Three stories worth your coffee today. The biggest chipmaker on earth has its CEO in a mess nobody saw coming, Anthropic just bought itself a power plant's worth of gas generators, and Meta shipped a model that costs basically nothing to run. Let's go.
01 of 03
Nvidia's board is investigating Jensen Huang
On Sunday, Nvidia's board announced an independent investigation into CEO Jensen Huang after reports that he was the target of a Chinese intelligence operation. Independent directors are running it with outside legal counsel.
Here is how we got here. The New York Times first reported on a memo that circulated among U.S. policymakers earlier this year. It claimed Alice Huang, an executive at Singapore-based chip company Megaspeed, was recruited by China's Ministry of State Security to spy on Nvidia's CEO. The Financial Times then reported the memo also described a romantic relationship between the two. The Times says U.S. intelligence watched Alice Huang for more than a year.
Nvidia's first response was that the interactions were "not unusual" and that it does not share classified information with customers. Then Huang sent a message to the company admitting an affair, without naming her. "I was unfaithful to my wife," he wrote, and "I owe my family and all of you an apology." The board's statement says it has "no reason to believe" national security or company secrets were compromised and still has confidence in Huang.
Wall Street did not wait around. The stock fell more than 4% Monday, and by that morning a whole lineup of shareholder firms had announced investigations: Schall, Brown & Schwartz, Rosen Law Firm (which says it is preparing a class action), Levi & Korsinsky, Bronstein, Gewirtz & Grossman, and the Law Offices of Frank R. Cruz. Stanford Law professor Robert Bartlett told TechCrunch that earlier "not unusual" line was "clearly a misleading statement," because a CEO's knowledge gets imputed to the company.
What I want to know is what Huang knew and when. The affair is his family's business. Whether the CEO of the company at the center of U.S. chip export fights knew he might be talking to a foreign intelligence asset is everybody's business. Nvidia has spent two years lobbying to loosen China export limits, and that pitch just got a whole lot harder to make with a straight face. If the investigation comes back with a shrug, nobody in Washington is going to buy it.
Source: TechCrunch, TechCrunch on the lawsuits
02 of 03
Anthropic buys $3.4 billion in gas generators
According to the Financial Times, Anthropic is spending $3.4 billion on Caterpillar natural gas generators that can put out up to 1.7 gigawatts. The twist is where they came from. Equipment supplier OnePWR ordered them a couple of years back but could not line up financing to finish the purchase. Anthropic originally planned to buy through OnePWR, then went straight to Caterpillar instead.
The reason anybody would pay that kind of money for somebody else's order is simple: the line is long. Turbine backlogs run into the early 2030s, and grid connection queues can be just as bad. Google, Amazon, Microsoft, and Meta have all announced on-site gas plants for data centers, so Anthropic is joining a crowd, not starting one. This also comes on top of the compute deal disclosed in Anthropic's S-1 back in May, where it agreed to pay xAI $1.25 billion a month through May 2029.
Here is where I land on it. I get it. If your whole business depends on electricity and the grid says come back in 2032, you buy your own. But every AI company that talked a big game about clean energy is now quietly lining up for gas, and one estimate cited by TechCrunch has U.S. data centers burning more natural gas than Germany and Japan combined by 2035. That bill lands on somebody, and I have a feeling it is going to be regular folks paying for it on their utility statement.
Source: TechCrunch, citing the Financial Times
03 of 03
Meta's Muse Ember costs a dime per million tokens
Meta Superintelligence Labs released Muse Ember on Monday, which Meta calls its lightest and most cost-efficient model yet. The price is 10 cents per million input tokens. Meta says that undercuts Anthropic's Haiku and OpenAI's Luna by a wide margin and puts it in the same price range as open-source options like DeepSeek V4.1 Flash and GLM 5.3 Flash.
Ember is not meant to be the smart one. It is the workhorse. Meta built it for the repetitive grunt work agents do all day: tool calls, coding subtasks, routing, data extraction, and parsing long documents. It takes text, images, and audio, so screenshots, charts, scanned docs, voice memos, and video are all fair game. The plan is to pair it with Meta's bigger Muse Wonder model, or as Meta put it, "Muse Wonder plans, Muse Ember handles the volume." Meta points to the Artificial Analysis Intelligence Index to argue Ember wins on cost per task, though that is Meta's claim, not an independent review.
This is the release I care about most as somebody who actually builds with this stuff. Big model plans, cheap model grinds is how real agent systems get built, because nobody can afford to run every little tool call through a frontier model. If Ember holds up outside of Meta's own charts, the price floor on agent work just dropped again. I will believe the benchmarks when I run it on my own junk.
Source: TechCrunch