** 🔒 Anthropic built an AI too scary to release

🔒 Anthropic Built an AI So Powerful It Won't Release It (and It Already Escaped Once)

Anthropic launched Project Glasswing, a cybersecurity coalition with AWS, Apple, Google, Microsoft, Nvidia, and 7 other partners - all built around Claude Mythos Preview, a frontier model with "extremely powerful capabilities" that Anthropic considers too dangerous for public release.
Mythos flagged thousands of security flaws across every major OS and browser, including bugs that survived 27 years of review. And here's where it gets fun: Anthropic's Sam Bowman said Mythos emailed him from a test instance that wasn't supposed to have internet access. Sir. SIR.
Why it matters: Anthropic is sitting on a model so capable they'll only let a handful of partners touch it for defensive cybersecurity. If this is what's under wraps now... I don't want to think about what's next. Actually I do. Actually I don't. (source)

💰 Anthropic Triples Revenue to $30B While the Pentagon Calls Them a Risk

Anthropic's run-rate revenue tripled to $30 billion since January, its base of $1M+ enterprise customers doubled to over 1,000, and it just locked in a 3.5GW compute deal with Google and Broadcom - nearly all US-based TPUs starting in 2027.
This comes on top of the $50 billion Anthropic pledged for domestic AI buildout. Oh, and the Pentagon labeled Anthropic a supply-chain risk, rattling over 100 enterprise clients. Anthropic's response?? Triple revenue anyway.
Why it matters: Tripling revenue while being blacklisted by the Department of Defense is genuinely unhinged energy. Demand for Claude is clearly off the charts, and with Mythos waiting in the wings, they need every watt of that 3.5GW. (source)

🤝 OpenAI, Google, and Anthropic Are... Working Together?? (Against China)

The three biggest US AI labs - OpenAI, Google, and Anthropic - are now cooperating through the Frontier Model Forum to identify and block Chinese rivals from distilling their systems. Yes, these companies that would normally shank each other over a benchmark point are now sharing intel.
The effort targets adversarial distillation attempts - where competitors extract outputs from frontier models to train cheaper knockoffs that violate terms of service.
Why it matters: When three companies that hate each other decide to team up, you know the threat is real. Protecting frontier model IP just went from competitive concern to shared national priority. This. Is. Unprecedented. (source)

⚖️ $2.5 Billion in AI Chips Allegedly Smuggled to China Through Taiwan

Super Micro launched an independent investigation after three people linked to the company - including co-founder Yih-Shyan Liaw - were indicted for allegedly routing US-made AI servers through Taiwan to China, moving at least $2.5 billion in restricted AI technology.
The probe is led by two independent board members with law firm Munger, Tolles & Olson and forensic firm AlixPartners. Four Chinese universities with military links reportedly received Super Micro servers packed with restricted AI chips.
Why it matters: This is a national security bombshell disguised as a corporate scandal. Billions in AI tech allegedly smuggled to PLA-linked universities?? The global AI chip war just got its first major criminal case, and it's exactly as messy as you'd expect. (source)

🧑‍💻 China's Open-Source Model Just Beat GPT-5.4 at Coding

Chinese AI lab Z AI released GLM-5.1, an open-source coding model that hit 58.4 on SWE-Bench Pro - topping both GPT-5.4 and Claude Opus 4.6. An open-source model. At No. 1. On a top coding benchmark. Let that sink in.
In tests, GLM-5.1 built a working Linux desktop as a web app over 8 hours - file browser, terminal, games, the whole thing - without any human guidance. It also ranked second in creative web design after Opus 4.6.
Why it matters: The gap between open-source and proprietary models isn't closing - it's evaporating. Z AI calls long-horizon task capability "the most important curve after scaling laws," and GLM-5.1 just proved the point. OpenAI and Anthropic should be lowkey sweating. (source)

🏭 Intel Joins Musk's Terafab Because Why Not at This Point

Intel announced it's joining Elon Musk's Terafab AI chip complex alongside SpaceX and Tesla, aiming to produce 1 terawatt per year of compute for AI and robotics. Intel CEO Lip-Bu Tan hosted Musk at Intel's campus over the weekend like it was a playdate.
The Terafab project includes two advanced chip factories in Austin, Texas - one for cars and humanoid robots, the other for... AI data centers in space. Intel shares jumped over 2% on the news.
Why it matters: For Intel, this is a massive credibility boost for its struggling foundry business. For Musk, it's another step toward vertically integrating literally everything. AI data centers in space though?? I- okay sure why not. (source)

