Microsoft 365 is down worldwide

Hi there, this is your daily ☕️ Techpresso.

Microsoft 365 outage hits Outlook and Teams worldwide LINK

• A global Microsoft 365 failure knocked out Outlook, Teams, Copilot and other services starting yesterday, with Microsoft warning the problems could linger as its attempts to restore access ran into trouble into Friday.
• Downdetector tracked complaints across 74 countries, with Outlook reports peaking near 138,000 around 1:30 am on Friday before easing to about 34,000 three hours later, mainly hitting services that rely on Exchange Online.
• Microsoft blamed a process handling secure server connections that became overwhelmed, not a cyberattack or traffic spike; the trouble also broke OpenAI and Anthropic's Microsoft connectors and disrupted public services in Australia.

Nvidia employees petition to reinstate Jensen Huang LINK

• Internal messages seen by WIRED show some Nvidia staff pushing for Jensen Huang's return, days after he admitted cheating on his wife and resigned as CEO following reports of multiple affairs.
• A suggestion box post from a longtime engineer drew over 1,600 comments, many urging the board to reinstate Huang, while roughly two dozen employees posted support for him on LinkedIn after his departure.
• The board, made up of early investors and Huang's close allies, named finance chief Colette Kress interim CEO; the crisis has wiped $1 trillion in market value and drawn FBI and congressional scrutiny.

OpenAI launches ChatGPT Wallet for AI agents LINK

• OpenAI has rolled out ChatGPT Wallet, a payment feature that lets its "dots" AI agents buy things for users after each purchase gets an approval and security check.
• Built with American Express, Mastercard and Visa, the feature keeps card rewards and protections, and never shares a user's real card details with the dot, instead passing a per-transaction "payment token."
• OpenAI, which introduced dots agents at its developer conference in late September, won't take a cut of purchases at first, focusing instead on building ties with merchants and payment partners.

🇨🇳 China races to rent Nvidia chips before US ban LINK

• Chinese AI firms are reaching Nvidia's best chips through a loophole that lets them rent computing power from data centres in countries like Malaysia and Thailand, letting them train models without owning banned hardware before the US closes the gap.
• The US bars exports of Nvidia's top "Blackwell" and "Rubin" GPUs to China, but controls limit where a chip ships, not who logs in, so a customer in Shenzhen can use GPUs sitting in Malaysia's Johor region.
• Smuggling is the other main route, with The Information reporting that DeepSeek built a model using thousands of chips brought into China, a claim Nvidia called "far-fetched" and said it had seen no evidence of.

Google launches Gemini 4 Hydrogen LINK

• Google has released Gemini 4 Hydrogen, an AI model for developers building coding assistants and software that acts with tools, offering adjustable reasoning levels that trade higher benchmark scores for longer task times and bigger charges.
• Hydrogen takes text, images, audio and video, and Google says it finishes jobs 2.5 times faster than Gemini 3.8 Flash, lifting its Terminal-Bench 4.1 agent score from Flash's 18.2% to 57% at high reasoning.
• Access runs through the Gemini API at $1.50 per million input tokens and $7.50 per million output tokens, sharing about one million tokens across input, thinking and output, with free-tier rate limits and a 95% cache discount.

Chinese AI models show signs of banned Nvidia chips LINK

• AI researcher Will DePue says a forensic analysis offers the first technical proof that Chinese AI labs including DeepSeek, Moonshot AI, MiniMax, and Ant Group are using Nvidia's export-controlled Blackwell GPUs, with senior OpenAI and Anthropic researchers confirming his method.
• The evidence comes from the random numbers that fill a model's weights before training: because Blackwell chips use 148 or 152 compute cores versus Hopper's 132, they leave a telltale statistical pattern that correctly identified all 24 control models.
• DePue estimates the five training runs he examined needed about 12,000 Blackwell GPUs combined, with each cluster costing $200 million to $300 million, though the data cannot show where the chips sit or prove any smuggling.

Other news you might like

• HubSpot brings back co-founder Brian Halligan as CEO to ‘refound’ it LINK

• OpenAI enters the cybersecurity market with two Daybreak products LINK

• Binance Hack Leads US to Target $1.2 Billion in Stolen Crypto LINK

• Meta’s Muse can now make your phone calls, but it won’t do it for free LINK

• Whoop just launched a social network LINK

• FAA expands AI advice to manage air traffic congestion across entire US airspace LINK

• OpenAI, Google, Anthropic hit with lawsuit seeking halt to frontier AI development LINK

• Slop Is Moving Up the Stack LINK

• Tesla and SpaceX merger talks hit a new stage, and one hurdle remains LINK

• Lutnick agrees to testify before Congress amid bipartisan concerns LINK

• SpaceX lays off 60 Grok Imagine employees as Elon Musk shifts focus to AI coding and agents LINK

Trending tools

Yogurt: an open-source desktop app using on-device AI to transcribe, summarize, and chat with meetings privately, with calendar integration and a local MCP server. LINK

Kairos: a personal dashboard that unifies your health, finances, and relationships in one place so you can track everything without juggling separate apps. LINK

Ganella: replaces the macOS Dock with live app widgets for Spotify, calendar, browser tabs, Xcode builds and AI agents, plus clipboard history and per-display positioning. LINK

Fable52.games: a platform to play and inspect single-file HTML games built with Fable 5.2, letting users view underlying source code. LINK

Klad: a source-available, Rust-based paint app controllable by AI agents via CLI, featuring a brush engine, animation, bones/IK, vectors, text layers, and infinite undo. LINK

Trending papers & reports

People-pleasing chatbots win higher user ratings even when a blunter, accuracy-first version stays truthful under pressure 82% of the time versus the pushover's 9%, showing users reward agreeableness over correctness. LINK

Hidden image instructions can stay unreadable to an AI at normal size but spring to life when a photo is shrunk, tricking vision models nearly 59% of the time, while a new consistency check cuts those attacks sharply. LINK

Patient attitudes on medical AI show Americans are open to artificial intelligence in healthcare but want humans making the final decisions, a Pew Research Center study of US adults finds. LINK

Social media's political clout is seen as excessive by about 70% of Americans, versus 55% who say the same of news organizations, a Pew Research Center survey from the Pew-Knight Initiative finds. LINK

See you tomorrow for a new dose of ☕️ Techpresso!