The Video Signal

Weights & Biases

CoreWeave Fully Connected 2026 Keynote: Introducing CoreWeave Forge | Built for Adoption

Published 2026-10-14 · Duration 1:51:33

Summary

The CoreWeave Fully Connected 2026 keynote focused on moving AI from the demo stage into reliable, scalable production environments. The central theme is the 'AI loop' (Run, Observe, Curate, Improve, Evaluate), which is necessary for continuous improvement of AI agents. Key announcements included the launch of CoreWeave Forge, a unified platform for managing the entire AI lifecycle, and the debut of the NVIDIA Vera Rubin NVL72, designed specifically for the demanding, multi-agent inference era. Industry leaders demonstrated how to apply these tools in high-stakes domains like finance (Capital One) and physical automation (Caterpillar).

Download summary

Key takeaways

  1. CoreWeave Forge: Unifying the AI Loop 22:06

    CoreWeave Forge brings all steps of the AI lifecycle—Run, Observe, Curate, Improve, and Evaluate—into one connected environment. This allows teams to move from a first agent to a best agent by ensuring context and evidence flow seamlessly between stages.

  2. NVIDIA Vera Rubin NVL72 Launch 27:08

    The Vera Rubin NVL72 is announced as ready for production, designed from the ground up for inference in the agentic era. CoreWeave operates the entire stack, ensuring consistency across all layers (power, cooling, firmware, network, and applications).

  3. Enterprise AI Evaluation Standards 38:49

    Capital One detailed their evaluation-first architecture, moving beyond legacy N-gram scorers (like BLEU/ROUGE). They utilize LLM-as-a-judge with Chain-of-Thought rubrics and track specialized metrics like 'actual state changes' (verifying API calls) to ensure trustworthy, auditable AI outputs.

  4. Agentic Failure Modes 30:59

    The complexity of agents introduces new failure modes: context/memory drift, behavioral traps (false claims of action), and systemic complexity (minor early errors compounding exponentially). Robust evaluation must account for these multi-step failure patterns.

Technical details

  • AI Agent Architecture & Performance 1426s

    Agentic workflows are complex, requiring multiple models (one for processing context, another for observing meaning) and tool calls. The process is a continuous inner loop of reasoning, consuming tokens, and performing actions, which is a data center scale problem.

  • CoreWeave Mission Control 1526s

    This system provides operational intelligence for AI workloads, monitoring everything from workload behavior to the physical state of the rack. It enables continuous optimization by identifying root causes (e.g., KV cache issues) and suggesting improvements (e.g., adjusting cache thresholds or scaling replicas) to boost throughput and efficiency.

  • Model Distillation and RL Rollouts 1326s

    To improve models, teams can use CoreWeave Model Distillation to create smaller, specialized models based on production traces. For Reinforcement Learning (RL), CoreWeave launched RL rollouts, allowing weights to be loaded straight into a running deployment without waiting for a full serving redeployment, accelerating the learning cycle.

  • Data and Infrastructure Composability 1628s

    The platform is designed to be composable, supporting six generations of NVIDIA accelerators (from Volta to Vera Rubin). This ensures that the infrastructure can adapt to evolving workloads (e.g., running both consumer-scale traffic and long-context coding agents) without requiring a full stack overhaul.

Mentioned resources

Channel & topics

Watch on YouTube · Back to latest

This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.