Good morning. There comes a moment in every billionaire’s life when the money is no longer enough. You need a blessing from the man in mittens who yells about billionaires for a living. Sam Altman reportedly asked to meet with Bernie Sanders after the senator proposed a 50% government stake in frontier AI labs. The profits would fund a sovereign wealth fund for the public. OpenAI floated a similar idea before, just without the part where Uncle Sam takes half the company. Altman wants the working class pope to bless the machine built to eat working class jobs.
Nvidia is turning AI agents into robot trainers. Teaching robots is painfully slow because researchers first have to build world models where robots can stumble, crash, and fail without causing chaos in real life. Nvidia’s new physical AI skills help build practice worlds faster. Agents can take driving footage and quickly reconstruct 3D roads for self-driving systems to train on. For factory inspection, they can generate fake defects so vision AI models learn what real problems look like before they happen. With robotics skills, AI agents can set up practice sessions, run simulations, and tune performance for the next round. Nvidia built all this around Cosmos 3 so physical AI can learn new skills without lab teams creating every scenario by hand. Nvidia already made itself the default toll booth for AI training. Now it wants the same role for robots. (Nvidia)
Google wants to pay you not to use electricity. It signed a deal with Voltus to build a virtual power plant for data center demand. VPPs sound absurd, like expecting a simulated power plant to keep the lights on. The reality is weirder. Voltus links homes, EV chargers, thermostats, batteries, and businesses into a flexible network so power can be shifted to data centers when the grid gets tight. That freed up capacity lets Google pull more from the grid without waiting years for new power plants. The hard part is finding people willing to give up control of their electricity so AI can jump the line. Maybe it would be easier for Google to subsidize solar panels and batteries. (MIT Tech Review)
Capture live traffic, fine-tune and optimize, then deploy your own checkpoints to dedicated GPU endpoints.
Choose hardware, set scaling limits, and select region. Stable latency, predictable cost, clear data residency.
👀 closer look
Researchers built an AI worm that learns as it spreads. Old school worms attack by repeatedly exploiting the same weakness. This one is smarter. It studies each machine it hits and builds a custom attack plan. If one approach fails, it tries another. When researchers tested it in a sandbox corporate network, the AI worm exploited an average of 23 machines and spread itself to 20. It carries an open weight model, so when it finds a GPU, it hijacks the compute to power the next wave. Somewhere, a CISO is begging ransomware to go back to being stupid. (arXiv)
Sam Altman is discovering fame doesn’t pay the compute bill. When OpenAI kicked off the AI boom, it looked like the obvious winner. But first mover advantage isn’t a moat. Anthropic doubled down on coding and enterprise customers, while OpenAI got distracted by side projects like Sora. Now Anthropic has passed OpenAI’s valuation and may beat it to the IPO starting line. That’s bad timing for a company still trying to convince investors it can grow fast enough to cover massive compute contracts. OpenAI might own the brand, but the money is moving to whoever wins corporate budgets. ChatGPT could end up being the Lyft of AI. Not exactly what investors want to hear before an IPO. (Sherwood)
Meta wants $200 a month for its version of OpenClaw. The product is called Hatch, and it’s supposed to be a personal AI assistant that builds tools and handles tasks. That’s an odd pitch from a company with no real productivity suite. Google has Workspace. Microsoft has Office. Meta has Facebook, Instagram, and WhatsApp, which might help if your job involves swiping left on Reels and arguing with strangers. Meta has been testing Hatch with Claude, but plans to shift to its own Muse Spark model at launch. The goal is clear. Meta needs new revenue since nearly all its business comes from ads while its AI spending keeps climbing. Personal superintelligence sounds a lot cooler than asking people to chip in for data centers. (The Information)
fun stats
📊 51%. US companies now paying for AI subscriptions, up from 47% in January. OpenAI leads at 35%, Anthropic trails closely at 31%, while Google and xAI are still under 7% combined.
🌊 $59 billion. Combined net worth of 19 new AI billionaires minted in the second wave of the AI boom. The group includes a published poet, college dropouts, and tech leaders in their 20s.
👭 74%. Consumers who told Accenture they’d trust a personal AI agent more than their best friend to make a purchase. GenAI now beats social media on trust, and 36% call it “a good friend.”