Back to feed
News Story
NVIDIA AI Blog
1 sources

From Training to Production, NVIDIA and CoreWeave Close the Loop on Agentic AI

At its Fully Connected event, CoreWeave announced production availability of NVIDIA Vera Rubin NVL72 systems with Spectrum-X 102.4T Ethernet networking. Cognition, the applied AI lab behind Devin, is the first customer running production workloads on Vera Rubin, reporting 4.8x higher token throughput in early benchmarks. CoreWeave also launched CoreWeave Forge and will offer NVIDIA Vera, the first CPU built for AI agents.

SynthePulse Insight · AI deep readingMembers

CoreWeave Pushes Vera Rubin Into Production: Behind the 4.8x Throughput, the Closed-Loop Battle for Agent Infrastructure

Version 1 · 1 source

At the Fully Connected conference, CoreWeave announced that NVIDIA Vera Rubin NVL72 is entering production availability, with Cognition becoming the first customer to run the platform on production workloads and measuring up to a 4.8x token throughput improvement over GB200 NVL72. In the same announcement, CoreWeave also rolled out the Vera CPU for agents and Forge, which unifies a train-evaluate-improve closed loop. What is really worth questioning is not any single multiplier, but which parts of this "closed loop" narrative are verifiable deliveries and which remain vendor self-accounts.

  • CoreWeave announced that NVIDIA Vera Rubin NVL72 paired with Spectrum-X 102.4T Ethernet is entering production availability, making it one of the earliest cloud providers to deliver the platform to customers.
  • Cognition is the first customer to run production workloads on Vera Rubin, with its Devin training, reinforcement learning and production inference all running on CoreWeave.
  • Using GB200 NVL72 as the baseline and real software engineering tasks drawn from FrontierCode for early testing, Cognition measured up to a 4.8x improvement in total token throughput on the SWE-2 inference workload.
Open section navigationOne Announcement, Three Layers: Hardware Delivery, Customer Validation, Closed-Loop Platform

One Announcement, Three Layers: Hardware Delivery, Customer Validation, Closed-Loop Platform

At the Fully Connected conference, CoreWeave announced that NVIDIA Vera Rubin NVL72 systems paired with Spectrum-X 102.4T Ethernet are entering production availability, making it one of the earliest cloud providers to deliver the platform to customers. This is not an isolated product listing: in the same announcement, CoreWeave also said it will offer NVIDIA Vera CPU and launched CoreWeave Forge—a connected environment that unifies training, evaluation and improvement of models and agents on NVIDIA accelerated computing.

Put these three things together and the structure of the announcement becomes clear: the bottom layer is delivery of a new-generation rack-scale system, the middle layer is customer validation on production workloads, and the top layer is a toolchain that tries to feed "production behavior back into training" as a closed loop. The key to understanding this announcement is not memorizing every multiplier, but distinguishing which layer has already landed and which layer remains a vendor self-account.

Free for now

Read the full analysis

5 more sections of analysis, plus the full takeaway

Loading

Credibility boundary

All facts in this article come from a press-release-style article on NVIDIA's official blog, and statements involving partners such as CoreWeave and Cognition carry a promotional stance. Among them, Vera Rubin NVL72 production availability, Cognition as the first production customer, the first rack delivery timeline, the Forge launch, Dynamo supporting managed inference and RL Rollouts, and the Sandboxes isolation mechanism are confirmed facts; the 4.8x throughput, thousands of GPUs in nine months, the Vera CPU rack specifications and the 3x/1.7x tests, Forge's first customers, the Agent Lens and serverless RL gains, V100 still running, and the Ennoble Care case are all source claims. All performance figures lack independent third-party reproduction, and the 4.8x result is based on a specific workload and early testing and cannot be extrapolated.

Sources for this version

  1. From Training to Production, NVIDIA and CoreWeave Close the Loop on Agentic AINVIDIA AI Blog

Primary report

NVIDIA AI Blog

Primary source