Huang Renxun Extramarital Affair Scandal Escalates as Nvidia Stock Plunges After Hours; OpenAI Posts $10.9 Billion Quarterly Loss | AI Morning Brief
Nvidia's stock plunged more than 3% in after-hours trading on November 12, as CEO Huang Renxun was revealed to have had an extramarital affair with an internal executive who reported directly to him; the board had previously launched an independent investigation and the company faced a shareholder class-action lawsuit. OpenAI's third-quarter revenue grew about 50% quarter-over-quarter to $10 billion, but its net loss for the quarter reached $10.9 billion. xAI and SpaceX launched the Grok Voice 3.0 speech-to-speech API, priced at $0.08 per minute.
Body · AI translation
Nvidia stock plunged after hours amid the escalating Jensen Huang affair; OpenAI posts a $10.9 billion quarterly loss | AI Morning Brief
Phoenix Tech Download the app for exclusive early access November 12, 2026 07:47:52 From Beijing
Phoenix Tech News, November 12 — A series of major AI industry stories broke overnight. Nvidia's stock suddenly fell more than 3% after hours as the personal controversy surrounding Jensen Huang continued to escalate; Elon Musk's xAI launched a new-generation Grok voice API, bringing real-time AI voice interaction into the era of 8 cents per minute; OpenAI's third-quarter revenue is said to have reached $10 billion, but its loss for the same period widened further to $10.9 billion. Meanwhile, Google is stepping up testing of a new version of Gemini, Anthropic's IPO process is drawing attention, and New York State has begun strengthening safety oversight of frontier AI models.
Domestically, this year's Double 11 became a concentrated showcase for AI products and digital-human services. JD.com disclosed that it sold more than 50 million AI products, and as demand for large-model API calls surged, data security risks behind low-cost token relay stations also began to draw attention.
01 Nvidia shares plunge more than 3% after hours; Jensen Huang reportedly had an extramarital affair with a subordinate executive
At 6:01 a.m. Beijing time on November 12, according to a Cailian Press wire, Nvidia's stock suddenly plunged more than 3% in U.S. after-hours trading. Related reports said Nvidia CEO Jensen Huang reportedly had an extramarital affair with an internal company executive who reported directly to him. This is the latest development in the controversy surrounding Huang personally and the company's governance, following Sunday's news that Nvidia's independent directors on the board hired a former federal prosecutor to launch an independent investigation and that the company faced a shareholder class-action lawsuit.
02 SpaceX and xAI officially launch the Grok Voice 3.0 speech-to-speech API, billed at $0.08 per minute
SpaceX and xAI jointly announced that their new-generation real-time voice model, Grok Voice 3.0, has officially launched in the speech-to-speech API. The interface opens end-to-end low-latency voice conversation capabilities to external developers and enterprise customers, with an official flat rate of $0.08 per minute.
The speech-to-speech API eliminates the intermediate loss of traditional speech-to-text followed by text-to-speech, pushing response latency down to the level of human conversation. The clearly posted price of 8 cents per minute is already close to the call-handling cost of traditional human customer service, making it a replacement option with a straightforward cost calculation for smart hardware, in-car voice assistants, and overseas call centers.
03 Google opens testing of a new version of its Gemini AI model to internal employees
Concept image of a Google internal engineering team conducting sandbox testing of a new-generation Gemini model
Google has begun allowing internal employees to test a brand-new version of its Gemini AI model. The broad internal real-world testing follows the previous day's full rollout of Gemini 4 Carbon to paid API and Ultra subscription users, aiming to collect feedback through real internal office and engineering scenarios and accelerate the pace of the next stage of model iteration.
Just after pushing Gemini 4 Carbon to the commercial API, Google opened company-wide testing of a new version in the early hours of the next morning, a clear acceleration in its model iteration cadence. Using the daily coding and documentation work of tens of thousands of internal engineers as a whetstone has become a standard move for Silicon Valley giants to eliminate model weaknesses before public release.
Beijing time, early morning November 12 — New York State officially launched a registration portal for frontier AI developers and simultaneously established the New York AI Safety and Responsibility Research Center, hiring about 80 researchers to conduct independent safety testing of mainstream large models using the Empire AI computing cluster. Under New York State's RAISE Act, frontier AI labs with annual revenue exceeding $500 million (including OpenAI and Anthropic) must complete registration by January 1, 2027, and report safety incidents within 72 hours. First violations carry a maximum fine of $1 million, and subsequent violations up to $3 million.
New York State is directly funding an 80-person scientist team and mobilizing a state supercomputer to test models itself, while also setting hard rules for mandatory 72-hour reporting and fines of up to $3 million. This means AI regulation at the U.S. local level no longer relies only on companies' self-inspection reports; mandatory third-party safety audits are becoming a fixed compliance cost for leading model makers.
05 Report: Anthropic IPO several times oversubscribed, expected to price on November 18 and list on Nasdaq on November 19
Anthropic IPO roadshow announcement interface, image source: Anthropic official
Citing people familiar with the matter, Anthropic has received several times oversubscription from investors after launching its IPO roadshow, with strong institutional demand. Under the plan, Anthropic intends to offer shares at $70 to $78 per share, with the public offering and a $26 billion private placement totaling up to about $90.4 billion, and a fully diluted valuation of up to $1.4 trillion. Goldman Sachs and Morgan Stanley are lead underwriters. The company is expected to set the final offering price on November 18 and officially list on Nasdaq under the ticker ANT on November 19.
