CoreWeave deploys Nvidia Vera Rubin racks to scale agentic AI infrastructure
CoreWeave, Nvidia, and Dell executives recently discussed the infrastructure requirements for agentic AI workflows. The conversation highlighted how the integrated rack-scale systems of the Nvidia Vera Rubin platform aim to improve inference efficiency and throughput for continuous AI reasoning tasks.
Key Takeaways
- Vera Rubin platform delivers 10x higher inference throughput per watt and reduces costs to one-tenth per million tokens compared to Blackwell.
- Integrated NVL72 racks unify 72 Rubin GPUs and 36 Vera CPUs to operate as a single distributed computer via high-speed interconnects.
- CoreWeave's 'Mission Control' software manages compute, network, and storage as a unified system rather than isolated deployment targets.
- Proprietary hardware like the 'Racky' rack manager and 'Valvey' assembly coordinate liquid cooling and power sensors for 100,000-GPU clusters.
Why It Matters
The shift toward agentic AI changes the fundamental unit of compute from the individual server to the integrated rack. For streaming and enterprise video providers, this reduces the economic barriers to deploying real-time agents for automated editing, metadata tagging, or content moderation at scale. By tighter coupling of hardware and orchestration software, infrastructure providers are moving into a 'continuous operations' model where the system repeatedly evaluates and optimizes agent performance based on production traces. Watch for early 2027 performance data as CoreWeave scales these 3nm-based Vera Rubin deployments to competitive enterprise customers.
Additional Context
The push into agentic infrastructure follows a period of aggressive expansion for CoreWeave. According to Forbes in January 2026, Nvidia invested $2 billion in CoreWeave, valuing the cloud provider's strategy to build out 5 gigawatts of 'AI factories' by 2030. This investment deepens a relationship where CoreWeave serves as a primary deployment partner for Nvidia’s bleeding-edge hardware, recently reaching a $5 billion annual revenue run rate faster than any other cloud provider in history per Constellation Research in February 2026. This scale is supported by a massive order backlog, which TIKR reported reached approximately $99.4 billion as of March 2026, anchored by deep commitments from Meta, OpenAI, and Anthropic. Technically, the Vera Rubin platform represents a major leap in silicon density. Per SemiAnalysis in February 2026, the architecture moves to TSMC's 3nm process, allowing for 336 billion transistors and the introduction of HBM4 memory. These specs allow for roughly 50 PetaFLOPS of FP4 inference throughput per GPU—a 2.8x improvement over the Blackwell generation—and 22 TB/s of memory bandwidth to alleviate the bottlenecks common in large-scale reasoning tasks. Beyond individual chips, the transition to the Rubin CPX reflects a trend toward disaggregated prefill and decode in inference pipelines, allowing providers to segment workloads for maximum cost efficiency. This modularity is a direct response to the massive capital requirements of the AI era; per CIO Dive in January 2026, total industry spending on AI infrastructure exceeded $430 billion in 2025 alone.
Read full article at siliconangle.com
Get this in your inbox → Subscribe
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source