Nvidia's board removes Jensen Huang and names Colette Kress interim CEO
Directors ended Jensen Huang's 33-year tenure after reports of two affairs, one with a woman alleged to have been a Chinese intelligence operative, and named the chief financial officer as interim chief executive.

On this page
Nvidia’s board announced on Thursday that Jensen Huang had resigned as president, chief executive and director. The resignation took effect immediately and ended 33 years at the company he co-founded in 1993. Chief financial officer Colette Kress will serve as interim chief executive while keeping her finance role. The board also appointed former Synopsys chief Harvey C. Jones as its first-ever chair. Directors said they have begun a search for a permanent CEO and will consider candidates inside and outside the company.
The decision followed reports of two extramarital affairs, the first with an executive at Megaspeed, a Singapore cloud provider and Nvidia customer. She is alleged to have passed information to Chinese intelligence, and Huang told employees over the weekend that he had been unfaithful. The second, reported by the New York Times on Wednesday, was with Nvidia’s vice president of healthcare, Kimberly Powell, who reported directly to him. Independent directors had retained former US attorney Joon H. Kim and his firm Wachtell, Lipton, Rosen & Katz on Sunday. The announcement did not say whether that investigation had concluded, and the board’s statement carried no comment from Huang. Nvidia still plans to report third-quarter results on 17 November.
Featured source: Nvidia , The Verge .
Other Stories
TSMC commits $1 billion to stakes in OpenAI and Anthropic
TSMC’s board approved up to $500 million for OpenAI preferred stock reserved for large investors. It approved another $500 million for Anthropic common shares at the company’s expected IPO price, Reuters reported. The chipmaker earned more than $23 billion in the quarter through September, its sixth consecutive record quarter. Analysts told the Wall Street Journal that the investment could steady sentiment toward AI stocks. They also warned that financing loops between suppliers, customers and investors make real demand harder to judge.
Filed from: Ars Technica .
AI agents produce a machine-checked proof of Signal’s core cryptography
Google DeepMind said its AlphaProof Code framework formally verified key security properties of Signal’s compiled protocol code in four weeks. Agents turned Signal’s security goals into Lean specifications and generated more than six million lines of Lean code. Agents then flagged about one percent of the codebase for human review. Researchers reintroduced two previously fixed bugs, and the system proved both versions insecure. The work covered the protocol’s core cryptography rather than the whole app, and Google declined to disclose its cost.
Filed from: Google DeepMind , Ars Technica .
Memory prices lift DRAM revenue 369 percent
Global DRAM revenue rose 369 percent year over year in the third quarter to just under $200 billion, Counterpoint Research reported. Bit shipments grew only about four percent, so prices did most of the work, and average selling prices reached $1.9 per gigabit. Samsung held 38 percent of revenue, SK hynix grew more than 40 percent from the previous quarter, and China’s CXMT slipped to nine percent. Counterpoint expects tight supply to persist because major new fabs will not start production until 2028.
Filed from: Counterpoint Research , TechRadar .
IBM plans a 1,040-qubit system by linking two chips
IBM said its Nighthawk r3 processor, due in 2027, will be built to execute 10,000 quantum operations across 520 programmable qubits. The chip uses 2,034 physical qubits. IBM plans to join two of the processors into a single 1,040-qubit system that year, linking chips instead of building one larger processor. Its current Nighthawk r2 has 120 programmable qubits and has run 34 billion circuits for clients including E.ON and Q-CTRL. Those clients report speedups of up to 19 times over earlier IBM hardware.
Filed from: IBM Newsroom , Ars Technica .
A Microsoft researcher hides a backdoor in a few dormant neurons
A Microsoft researcher, Dimitris Papailiopoulos, demonstrated an attack called Lazarus that turns a language model into a sleeper agent. It works by changing the weights of neurons that are rarely active in ordinary use. The model behaves normally until a trigger appears, such as a date, a keyword or an error message. It then follows attacker instructions, which could include running remote code or copying data out when the agent holds the right permissions. The technique requires write access to model weights, and a public paper was not immediately available.
Filed from: TechRadar .
Uber, DoorDash and Airbnb report almost no bookings from AI agents
Executives at three consumer platforms said agents still account for a negligible share of transactions. Uber counted 320 million rides booked by people during the week of Halloween and four booked by agents. DoorDash chief executive Tony Xu said agents placed several hundred orders. Airbnb’s Brian Chesky said agent bookings were a tiny fraction of the total. The figures cover a short window and do not establish a trend.
Filed from: TechRadar .
From the Community
eMudhra traces a certificate failure to its own script error
eMudhra has published a statement on the flaw that led browser root programs to distrust one of its root certificate authorities. The company says a script error caused it to issue intermediates with weak keys. A researcher reported on 17 October that the private keys of eight emSign G3 intermediate TLS CAs could be recovered. All eight were revoked the following day, none had issued a publicly trusted certificate, and the company has appointed KPMG as an independent external reviewer.
Filed from: eMudhra statement .
A shuffle that never stores the array
Andrew Helmer wrote up Andrew Kensler’s permute(), a function that walks an array in pseudorandom order without building a shuffled copy. It hashes an index over a power-of-two domain and uses cycle walking to cover arrays of any length. That gives O(k) time and O(1) memory for k draws. Helmer credits Kensler’s sampling paper and notes the method is not cryptographically secure and works best below domains of about 2^27.
Filed from: Helmer’s Blog .



