莫 大

Google Cloud just said the quiet part out loud at Goldman's tech conference: AI servers pay for themselves in under two years on average. Under one year if you run Google silicon. Thomas Kurian gave the number. An hour earlier, Hock Tan at $AVGO said the industry spends ~$200B/year producing inference tokens against ~$150B of model revenue. The industry book is underwater. Google says its own kit pays back inside twelve months. Both can be true. Kurian's edge claim is vertical integration. 2.7x better price-performance on training. 80% better on inference. 30% better on CPUs. Chip and model designed together. The second disclosure is the $NVDA problem. Google is now selling TPUs as capex. Customers buy the boxes, run them in their own buildings. There's also a neocloud line with Blackstone. Kurian's reason was blunt: when the customer owns the hardware, Google doesn't spend the capex, doesn't run the power or the floor. Revenue without the buildout. That's a merchant silicon business. Kurian said Google's accelerator and CPU book is more than 2x the next-largest hyperscaler's. Same session: 17 product lines above $1B. $100M+ deals up 2x YoY and QoQ. Customers who commit typically spend 50%+ above the commit. 90% of the Fortune 100 on Gemini Enterprise. The AI trade has been who sells the compute. Google is arguing it can sell the chip, rent the chip, and sell the model that runs on it. Watch Cloud margins next quarter. That's where a capital-purchase mix stops being a conference line. $GOOGL $NVDA $AVGO