The Future of AI Infrastructure with CoreWeave

What It Covers

Corey Sanders, SVP of Product at CoreWeave, explains why AI workloads require purpose-built infrastructure distinct from general-purpose cloud, covering training optimization, inference architecture, the AI development loop, agentic workflows, and CoreWeave's ARIA research agent and Slurm-on-Kubernetes platform called Sunk.

Key Questions Answered

• •AI Infrastructure Specialization: Training workloads spanning hundreds to tens of thousands of interconnected GPUs require pre-planned, purpose-built infrastructure using specialized networking like InfiniBand or RoCE. General-purpose cloud's fungible, deploy-on-demand model actively slows AI workloads. Organizations running large training jobs should audit whether their infrastructure is co-designed for deep GPU interconnection rather than adapted from commodity compute.
• •GPU Straggler Detection: When a training job runs across thousands of GPUs, a single underperforming GPU degrades the entire job's output without obvious visibility. CoreWeave's straggler detection tooling identifies which specific GPU is slowing down. Teams running large-scale training should implement hardware-level observability that distinguishes hard failures from soft performance degradation before attributing slowdowns to model or data issues.
• •ARIA Research Agent Workflow: CoreWeave's ARIA (AI Research and Iteration Agent) continuously analyzes training experiments and recommends next steps, replacing manual review of line charts across multiple runs. Practitioners can receive experiment results and iteration recommendations via mobile overnight. This shifts the researcher's role from dashboard interpretation to judgment calls on agent-generated recommendations, compressing experimentation-to-production timelines.
• •Sunk: Slurm on Kubernetes: CoreWeave built Sunk (Slurm on Kubernetes) to combine Slurm's job scheduling, familiar to AI researchers, with Kubernetes' orchestration and infrastructure failure management. Sunk Anywhere extends this to other clouds and on-premises environments via Kubernetes operators. Teams already using Slurm for research workloads can adopt Sunk to reduce infrastructure management overhead without abandoning existing scheduling workflows.
• •The AI Development Loop: Production AI applications require continuous iteration across prompt tuning, model swapping, fine-tuning, and reinforcement learning rather than one-time deployment. CoreWeave's integrated stack connects inference traces via Weights & Biases, evaluation sandboxes, model registries with version lineage, and ARIA recommendations. Teams should structure development pipelines so production trace data feeds directly back into improvement cycles rather than treating deployment as a terminal step.