After voluntarily cutting the valuation to a range of $1 trillion to $1.4 trillion, Wall Street institutions immediately responded with several times oversubscription. With only one week left before the November 19 listing, this capital feast totaling nearly $90 billion will become the watershed that tests the secondary market's true risk appetite for generative AI.
06 DeepL launches voice AI translation model Aure, reproducing the speaker's original timbre and tonal fluctuations in real time
Concept image of cross-language real-time voice translation preserving original timbre and emotional fluctuations
On the evening of November 11 Beijing time, DeepL officially released Aure, its most dynamic voice AI translation model. As the underlying core of DeepL Voice, Aure is built on original timbre preservation technology and can synchronously reproduce the speaker's tonal fluctuations, emotional shifts, stress emphasis and speaking pace in cross-language real-time interpretation. The model has been launched on DeepL Voice web conference edition, face-to-face conversation edition and voice API, initially supporting 8 languages including English, German, French, Japanese and Korean, but not yet Chinese.
The biggest fear in multinational meetings is translation software reading tentative or emphatic tones as flat, machine-like speech. DeepL packs stress and emotional changes into the real-time interpretation stream, precisely hitting the pain point of cross-border business communication; the only regret is that Chinese is still missing from the initial list.
07 OpenAI's third-quarter revenue grows 50% quarter-on-quarter to $10 billion, single-quarter loss widens to $10.9 billion
According to investor financial data disclosed by Forbes, thanks to a surge in enterprise subscriptions and API sales, OpenAI's revenue in the third quarter of 2026 grew about 50% quarter-on-quarter to $10 billion, but dragged down by high computing power and model R&D spending, its single-quarter net loss reached $10.9 billion. Microsoft's disclosed financial data also confirms this scale: in the fiscal quarter ended September 30, Microsoft's OpenAI-related commercial revenue reached $10 billion, accounting for nearly 11% of Microsoft's total revenue for the quarter.
Taking in $10 billion in a single quarter while losing $10.9 billion at the same time intuitively shows the coexistence of high growth and high cash burn at leading large-model labs. The enterprise service fees earned are almost entirely converted into bills paid to computing centers and chip manufacturers. How to reduce unit inference costs will be a problem OpenAI must solve before its sprint to listing next year.
08 Anthropic roadshow discloses that Claude completed AMD chip inference adaptation in only 3 days
Large model autonomously completes AMD GPU computing chip inference adaptation across platforms
According to Anthropic's IPO roadshow materials, Anthropic's Claude Mythos Preview model took only about 3 days to autonomously figure out a method for another AI model to run efficiently on AMD MI355X GPUs. Under traditional engineering processes, completing model inference adaptation for a brand-new computing platform usually takes 6 to 12 months.
Letting the large model write the underlying operator adaptation code itself compresses the AMD graphics card migration cycle, originally starting at half a year, to 3 days. This not only demonstrates engineering efficiency to investors but also substantially weakens the binding threshold of a single hardware ecosystem. When models can autonomously run across chips as substitutes, downstream buyers' bargaining power over computing power will rise accordingly.
09 The 18th Double 11 concludes: JD.com sells over 50 million AI products, JoyAvatar digital humans stream for over 9 million hours
Double 11 promotion AI hardware bundle and digital human livestream selling
As of 23:59 Beijing time on November 11, the 18th Double 11 officially concluded. Tmall and JD.com did not disclose total transaction volume for the full cycle for the fifth consecutive year, with AI software and hardware and digital human services becoming core growth points. JD.com's final report shows that during the full Double 11 cycle, more than 50 million AI products were sold cumulatively, with transaction volume of AI computers, AI workstations, AI glasses and AI recording devices growing more than 100% year-on-year. JoyInside completed more than 70 million household interactions cumulatively, and JoyAvatar digital humans streamed for more than 9 million hours cumulatively and contributed more than 1.3 billion yuan in sales.
As traditional GMV figures retreat into the background, 50 million AI hardware units and 9 million hours of digital human store livestreaming have become the most tangible incremental indicators of this year's Double 11. Consumers buy AI hardware for efficiency and peace of mind, while small and medium merchants put digital humans on late at night to capture idle traffic in order to save high livestreamer slot fees. AI has thoroughly settled the input-output account in e-commerce scenarios.
10 Media investigation into Double 11 large-model token gray-market relay station chaos; China's AI inference data volume surpasses training data volume for the first time
Explosive growth of large-model token calls and security risks of unofficial API relay stations
A deep investigation published by Science and Technology Innovation Board Daily pointed out that data from the China Center for Information Industry Development shows China's AI inference data volume has surpassed training data volume for the first time, with average daily token calls exceeding 175 trillion, more than 1,000 times higher than at the beginning of 2024. As Taobao and JD.com launched official AI recharge stores for Double 11, some unofficial AI relay stations with hidden entities and no filing information are soliciting customers on e-commerce platforms under gimmicks such as 30% of the official price and one yuan for 10 million tokens. Experts from the State Information Center recommend unifying token measurement standards as soon as possible and introducing third-party billing verification.
The volume of inference data has surpassed the volume of training data, marking that the center of gravity of China's AI industry has officially shifted from training large models to using models at high frequency. But buying API keys from unregistered relay stations that offer a 70% discount, in an attempt to save a few dozen yuan, actually exposes enterprise code and internal documents entirely to unregistered intermediaries; the cost of data leakage far exceeds the computing fees saved.
"Special statement: The content of the above works (including videos, pictures, and audio if any) is uploaded and published by users of "Dafeng Hao," a self-media platform under Phoenix Net. This platform only provides information storage space services.
Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."