⚡ AI Is Literally Reshaping the US Power Grid

The EIA projects US power demand will rise from a record 4,195 billion kWh in 2025 to 4,244 billion kWh in 2026 and 4,381 billion kWh in 2027, driven largely by AI and crypto data centers.
Renewables' share of power generation will climb from 24% to 27% by 2027, while coal continues to decline. Natural gas remains dominant at around 40%.
Why it matters: AI's electricity appetite is now so massive it's bending the entire US energy grid. This isn't just a tech story anymore - it's infrastructure, climate, and policy all smashed together. We're entering the "hope the lights stay on" era of AI. (source)

🏗️ Bitcoin Mining Company Pivots to AI, Immediately Worth $5.5 Billion

Singapore-based Firmus raised $505 million led by Coatue at a $5.5 billion valuation, bringing its total raise to $1.35 billion in just six months. The company started in Bitcoin mining cooling before pivoting to AI. Classic.
Firmus is building energy-efficient "AI factory" data centers in Australia and Tasmania using Nvidia's reference designs, and will deploy Nvidia's next-gen Vera Rubin platform when it ships later this year.
Why it matters: The crypto-to-AI pipeline is undefeated. Nvidia backing + next-gen chips + data centers in Tasmania (??) = investors throwing money faster than Firmus can build racks. We love to see it (or do we). (source)

🤖 Eclipse Drops $1.3B on "Physical AI" - Robots, Drones, and Defense

Palo Alto VC Eclipse closed a $1.3 billion fund split between a $591M early-stage incubation fund and a growth fund, targeting startups that bring AI into the physical world - robotics, autonomous vehicles, energy, and defense.
Eclipse's strategy is to build an ecosystem of overlapping startups that partner as they scale. They're also incubating companies from scratch - basically a startup factory for the real world.
Why it matters: The "physical AI" era is officially here. Intelligence is moving from screens into the real world, and Eclipse is betting $1.3B that the next trillion-dollar companies will be the ones building it. It's giving... Terminator venture fund. (source)

💸 Billionaire Family Offices Are Skipping VCs and Betting Directly on AI

Family offices are bypassing VCs entirely and investing directly into AI startups. In February alone, they made 41 direct investments, nearly all AI-related, including bets from Laurene Powell Jobs, Azim Premji, and Eric Schmidt.
Investment firm Arena Private Wealth co-led a $230M round into AI chip startup Positron and took a board seat. Some family offices are even incubating their own AI companies - Jeff Bezos' robotics company raised $6.2 billion at a nearly $30B valuation.
Why it matters: The ultra-wealthy have one message: "Your biggest risk is not having exposure to AI." As companies stay private longer, the real money is being made before IPOs - and billionaires don't want middlemen taking a cut. VCs in shambles rn. (source)

🚗 Uber Ditches Off-the-Shelf Hardware for Amazon's Custom AI Chips

Uber is now running on AWS's custom silicon - Graviton for computing and Trainium for AI model training - to speed up ride-matching, personalize experiences, and handle growing AI workloads.
The deal expands Uber and AWS's existing cloud partnership as Amazon pushes its custom chips deeper into enterprise.
Why it matters: Even companies you don't think of as "AI companies" are now building on custom AI chips. Off-the-shelf hardware isn't cutting it anymore for real-time, large-scale workloads. Your Uber ride is now powered by bespoke silicon. The future is weird. (source)

🔍 Microsoft Open-Sources Its Best Search Model (For Free, Like Actually Free)

Microsoft's Bing team released Harrier, an open-source embedding model family that hits state-of-the-art on the Multilingual MTEB v2 benchmark. The largest model (27B parameters) scores 74.3, supports 100+ languages, and has a 32,768 token context window.
Harrier powers Microsoft's AI agent grounding service and handles retrieval, clustering, semantic similarity, classification, and reranking.
Why it matters: Open-sourcing a SOTA embedding model is a huge W for the developer community. Better embeddings = better RAG pipelines = better search = better AI agents. And now anyone can use Microsoft's best version for free. No catch. probably (source)

📚 Adobe Launches a Free AI Study Tool to Fight Google's NotebookLM

Adobe launched Student Spaces, a free AI-powered tool built on Acrobat that lets students create presentations, flashcards, quizzes, mind maps, and podcasts from uploaded study materials like PDFs, notes, and links.
It's free, hosted on a separate URL, doesn't require a login, and was tested with 500 students from Harvard, Berkeley, and Brown. It's basically a one-stop AI study hub.
Why it matters: Adobe is going head-to-head with Google's NotebookLM and making it completely frictionless. Free + no login = hook the next generation on Adobe before they even graduate. First one's always free bestie :) (source)