。11 月 3 日创建,约 420 星;GitHub 识别许可证为「其…
集智俱乐部解读 Nature Communications:清华团队让数千个 LLM 智能体自由交友讨论政治,没有预设规则也会自发抱团、观点极化
清华电子系李勇团队 10 月发表在《自然·通讯》的论文,模拟数千个大模型智能体自己建立社交关系、讨论政治议题,结果在不预设任何极化规则的情况下,智能体自发形成同质化小圈子并出现观点极化,和人类社会的现象很像。作者认为这样的智能体网络可以当作研究舆论动态和干预手段的「合成试验场」。…
新智元解读 EMNLP 2026 论文 Auto-Robotist:让 LLM 智能体设计机器人,每次试错都沉淀成可复用的「设计技能」
密歇根大学团队的工作:智能体提出机器人形态,在仿真里训练控制器并评估,再把成功和失败的经验提炼成自然语言写成的设计技能,下一次设计直接调用,搜索过程因此越跑越快。论文已被 EMNLP 2026 主会录用并开源。它和最近很多「自进化技能库」的智能体思路一致,只是落在了机器人协同设计…