TBPN FULL INTERVIEW: Dylan Patel Says We’re Still Underestimating AI
153 snips
Feb 3, 2026 
Dylan Patel, founder and CEO of SemiAnalysis and semiconductor/AI infra analyst, walks through chips, data centers, and geopolitics shaping AI. He discusses data centers in space and cooling challenges. He compares specialized chips like NVIDIA, Cerebras, and Google TPUs. He debates capacity, power, supply-chain bottlenecks, export controls, and how AI reshapes finance and products.
AI Snips
Chapters
Transcript
Episode notes
Space Data Centers Face Reliability Limits
- Putting large AI clusters in space is plausible as launch and solar costs fall, but design constraints remain profound.
- Satellites face size, reliability, and servicing limits that break assumptions about huge fully connected racks in orbit.
NVIDIA's Multi-Chip Strategy
- NVIDIA is diversifying beyond one-size-fits-all GPUs toward specialized chips across different points on the Pareto curve.
- Multiple architectures (GPUs, CPX, Grok chips) reflect uncertainty about which hardware optimally serves future AI workloads.
Pay For Faster Completion, Not Strict Latency
- Prioritize latency-insensitive placement less than end-to-end inference speed improvements for long-running agent tasks.
- Pay up for faster hardware when a job completes 10x faster and that time reduction is worth the extra cost.