A short seller flagged Nvidia CEO's ties to a suspected Chinese spy six months before the scandal
Six months before the Nvidia spy scandal emerged, the New York-based short seller firm Culper Research had already raised questions about Nvidia CEO Jensen Huang’s unusually close relationship with Alice Huang, an executive at Megaspeed, a Singapore-based company that buys Nvidia chips. She is now suspected to be a Chinese spy.
The Nvidia CEO admitted to an extramarital affair over the weekend, and the company has now opened a formal investigation into his relationship with Alice Huang (who shares his last name but is not related). U.S. intelligence officials believe she was recruited by the Chinese government to collect intelligence on Jensen Huang and influence his position regarding U.S. export controls on advanced AI chips.
Culper’s May 13 report is resurfacing because it picked up on the connection between Jensen Huang and Alice Huang while investigating Megaspeed’s ties to Nvidia.
In particular, the report called the relationship between the two executives “uncomfortably close,” noting that the Nvidia CEO was spotted partying and socializing with Alice Huang on multiple occasions. The report even claimed that some Nvidia employees wondered whether they had a “special relationship.”
Culper Research also theorized the China business that Nvidia lost to export controls had been rerouted through Southeast Asia, and that Megaspeed, a major chip buyer in the region, was secretly financed by Chinese tech giant Alibaba through shell companies. It pointed to financial records that showed Megaspeed’s total liabilities and equity had grown from $33 million at the end of 2023 to around $3 billion by the end of 2024, much of it in the form of refundable deposits from an unidentified source.
Financial Times reported Monday that documents show Alibaba lent money to Megaspeed through subsidiaries.
Culper is a short seller, so it stood to financially benefit from making these allegations, and those about possible illegal chip diversion have not been independently verified. Its report is available on the web and was posted to X back in May.
https://techcrunch.com/2026/11/09/a-short-seller-flagged-nvidia-ceos-ties-to-a-suspected-chinese-spy-six-months-before-the-scandal/Reposted. Will be removed on request. All rights belong to the original author.
More from this batch
The race is on to sue Nvidia
Nvidia's stock slid more than 4% after its board opened an investigation into Jensen Huang's relationship with a suspected Chinese spy, and …
Jensen Huang pulls out of ElevenLabs Summit amid Nvidia controversy
The Nvidia CEO also pulled out of an Anthropic event earlier.
中国能从黄仁勋身上得到什么?Verge 梳理他对出口管制和特朗普的独特影响力
英伟达持有约 990 亿美元前沿 AI 公司股权,黄仁勋又能直接接触特朗普,因此是外国情报机构的高价值目标。文章引用 CSET 分析师的看法,讨论这段关系可能如何影响他推动对华卖先进芯片,以及 Megaspeed 背后批量获取芯片的公司网络。
特朗普在白宫记者会上力挺黄仁勋:「他说他根本不知道那是间谍,我相信他」
特朗普没说清自己何时得知此事,只说「不久以前」,并称 FBI 在调查「别的事」。此前《纽约时报》报道他约一年前就收到过相关情报简报。这关系到英伟达对华芯片政策会不会因此变化。
数据中心遭抵制,核心可能只是电:并网推高电价、自建燃气电站又招来污染争议
记者走访弗吉尼亚「数据中心走廊」后认为,反对声浪不像 OpenAI 的 Chris Lehane 说的那样是一场文化战争,而是实实在在的用电问题:当地数据中心平均功率 54 兆瓦,超过丹佛国际机场。对做 AI 基础设施选址和私有化部署规划的人,电力已是硬约束。
Meta 把 Muse Ember、Muse Wonder 和 Muse Spark 1.3 带到 Google Cloud
Meta 的闭源 Muse 系列模型除了自家 Model API,现在也能在 Google Cloud 上用。对已经在 GCP 上的企业,多了一个低价跑智能体杂活的选项。
Artificial Analysis 智能指数:Fable 5.2 以 65 分登顶,比 Gemini 4 Carbon 高 5 分、比 Fable 5.1 高 15 分
进步在哪:最高算力档拿到 65 分,较低算力档也以 62 分在每一档都领先,而且消耗的 token 少得多;但最高质量平均每个任务 26.5 美元,约是 Opus 5.5 的 2.6 倍。靠什么:Anthropic 未公布训练方法,文章提到 Anthropic 核心代码库合并的改…
Fable 5.2 完成了物理学家搁置 23 年的一个黑洞计算,已有物理学家核验
Rohan Paul 转述 Anthropic 的预发布案例:问题源自 1983 年霍金与 Don Page 关于黑洞的工作,Fable 5.2 补完了当时没算完的部分。这是模型在基础科学上产出新知识的又一个例子,细节以 Anthropic 官方案例页为准。
苹果新论文:直接让 Claude Opus 5.5 看摄像头、控制机械臂,抓取物体比用示范数据训练的机器人模型更稳
进步在哪:通用前沿模型不经过机器人专门训练,就在操作任务上超过了从人类示范学习的机器人策略模型。靠什么:把多视角摄像头画面直接给模型,让它直接输出机械臂控制,依靠通用视觉推理能力而不是专门的示范数据。
UC Berkeley 论文:让一个「主管」不断重组编程智能体团队,GPU 内核性能超过专家手写代码
进步在哪:单个编程智能体写 GPU 内核时前期进展快,之后就停在较简单的设计上,离专家水平很远;在 MLA 等内核上,多智能体方案突破了这个平台期。靠什么:主管智能体持续调整团队分工和探索方向,避免整个团队过早收敛。
LambdaMix:一次训练得到一整族可调偏好的策略,部署时用一个偏好向量切换取舍
进步在哪:常规后训练只能为固定的目标权重得到一个模型,用户偏好变了就得重训。靠什么:在冻结的基座上为每个目标各训一个 LoRA 适配器,用组相对在线强化学习联合训练,推理时用偏好向量决定适配器的混合比例,并分别归一化各目标奖励。对需要在「有用 vs 安全 vs 简洁」之间按客户调…
LingBot-Touch:把不同触觉传感器的信号统一成 token,用 JEPA 式隐空间预测训练,胜过 MAE 和 VAE
进步在哪:异构触觉传感器的数据第一次能用一个共享 Transformer 处理。靠什么:掩码隐空间预测(JEPA 路线),在同一份数据上和 MAE、VAE 目标做了对比,论文称隐空间预测更好。适合关注具身智能和 JEPA 路线的人。