1,000x cheaper AI power, Google's talent drain, and video games training robots
A startup claims 1,000x AI power reduction, Google keeps losing top researchers to Anthropic, and $320M bets that Fortnite can teach robots to think.
Subscribe free →A Databricks Veteran Thinks He Can Cut AI's Power Bill by 1,000x
Ali Ghodsi's departure from Databricks led to something wild: a startup claiming its architecture can slash AI's energy consumption by three orders of magnitude. The company's first public demo is Un0, an image-generation tool that reportedly replicates conventional AI output while running on a fraction of the power. If even a fraction of the 1,000x claim holds up at scale, this reshapes the economics of running AI entirely.
Why it matters: Power costs are the biggest ceiling on AI scaling — a credible path to 1,000x efficiency would be the most important hardware story of the decade.
Google's AI Brain Drain Keeps Getting Worse
Two more top Google AI researchers — Jonas Adler and Alexander Pritzel — are heading to Anthropic, adding to a growing list that includes Noam Shazeer and Nobel Prize winner John Jumper. The exits span Google DeepMind and Google Research, suggesting this isn't a one-team problem. When your best people keep walking out the door to a competitor, no amount of compute budget fixes the morale math.
Why it matters: Talent concentration at Anthropic is accelerating at exactly the moment the frontier model race is most competitive.
General Intuition Just Raised $320M to Train AI on Millions of Hours of Gameplay
General Intuition — now valued at $2.3B — is betting that gameplay data from titles like Fortnite is the missing ingredient for teaching AI agents real-world physical intuition. The theory: action-dense video game footage gives models a richer signal about cause, consequence, and physical dynamics than curated robotics datasets. The $320M Series raise will go toward scaling that training pipeline.
Why it matters: If gameplay-trained models can generalize to physical robotics, it unlocks a virtually unlimited supply of training data without expensive real-world collection.
Amazon Drops Another $13B on AI Infrastructure in India
Amazon's latest India commitment brings its total announced investment in the country to a staggering figure, as AWS races to plant data center flags ahead of Microsoft and Google. India's combination of surging enterprise cloud demand, government AI ambitions, and a massive developer base makes it the most contested AI infrastructure market outside the US and China right now.
Why it matters: Whoever owns India's AI infrastructure layer in the next five years will have a structural advantage in one of the world's fastest-growing tech economies.
Companies Are Rationing AI Tokens After Employees Go Wild
The tokenmaxxing era is over. Enterprises that handed out unlimited AI access are now discovering employees burning $30,000/year on tasks like "summarize my calendar" — and the backlash is a wave of usage caps, approval workflows, and spend dashboards. Rippling CEO Parker Conrad is among those publicly surfacing the data: the distribution of AI ROI across employees is extremely uneven.
Why it matters: How companies govern AI spend will define whether the productivity narrative survives contact with CFOs — and it's shaping a new category of AI cost-management software.
Engineers Were Supposed to Be AI's First Casualties. They're Actually Thriving.
New data from SignalFire shows engineers are making up a *larger* share of new hires even as AI dominates the tech layoff narrative. The pattern suggests AI is so far augmenting engineering output rather than replacing headcount — and that companies are competing harder for people who can build and deploy AI systems, not fewer of them.
Why it matters: The "AI kills coding jobs" thesis is getting its first real data challenge, and the answer has significant implications for how we think about AI's labor market impact.
Quick Hits
- →Cerebras stock tanked after its first post-IPO earnings report, with investors spooked by narrower gross margin guidance — the CEO says they got it wrong. TechCrunch
- →Adobe acquired AI enhancement tool maker Topaz Labs and plans to bake its upscaling and restoration tech into Photoshop, Premiere, and the rest of Creative Cloud. TechCrunch
- →The math on OpenAI's custom Jalapeño ASIC chip (built with Broadcom) shows just how much inference cost reduction it needs to hit its financial targets. AI News
- →Europe is pushing back hard on the US MATCH Act, which would restrict exports of older-gen chip tools to China — ASML's CEO is among the loudest critics. TechCrunch
- →a16z backed Netris with a $15M Series A to help AI neocloud operators get their infrastructure live faster — unsexy but critical picks-and-shovels infrastructure. TechCrunch
AI TLDR