Skip to content

Collections

Best Chain of Thought Episodes on AI Infrastructure & Hardware

What AI runs on: power, GPUs, networks, kernels and small models.

Every model runs on physical systems, and these conversations cover the layers underneath. Jeetu Patel starts with the limits on power, compute, network bandwidth and data center capacity. Anush Elangovan appears twice, first on AMD’s open-ecosystem bet and then on using AI agents to tune GPU performance. Maxime Labonne explains how Liquid AI designed models for phones and wearables, and Andy Masley closes by disputing a widely shared claim about AI water use.

  • 5 episodes
  • 5 guests
  • 5 hours of listening
  • 5 jump-to moments
  1. 1 EP 44 1 hr 18 min Transcript The Critical Infrastructure Behind the AI Boom | Cisco CPO Jeetu Patel Jeetu Patel, Cisco Cisco’s Jeetu Patel names infrastructure as the first constraint on AI: not enough power, compute, network bandwidth or data center capacity. He also explains why GPUs at training scale depend on the network that joins them. Jump to the moment · 11:39 · Infrastructure is the input →
  2. 2 EP 28 49 min Transcript AMD's Challenge to NVIDIA: The Open Ecosystem Bet | Anush Elangovan & Sharon Zhou Anush Elangovan & Sharon Zhou Recorded at AMD’s Advancing AI event, Anush Elangovan argues that an open ecosystem across hardware, networking and software moves faster than a closed one, and describes outside developers helping bring PyTorch to Windows on AMD hardware. Jump to the moment · 2:04 · Why open moves faster →
  3. 3 EP 72 53 min Transcript Tuning GPU Performance with AI Agents | AMD’s Anush Elangovan on ROCm 10 Anush Elangovan, AMD Anush Elangovan returns for ROCm 10, where agents handle installation and Hyperloom uses LLMs to explore the performance-tuning search space. He cites one case of optimizing about 14,000 models in a single pass. Jump to the moment · 25:07 · LLMs search the tuning space →
  4. 4 EP 43 53 min Transcript Beyond Transformers: How Liquid AI Is Rethinking LLM Architecture | Maxime Labonne Maxime Labonne, Liquid AI Maxime Labonne on designing Liquid AI’s models for phones and wearables: a hybrid with more short convolution layers than attention layers, which he says runs faster and keeps long-context memory use down without losing the quality of a pure transformer. Jump to the moment · 3:08 · More convolution than attention →
  5. 5 EP 45 59 min Transcript Debunking AI's Environmental Panic | Andy Masley Andy Masley, Effective Altruism DC Andy Masley argues that a book chapter overstated a 2027 projection by an order of magnitude. Using the water-consumption definition, he says the study the book cites projects global AI at about 5% of Britain’s water use rather than half, while noting that 5% is not nothing. Jump to the moment · 5:24 · Half of Britain or 5 percent →

More collections

All collections →

New here? Start with five → Browse by topic →