#

Key takeaways

• CoreWeave announced Vera Rubin NVL72 systems on CoreWeave Cloud.

• Cognition is named as the first customer running production workloads on this hardware.

• CoreWeave will also offer NVIDIA's Vera CPU for agent workloads.

• Forge combines training, evaluation, and improvement tools for models and agents.

• The source reports no Moroccan deployment or access.

What NVIDIA and CoreWeave announced

NVIDIA reported on September 30 that CoreWeave announced availability of Vera Rubin NVL72 systems on CoreWeave Cloud with Spectrum-X 102.4T Ethernet. The announcement positions the hardware for both training and production use. It also says CoreWeave will offer NVIDIA's Vera CPU for agent workloads.

The source frames this as a broader infrastructure and software update. It is not presented as a general market claim. It is a specific product and platform announcement.

Production use starts with Cognition

Cognition, the maker of the Devin software engineering agent, is identified as the first customer running production workloads on this hardware. The source also says Cognition compared Vera Rubin NVL72 against a GB200 NVL72 baseline. It used sampled software engineering tasks for that comparison.

NVIDIA says the early SWE-2 inference tests showed up to 4.8 times more total token throughput. The source is careful here. It says this is a company-reported workload measurement, not a universal benchmark.

CoreWeave Forge and related tools

The announcement also includes CoreWeave Forge. Forge is described as an environment for training, evaluating, and improving models and agents using NVIDIA accelerated computing. It brings together Weights & Biases, OpenPipe post-training expertise, and the marimo notebook project.

Related capabilities include ARIA for analyzing experiments, Agent Lens for tracing production agents, and isolated Sandboxes. The source also says capacity can be operated through CoreWeave's Kubernetes service, SUNK, Mission Control, Sandboxes, and Inference products. These details suggest a stack that spans development, testing, and deployment.

Reported performance notes

The source includes two additional performance claims. It reports more than three times faster sandbox startup in tests using Vera CPUs. It also reports a 1.7 times performance gain on passing Terminal-Bench tasks.

These figures should be read in context. They come from the announcement and describe specific tests. The source does not present them as universal results across all workloads.

Why this matters for agentic AI

The main theme is continuity. The announcement connects training, evaluation, and production in one infrastructure story. That matters for teams building agents, because the same platform now appears to support multiple stages of the workflow.

The source also emphasizes observability and isolation. ARIA, Agent Lens, and Sandboxes point to a setup where experiments and production agents can be inspected and separated. That can help teams manage iteration without losing track of what is running.

Morocco relevance

The source reports no Moroccan deployment or access. For readers, the global lesson is simple: when evaluating agentic AI infrastructure, look for tools that connect training, testing, tracing, and production in one flow.

Bottom line

This announcement is about closing the loop on agentic AI infrastructure. CoreWeave is adding hardware, CPU support, and software tools around the same workflow. NVIDIA's report presents it as a step from training toward production use, with Cognition as the first named customer on the new hardware.