Morning Brief: Microsoft 365 recovers after 19-hour global outage; DeepSeek, Kimi and other models alleged to carry…
Microsoft 365 cloud services fully recovered after a global outage of about 19 hours that affected roughly 20 services including Outlook, Teams and Copilot across more than 70 countries and regions; Microsoft confirmed the cause was a technical issue introduced by a recent upgrade. Unemployment Labs, through weight forensics, found that six Chinese frontier models including DeepSeek-V4-Flash, MiniMax-M3 and Kimi K3 carry features of NVIDIA Blackwell architecture GPUs, but the report says this does not constitute direct evidence of export control violations.
Body · AI translation
Monday 16 Morning Brief: Microsoft 365 service recovers after 19-hour global outage 07:24 10-minute read Morning Brief at a Glance
Microsoft 365 service recovers after 19-hour global outage5
Microsoft 365 service recovers after 19-hour global outage
According to multiple media reports, in the early hours of November 14 local time, Microsoft announced that its Microsoft 365 cloud service had fully recovered after a global outage of about 19 hours, and the backlog of emails had also been processed. The failure began at noon on the 12th and within an hour rapidly spread to more than 70 countries and regions worldwide, affecting a total of about 20 services including the email system Outlook, the collaboration software Teams and the AI assistant Copilot.
Microsoft ruled out the possibility of a cyberattack or an AI agent causing the failure, confirming that the incident stemmed from a technical issue introduced by a recent upgrade. People familiar with the matter revealed that Exchange, the underlying service supporting email and calendars, would rapidly exhaust available resources and become unresponsive after server startup. Microsoft's cloud business head and other executives convened more than a thousand engineers to troubleshoot overnight, adopting a phased recovery strategy to prevent the system from being overwhelmed again by a sudden influx of requests.
The incident caused widespread disruption to government affairs, healthcare, the judiciary and business operations worldwide. Government agencies and court offices in the United States, the United Kingdom, Japan, Singapore, Australia and other countries were hindered; a court in Manchester, UK, was forced to adjourn, and the UK National Health Service (NHS) postponed some medical procedures and had difficulty sending electronic prescriptions to pharmacies; some banks and trading firms in Hong Kong had their operations disrupted; Euronext and shipping giant Maersk were also affected. In addition, Switzerland's Richemont and US commercial vehicle company Workhorse were forced to postpone the release of their financial reports, and clothing giant PVH Group's planned layoff communications in Asia were also forced to be put on hold. A large number of institutions and individuals had to fall back on using telephones and paper notices.
Microsoft 365 currently has more than 450 million paid commercial users, with revenue of nearly $28 billion in the previous fiscal quarter. After the incident, Microsoft's stock price remained stable. Some analysts pointed out that given the high cost of replacing office workflows, this incident will not shake Microsoft's market position in the short term, but in the long run it may prompt some enterprises to consider alternatives.
In recent years, major global technology infrastructure has repeatedly suffered large-scale disruptions due to internal maintenance or software updates. For example, Amazon Web Services (AWS) suffered a 15-hour outage in 2025, and cybersecurity company CrowdStrike also caused 8.5 million Windows systems worldwide to crash in 2024 due to a faulty software update. Clothing giant PVH Group's planned layoff communications in Asia were also forced to be put on hold—so from this perspective, did Microsoft actually do a good thing? This headline's word order is really strange. The sentence "After the incident, Microsoft's stock price remained stable" is quite amusing. With a service like this, they're still thinking about reducing the 365 cloud storage? They're really courting disaster1
Domestic frontier models carry Blackwell GPU signatures
On November 12, the nonprofit research organization Unemployment Labs released a report stating that through forensic analysis of public model weights, it found that six frontier large models released since April 2026 by DeepSeek, Moonshot AI, MiniMax and Ant Group carry signatures of Nvidia Blackwell architecture GPUs.
The report shows that the weight initialization signatures of DeepSeek-V4-Flash and MiniMax-M3 match those of B200 or B300 chips with 148 streaming multiprocessors (SMs); DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ant Group's Ling-3.1-flash match GB200 or GB300 chips with 152 SMs. By contrast, earlier versions released by these four organizations (such as DeepSeek-V3 and Kimi K2.5), as well as the latest open-source models from Alibaba, Zhipu, Tencent, Xiaomi and StepFun, still show signatures of the previous-generation Hopper architecture (132 SMs).
The principle of the study lies in the mechanism by which the PyTorch framework calls Nvidia's cuRAND library to initialize model parameters: when the pseudorandom number generator Philox generates initial weights, it allocates parallel threads according to the number of GPU SMs, thereby leaving a specific arrangement pattern in the parameter matrix. During model training, rare tokens in the vocabulary or unused reserved tokens are rarely updated, so the embedding layer weights retain their initial state. By comparing about 6 trillion combinations of random seeds and hardware layouts, the researchers reverse-engineered the GPU architecture that generated these weights, and reproduced the initial values of some models on physical B300 and GB200 chips.
