CoreWeave on AI-Native Infrastructure and the AI Engineering Loop
Corey Sanders, SVP of Product at CoreWeave, joins the Practical AI podcast to explain why AI workloads demand infrastructure that is fundamentally different from general-purpose public cloud. He contrasts CoreWeave's vertically integrated, AI-centric stack—optimized for the expensive, highly interconnected GPUs used in large-scale training—with traditional hyperscaler clouds where GPUs were simply added as an option. Key differentiators include specialized observability (such as GPU straggler detection), AI-aware storage caching, Kubernetes-based orchestration tuned for bare-metal GPU performance, and SunK, CoreWeave's Slurm-on-Kubernetes scheduler that is now available as a portable "SunK Anywhere" offering. Sanders argues that the real productivity gains come from closing an end-to-end "AI loop" spanning production inference, tracing (via the Weave platform built on Weights & Biases), evaluation, and continuous model improvement. He introduces ARIA, CoreWeave's AI Research and Iteration Agent, which analyzes experiment results and recommends next steps, and predicts that within five to seven years, AI engineering teams will outsize legacy cloud teams at most enterprises, with agent-driven interfaces replacing today's click-based UIs. He also discusses robotics workloads, the Monolith acquisition, and CoreWeave's commitment to multi-cloud portability—delivering open, Kubernetes-based services that customers can run on any infrastructure rather than locking them in.
Original sources
Podcast Practical AI
YouTube Practical AI
Video:The Future of AI Infrastructure with CoreWeave |Episode #364| ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.