Nvidia Investigates Its CEO Over Affair With Suspected Chinese Spy

Read the episode (Narration script)

Good morning and welcome to Cyber World News. The company that sells the world most of its AI chips woke up to a spy story, and it's a strange one. After that, a once a minute wobble in the world's power grids finally gets explained, two researchers find a way into other people's cars, and a giant AI model squeezes onto a desktop. Let's get started.

Nvidia's chief executive, Jensen Huang, told employees on Sunday that he had been unfaithful to his wife. "I owe my family and all of you an apology," he wrote. That would normally be private. It isn't, because of who the other person reportedly is. According to the New York Times and the Financial Times, a memo that circulated among Trump administration officials this spring said US intelligence believed China's Ministry of State Security had recruited a Chinese citizen named Alice Huang to gather information on Jensen Huang, and described a romantic relationship between them. His message didn't name her.

Here's why this is bigger than a scandal. Alice Huang has been an executive at Megaspeed, a Singapore data center company and an Nvidia customer. Last year the Times reported that Megaspeed bought about two billion dollars of Nvidia chips, and US officials suspect the company of helping China get around the ban on advanced AI chip sales. Nvidia says those allegations were investigated and debunked. So the question is whether a foreign intelligence service got close to the man running the company at the center of the US and China contest over AI.

Nvidia's board has opened an independent investigation, run by its outside directors with their own lawyers. The board says it has no reason to believe national security or company secrets were compromised, and that Huang is cooperating. Before Sunday, Nvidia had called the pair's interactions "not unusual." China's embassy in Washington calls the allegations "groundless smears." Nobody has been charged.

Washington is moving anyway. Three House committee chairs have asked Huang for a closed door briefing on November 18th, the day after Nvidia reports earnings, and Senator Chris Coons says Congress will investigate. Nvidia shares were down about four percent before the market opened Monday. Huang doesn't hold the special voting shares that protect some founders, so this board really does hold his future. Yale's Jeffrey Sonnenfeld, who studies chief executives, puts his odds of keeping the job at "about a 50-50."

Now a mystery with a much nerdier culprit. For years, the frequency of power grids from Europe to Australia to Malaysia has dipped slightly every sixty seconds, at the same moment, on different continents. Germany's biggest grid operator brought the puzzle to Dirk Witthaut at Forschungszentrum Jülich, who calls what followed years of detective work. Hydropower resonance was the first suspect, but resonances fade, and this never did. Something was running on a clock. So the team scattered more than eighty power meters around their own campus and found the same once a minute spike beside a big battery.

The culprit: large batteries, including the backup power systems in data centers, check their charge level every minute, on the minute. Each check draws a sliver of extra power, and because these machines are online, their clocks are synchronized, so thousands do it at once. In Ireland, where data centers now use nearly a fifth of the country's electricity, that minute dip grew fourfold between 2016 and 2021. It's still well inside safe limits, but it wears on hydropower turbines. The fix is almost comically simple: have each battery check at a random second. Getting every manufacturer in every country to do it is the hard part.

Security researchers Sam Curry and Shubham Shah published a chain of bugs last week at two carmakers, which they agreed not to name, that let them take over any customer's account. For a connected car, that meant whatever the owner's app can do, including finding the car, unlocking the doors and starting the engine. In Australia, Shah writes, the starting point was nothing more than a number plate, looked up through public registration sites. The crucial step was slipping past a web firewall with an oddly written web address. Shah says he and Curry both tried to find that trick by hand and failed, then gave the login software's source code to OpenAI's GPT 5.6 Sol Ultra. It worked on the problem for about six hours and found it. Curry says both carmakers fixed what they reported within about two hours.

A project called llama-modchip, popular on Hacker News this weekend, says it runs DeepSeek's 284 billion parameter V4 Flash model on a desktop with 64 gigabytes of ordinary memory and no graphics card at all, at roughly nine to fifteen tokens a second. Models that size usually demand expensive graphics cards. This one gets there by squeezing each of the model's numbers to under two bits, and the speed figures are the developer's own.

Speaking of memory, after Sam Altman's podcast remark that home computer memory mostly sits idle, a developer released a script called Ram Altman. It checks your computer every five minutes, and if more than ten percent of the memory is free, it emails Altman to let him know, at most once a day. His address isn't filled in by default. The project's own description: "One email a day from one person is a joke. Ten thousand of them is a denial of service attack with better material."

That about wraps it up for today, friends. Thanks for listening; see you tomorrow!