Live from Transition-AI 2026: Inside Google’s massive AI CapEx

In this live episode from Transition-AI 2026, Shayle Kann interviews Amin Vadat, Google's Chief Technologist for AI Infrastructure, about the company's massive $175–185 billion capital expenditure plan for 2026—equivalent to Hungary's GDP, five to seven times U.S. transmission infrastructure spending, or seven NASA budgets. The conversation dives into the evolving role of scale in AI infrastructure, particularly as the industry shifts from training to inference. Vadat argues that while gigawatt-scale data centers are essential for training, inference can be efficiently handled at much smaller scales (tens to hundreds of megawatts), especially as older training capacity is repurposed. He challenges the traditional need for extreme reliability (four nines), noting that with compute now a dominant cost, customers often prefer more capacity at lower reliability. Google is exploring behind-the-meter power, demand response agreements with utilities, and mobile bridge power to accelerate deployment while reducing grid strain. The discussion also touches on the future of microgrids, workload differentiation, and the critical role of vertical integration in enabling flexible, co-optimized systems. Vadat ranks labor, power, and chips as equally critical rate limiters, with no single bottleneck easily solvable. Looking long-term, he emphasizes continuous cost optimization through co-design of hardware, software, and data center architecture to reduce future CapEx by up to 40%.