Nvidia's Blackwell series chips are currently subject to U.S. export controls and may not be exported to China without a license. The report estimates by model scale that the five main training runs above would collectively require about 12,000 Blackwell GPUs, with the largest single cluster (such as Kimi K3) needing 3,000 to 4,500 GPUs, and hardware costs exceeding $500 million. However, the report also notes that weight forensics can only reflect the hardware configuration at the model initialization stage; it cannot prove that the same cluster was used throughout subsequent training, nor can it determine whether the relevant chips are deployed in China or rented through overseas data centers, so it does not constitute direct evidence of violating export controls. The question is, DS V4.1 Flash isn't on this list—has it really switched to training on domestic chips? That's the change 421
Cable replacement grounds all U.S. flights
According to The Wall Street Journal, on November 14, U.S. Transportation Secretary Sean Duffy said that the system failure that grounded flights nationwide for more than four hours on Friday stemmed from routine maintenance work by defense and aerospace contractor L3Harris at a remote facility, where copper cables were being replaced with fiber optics. The Federal Aviation Administration (FAA) has opened an investigation, and L3Harris said it is actively cooperating.
Duffy noted that after L3Harris replaced a router during the construction work, the device sent up to 1,000 connection requests every four to five seconds to other routers in the air traffic control system, disrupting the transmission of aircraft communications and radar signals. FAA Administrator Bryan Bedford said there is currently no evidence that the failure was related to AI or a malicious attack, and that the agency issued the ground stop out of safety considerations.
The ground stop caused widespread flight delays across the country and disrupted aircraft and crew scheduling. About 9,000 commercial flights and 3,000 private jet flights were canceled on Friday; more than 6,000 flights were canceled on Saturday, and 530 had already been canceled in advance for Sunday. Major airlines such as Delta, United, and American have waived rebooking fees and are scrambling to adjust staffing. The FAA expects flight schedules to return to normal by Sunday evening.
Nationwide ground stops are extremely rare; the last one occurred in 2023 (also caused by a system failure), and before that during the September 11 terrorist attacks in 2001. Duffy used the incident to call on Congress to approve $50 billion in funding to comprehensively upgrade the U.S. air traffic control system, which still relies on 1980s technology. A conspiracy, there must be a conspiracy. 😏1
Over 100,000 sign petition demanding Jensen Huang's reinstatement
According to Business Insider, the resignation of Nvidia co-founder Jensen Huang has prompted widespread support from Silicon Valley executives, investors, and employees. As of the afternoon of November 14, the open letter Stand with Jensen, which calls for restoring Huang's position and ensuring due process, had gathered more than 108,000 signatures (organizers are cleaning up fake signatures), including more than 58,000 shareholders and over 5,600 current employees, while comments on support posts on the company intranet have exceeded 3,200. Industry figures including Musk, Y Combinator head Garry Tan, and Replit CEO Amjad Masad also signed.
Musk warned that many of Nvidia's supply chain partnerships depend on Huang's personal relationships rather than written contracts, and that his departure could cause key deals to be lost. Shareholders Altimeter Capital and Atreides Management wrote to the board on Friday, urging it to avoid a hasty leadership change and to keep open the possibility of Huang's return. Epic Games CEO Tim Sweeney compared the move to Apple ousting Steve Jobs in 1985, while Telegram CEO Pavel Durov and venture capitalist Elad Gil used the occasion to remind founders to be wary of board power.
A Forbes analysis pointed out that the reason Huang fell from power so quickly is that Nvidia adopted a traditional one-share-one-vote structure when it went public in 1999, rather than the dual-class (super-voting) system that became widespread in tech only after Google's 2004 IPO. Huang currently holds only 3.6% of Nvidia's shares and an equivalent proportion of voting rights, less than half that of institutional investors such as BlackRock or Vanguard, and the company had previously not formulated a succession plan because Huang opposed it.
According to the Financial Times, Huang is discussing with close associates the possibility of returning to helm the company. I hereby sign, demanding Jensen Huang's reinstatement. ꧁𝓙𝓮𝓷𝓼𝓮𝓷 𝓗𝓾𝓪𝓷𝓰꧂ But even if he returns, after this incident, the supply chain will no longer be the same as before 311
Streaming users complain it is harder to find the works they want to watch
According to Ars Technica, Gracenote, an entertainment data company under the market research firm Nielsen, recently released the "2026 State of Play" report. The report shows that although the number of global video streaming platforms and the scale of their content libraries continue to expand, more than half (51%) of surveyed users said that the vast amount of content and the scattered platforms have instead made it harder to find something to watch.
The survey covered 3,000 adult streaming users in six countries: the United States, the United Kingdom, France, Germany, Brazil, and Mexico. The data shows that as many as 83.4% of respondents do not always know clearly what they want to watch when they open a streaming service. Due to difficulties in content retrieval, 19.2% of users often give up watching and turn to other activities because they cannot find suitable content; another 27.4% of users, when looking for a specific show, often ultimately give up because they cannot confirm which platform it airs on.
The report points out that the core problem currently facing the streaming industry is not a lack of high-quality content, but platform fragmentation and lagging content discovery mechanisms. Although major streaming service providers are generally using user data to develop personalized recommendation algorithms in order to improve user stickiness and retention, existing user experience designs still fail to effectively help viewers locate target content within vast libraries.
