OpenAI's Pentagon deal costs them 2.5 million users. Meanwhile, the godfather of deep learning bets a billion dollars that every LLM company is wrong, and NVIDIA hands the agent-building playbook to everyone.
In this week's edition:
OpenAI's Pentagon deal triggers the biggest user revolt in AI history
A billion-dollar bet that the entire AI industry is building on sand
NVIDIA opens the agent-building playbook to everyone
Who's cutting thousands of jobs to feed the machine
The tools and tidbits you almost missed
OpenAI's Pentagon Deal Triggers Mass Exodus — Claude Takes the Crown
The AI industry's trust fault line cracked wide open this week. OpenAI signed a $200 million contract with the Pentagon on February 28. Within days, 2.5 million users pledged to cancel their subscriptions, and ChatGPT uninstalls spiked 295% in a single day.
Need to know:
The deal allows the Department of Defense to deploy OpenAI's models on classified networks for "all lawful purposes." Sam Altman says the agreement includes guardrails prohibiting domestic mass surveillance and requiring human oversight for autonomous weapons. Critics aren't convinced.
Anthropic CEO Dario Amodei refused the same terms, saying the company cannot "accede to their request." President Trump called Anthropic a "Radical Left AI company" and ordered federal agencies to phase out its technology within six months. Defense Secretary Pete Hegseth labelled Anthropic a "supply-chain risk to national security."
Claude shot to #1 on the U.S. Apple App Store. Anthropic reported a 60% increase in active users since January, with daily sign-ups quadrupling to over one million new users per day.
The #QuitGPT movement is building toward a physical AI accountability march scheduled for March 21, organised by the same coalition that demonstrated against Google DeepMind in 2025.
This is bigger than a product swap. When millions of people actively choose which AI they trust, they're choosing whose values shape the answers they get, the recommendations they see, and the information they act on. For businesses building visibility through AI channels, which model your customers use just became a strategic question.
Signals
Yann LeCun Bets $1 Billion That the AI Industry Got It Wrong
The godfather of deep learning thinks large language models are a dead end. In November, Yann LeCun walked into Mark Zuckerberg's office and told his boss he was leaving Meta. Four months later, his Paris-based startup AMI Labs closed a $1.03 billion seed round — Europe's largest ever.
Need to know:
AMI Labs is building "world models" — AI systems that understand physics and real-world interactions rather than predicting the next token. LeCun argues this is the path to genuinely capable AI.
The $3.5 billion valuation came from Cathay Innovation, Greycroft, Hiro Capital, and Bezos Expeditions. Day-to-day operations fall to CEO Alexandre LeBrun, previously founder of medical AI startup Nabla.
LeBrun's own prediction is telling: "In six months, every company will call itself a world model to raise funding."
If world models gain traction, the rules for AI visibility change completely. Today's game is about how LLMs interpret and cite text. A world model that understands real-world context could weigh trust signals entirely differently. The playbook hasn't been written yet.
Breakthrough
NVIDIA Drops Nemotron 3 — Open Models Built for the Agent Era
NVIDIA launched the Nemotron 3 family at GTC — open models designed specifically for agentic AI systems running at scale. The headline model, Nemotron 3 Super, packs 100 billion parameters but only activates 10 billion at any given time.
Need to know:
Three sizes: Nano (30B parameters, 3B active), Super (100B, 10B active), and Ultra (500B, 50B active). The hybrid mixture-of-experts architecture makes them fast enough for complex multi-agent workloads.
Nano is available now with a 1-million-token context window and 4x higher throughput than its predecessor. Super and Ultra land in H1 2026.
NVIDIA also released 3 trillion tokens of open training data plus safety evaluation tools. Early adopters include Cursor, Perplexity, Palantir, ServiceNow, and Zoom.
Open agentic models mean more AI systems making autonomous decisions — including which businesses to recommend, which products to shortlist, and which sources to trust. Perplexity is already on board. Every new agent built on Nemotron is another machine customer your business needs to be visible to.
The Roundup: The AI Cash Crunch
Oracle plans 20,000-30,000 layoffs to generate $8-10 billion for AI data centres. The cuts stem from a $156 billion OpenAI infrastructure deal requiring 3 million GPUs over five years. Banks are pulling back on financing. Implementation starts this month.
Meta weighs 20% workforce cut — roughly 16,000 employees — as AI spending balloons to $115-135 billion in 2026. If enacted, it would be Meta's largest reduction since the 2022 "year of efficiency." The company calls the report "speculative."
China's Moonshot AI quadruples valuation to $18 billion in three months, seeking $1 billion in an expanded round. The Kimi chatbot is gaining ground fast — another AI surface where brands either show up or don't.
Apple paying $1 billion a year for Google Gemini to power a rebuilt Siri arriving with iOS 26.4. When Siri starts answering questions properly, every iPhone becomes an AI discovery channel. If your business isn't visible to Gemini, it's invisible to a billion Apple users.
Legora raises $550 million Series D at $5.55 billion for its collaborative legal AI platform. Led by Accel with Benchmark and Bessemer. Legal tech is suddenly a top-tier AI vertical.
Mind Robotics closes $500 million Series A for AI-enabled industrial automation. NVIDIA's open robotics data (500,000 trajectories) and startups like this are converging on the factory floor.
Tool Shed
Krisp: Convert accented speech in real-time during calls using on-device AI processing with under 200ms latency and zero cloud storage.
Viktor: Build and manage AI automations natively inside Slack — context-aware scripts, scheduled tasks, and workflow triggers without leaving chat.
Aident AI: Automate open-world browser tasks described in plain English with visual confirmation steps for reliability on any website.
Robby: Uncover hidden revenue for home services businesses using AI to identify upsell opportunities and generate qualified leads for technicians.
Manicule: Build developer-facing technical documentation with an AI-native agency approach that keeps docs accurate as codebases evolve.
CalendarJet: Create branded scheduling pages on custom domains with multi-calendar sync, payment collection, and API webhooks built in.
Timelapse: Monitor real-time marketing performance metrics across channels with AI-generated insights that surface anomalies before they become problems.
NeedleVibe: Automate repetitive business workflows using visual AI pipelines that connect your existing tools without writing code.
Quick Bytes
A critical n8n vulnerability left 24,700 instances exposed to remote code execution — CISA added CVE-2025-68613 to its exploited vulnerabilities list with a March 25 federal patch deadline.
The Commerce Department published its evaluation of state AI laws on March 11, identifying which ones conflict with federal policy — setting the stage for unprecedented federal preemption of state AI regulation.
Washington state passed two AI bills this week: HB 1170 requiring AI disclosure and HB 2225 establishing chatbot safety protections for children.
Morgan Stanley warns an AI breakthrough is imminent in the first half of 2026, driven by unprecedented compute accumulation at America's top AI labs.
Apple unveiled the MacBook Neo at $599 with an A18 Pro chip — positioning a mass-market AI device for consumers who don't need a Pro machine.