Catalyst with Shayle Kann Live from Transition-AI 2026: Inside Google’s massive AI CapEx
28 snips
Apr 23, 2026 
Amin Vahdat, Google’s chief technologist for AI infrastructure who designs data centers, chips, and power systems. He talks about Google’s massive 2026 CapEx and the shift from training to distributed inference. Conversations cover rethinking reliability to favor more compute, using on-site power as a bridge, microgrids and software control, and co-designing chips, buildings, and models.
AI Snips
Chapters
Transcript
Episode notes
Inference Doesn't Need Gigawatt Data Centers
- Inference workloads generally don't require gigawatt-scale individual data centers and can run effectively in much smaller deployments.
- Amin notes racks trend toward hundreds of kilowatts and that serving can be done in tens of megawatts with co-located compute, storage, and networking.
Capacity Will Shift From Training To Serving
- We're entering an age where most capacity will shift from training to serving as models proliferate and efficiency improves.
- Amin compares it to search: compute once dominated index-building but quickly shifted to serving the index at scale.
Reliability Tradeoffs Shift With Compute Cost
- High reliability requirements grew when compute was a small cost, but with compute now dominant, customers often prefer more capacity over ultra-high availability.
- Amin says internal customers will trade fewer nines for double capacity in many cases.