English text is being prepared.
According to Electrek, Tesla recently opened personal orders for its driverless two-seat Cybercab to some invited customers, with a starting price of $29,990, but the simultaneously disclosed "Autonomous Driving Subscription Agreement" contains many harsh terms.
According to the subscription agreement, because the vehicle is not equipped with a steering wheel or pedals, owners must pay a monthly subscription fee of $299, otherwise the vehicle will be completely unable to operate. Calculated over a five-year use cycle, the purchase price plus 59 months of paid subscription totals $47,631 (the first month is free), and does not include insurance, registration, charging, and taxes.
The subscription agreement also sets strict restrictions on operating areas, usage permissions, and modifications. The Cybercab is limited to driving within Tesla-enabled Robotaxi geofences, and leaving the geofence will be technically blocked, while Tesla has the right to expand, shrink, or relocate operating areas at any time. At the same time, Tesla only needs to give one month's notice to adjust the subscription price. If an owner has three or more records of improper or unsafe use, autonomous driving service will be suspended for one week, during which the vehicle cannot be used and fees will not be refunded on a pro rata basis. In addition, Tesla can remotely access, restrict, or control the vehicle, and by default retains image data from the camera above the interior rearview mirror for 30 days; unauthorized changes to the car wrap, window film, or wheels will also lead to suspension of the subscription.
Although Musk has long promoted that owners can add their vehicles to a shared fleet to make money, the agreement clearly stipulates that the Cybercab is limited to personal non-commercial use, and without Tesla's written authorization, it may not be connected to Tesla's Robotaxi network or third-party ride-hailing, delivery, and rental platforms such as Uber.
Brief news for a quick look
Mark Gurman claims—
Apple is preparing to release its first smart glasses (internal codename N50) at WWDC in June 2027, and plans to ship them before the end of that year. The product's core selling point is combining Siri AI with the existing third-party app ecosystem on iPhone, allowing users to complete tasks such as ordering food, hailing rides, shopping, and buying tickets via voice commands without taking out their phones. To support this hardware plan, Apple added a new feature in this month's iOS 27.3 beta that lets Siri precisely control apps, and is expected to support official apps first in December this year, then fully open it to third-party developers next year. Developers do not need to develop software separately for the glasses; they only need to connect specific functions of existing iOS apps to Siri. The first-generation product does not have augmented reality (AR) image overlay display and must be paired with a new iPhone. The engineering prototype is slightly thinner than Meta's second-generation Ray-Ban smart glasses, is equipped with a processor customized based on the Apple Watch chip and a streamlined version of the visionOS 28 system, features a camera flush-mounted inside the frame, a Wi-Fi module, and a magnetic charging port at the end of the left temple, and supports wearing while charging;
In addition, Apple is stepping up development of an external-screen Face ID module and an action button for the second-generation iPhone Duo (codename V78), planned for release next fall. However, as the hardware finalization deadline approaches, whether Face ID can make it to next year's mass production remains uncertain;
Apple also plans to launch in 2027 a number of new products including AirPods equipped with cameras, a desktop HomeView with a robotic arm, and a MacBook Pro equipped with M7-series chips.
According to Now News, on November 14, due to suspicion that ticket QR codes had been stolen, the Hong Kong Disneyland website showed that system maintenance was underway, suspending ticket purchases and hotel bookings. If travelers need to buy tickets, they are welcome to go to partner travel agencies and platforms or the ticket office at the main entrance. It is understood that the park's tickets are sold online by Cityline. The park's Information Technology Department detected abnormal activity in the company's computer system and found that 17,000 QR codes had been stolen from the online ticketing system. The park said it discovered the cyber incident on Friday, has strengthened ticket verification procedures and reported to the police, will notify the Privacy Commissioner's Office, and strongly recommends that guests buy tickets directly through the park's official channels or through partner travel agencies and platforms, and not patronize sellers from unreliable sources, to avoid inconvenience and financial loss. Police said they received a report from the park and have classified the case as dishonest use of a computer, handing it to the Lantau District Crime Squad for follow-up. Some in the tech industry believe this incident was caused by hackers attacking the ticketing platform and suggest that companies strengthen precautions.
Recent updates from SSPAI
We have opened a special page to summarize and display the "Outstanding Author Selection" awards and results. "Outstanding Author Selection" is a mechanism implemented by the editorial team since 2023. After the end of each quarter, it selects existing authors who made outstanding contributions during that quarter, as well as emerging new authors with potential. Winners receive exclusive badges and prizes. Go to view
Limited-time offers on SSPAI membership, genuine software, and paid columns. Don't miss out
Good articles you may have missed
Aerial Switcher: Put aerial wallpaper schedules on your Mac
Nostalgia Notes | Sony CLIE: A high-end Palm OS PDA of the millennium
What to watch this week | 10 recent works worth watching
In the fast AI era, how did I make Mole a "slow product"?
Editor of this article:
@PlatyHsu
1 SSPAI friend has already charged