NVIDIA Vera Rubin NVL72 and Vera CPU Launch on CoreWeave

CoreWeave announced the availability of NVIDIA Vera Rubin NVL72 systems and the NVIDIA Vera CPU at its Fully Connected event in San Francisco on September 30, 2026. The platform delivers up to a 4.8x increase in total token throughput for SWE-2 inference workloads compared to GB200 NVL72, and it matters to infrastructure architects seeking to scale agentic AI. (token throughput: the rate at which an artificial intelligence model processes and generates units of text.) (agentic AI: artificial intelligence systems designed to perform autonomous actions and complex reasoning tasks independently.)

NVIDIA Vera Rubin NVL72 provides 4.8x throughput for software engineering

CoreWeave announced that the NVIDIA Vera Rubin NVL72 is now available on CoreWeave Cloud, featuring Spectrum-X 102.4T Ethernet networking. Cognition, the lab behind Devin, is the first customer running production workloads on this hardware. To verify performance, Cognition benchmarked the Vera Rubin NVL72 against a GB200 NVL72 baseline using a real-world software engineering workload sampled from FrontierCode.

Cognition reported that Vera Rubin NVL72 delivered up to a 4.8x increase in total token throughput for SWE-2 inference workloads over the GB200 NVL72. For the Devin AI software engineer, these gains result in faster real-time code generation and more responsive multistep reasoning. Silas Alberti, a founding team member at Cognition, stated that cost per token decides what can be shipped for complex agentic coding workloads involving long contexts and high concurrency.

Vera Rubin vs. GB200 Performance

NVIDIA Vera CPU offers purpose-built agentic concurrency

NVIDIA Vera CPU is the first processor specifically built for AI agents, which require both low-latency compute at scale and thousands of isolated environments for post-training. CoreWeave’s deployment of the Vera CPU puts 128 CPUs and 11,264 cores in a single rack. This density allows for more than 11,000 concurrent environments at one core each.

The architecture relies on hardware-isolated sandboxes and Spectrum-X Ethernet switches

The architecture relies on hardware-isolated sandboxes and Spectrum-X Ethernet switches with BlueField-4 DPUs to ensure secure, low-latency agent communication. (BlueField-4 DPUs: specialized hardware processors designed to offload networking and storage tasks from the main processor.) In testing, CoreWeave achieved more than 3x faster agent sandbox startup times on NVIDIA Vera CPUs compared to previous methods. On Terminal-Bench, the company saw a 1.7x performance gain on Vera CPU across all passing tasks.

CoreWeave Forge unifies the AI training and production loop

CoreWeave Forge is a connected environment that unifies Weights & Biases, post-training expertise from OpenPipe, and the marimo notebook project. The platform closes the loop by helping ensure production behavior informs the next training run without signal loss at handoffs between different vendors. The tool stays open across different models, frameworks, and clouds.

The Forge environment includes several new capabilities: CoreWeave ARIA for analyzing runs and proposing experiments; CoreWeave Agent Lens for production agent observability; and CoreWeave Sandboxes for serverless execution of RL and tool calls. The system also includes RL Rollouts, which allows for loading new checkpoints into a live deployment while it’s running so that reinforcement learning continues without redeployments.

Economic advantages of serverless RL and inference

Serverless RL on CoreWeave provides a significant path for developers to improve model quality and cut costs without requiring a dedicated training cluster. CoreWeave stated that serverless RL trains 1.4x faster and at a 40% lower cost than a self-managed setup. This allows for rapid experimentation with training recipes using production signals.

The infrastructure is designed to scale efficiently as agentic AI moves from simple chat to complex reasoning. By utilizing the high density of the Vera CPU, organizations can avoid the costs of managing large, underutilized clusters. The cost of being wrong on this infrastructure is high; choosing a standard cloud instance without Vera CPU capabilities might lead to significant queuing of agent runs, slowing iteration speed.

Who this is for: AI infrastructure architects

AI infrastructure architects should use the Vera CPU and Vera Rubin NVL72 to achieve the density required for scaling to thousands of concurrent agent environments. MLOps engineers can use CoreWeave Forge to reduce the complexity of reinforcement learning and cut costs by 40% compared to self-managed setups. Software engineering AI developers, such as those building tools like Devin, can leverage the 4.8x throughput increase for SWE-2 workloads. Cloud specifiers can prioritize CoreWeave to access the latest NVIDIA Vera Rubin hardware earlier than many competitors.

Who this is not for: Home PC gamers

Home PC gamers can ignore this as it is enterprise-grade AI infrastructure rather than consumer hardware; they should instead look at NVIDIA GeForce GPU reviews. Mobile app developers should skip this as the focus is on backend AI factory infrastructure; they should instead look at mobile NPU or edge AI deployment guides.

Tariq’s scenario for scaling coding agents

Tariq runs a startup building a fleet of 5,000 autonomous coding agents and requires high concurrency and low-cost training. Because the Vera CPU supports 11,000 concurrent environments per rack, Tariq can run all 5,000 agents simultaneously in hardware-isolated sandboxes on CoreWeave. He uses CoreWeave Forge to feed production errors from these agents directly back into his training loop.

If Tariq used a standard cloud instance without the Vera CPU, he would likely have to queue his agent runs, significantly slowing his iteration speed. He would also face higher costs for RL training if he chose to manage his own cluster rather than using CoreWeave’s serverless RL. This infrastructure helps ensure his startup can scale without linear cost growth.

Applications for agentic AI infrastructure

• Deploy 11,000+ concurrent agent sandboxes for large-scale AI labs.
• Reduce reinforcement learning costs by 40% for budget-conscious AI startups.
• Close the feedback loop from production inference to training for model improvement teams.
• Perform high-throughput inference for software engineering tools using the 4.8x throughput jump.
• Run live RL without redeployments using RL Rollouts for online learning researchers.
• Scale clinical documentation for healthcare providers using reserved RTX PRO 6000 GPUs.

What this replaces: CoreWeave's serverless RL reduces

CoreWeave’s serverless RL reduces the need for self-managed RL setups, which are 1.4x slower and 40% more expensive. By moving to serverless infrastructure, teams stop needing to manage the underlying complexity of RL clusters while maintaining the ability to run code in isolated environments.

What is not yet known: CoreWeave did not specify the geographical regions where the Vera Rubin

CoreWeave did not specify the geographical regions where the Vera Rubin NVL72 is currently available. The company also did not provide specific power requirements, socket types, or required operating systems for the Vera CPU. Furthermore, while Cognition provided a 4.8x throughput figure, it remains unconfirmed whether these results hold across all software engineering use cases beyond the SWE-2 benchmark.

Forward-looking judgement

The shift toward the Vera CPU signals a move from general-purpose AI accelerators to silicon purpose-built for agentic concurrency. As AI agents begin to perform more autonomous actions, the ability to host thousands of isolated environments on a single rack will become the primary metric for infrastructure providers.

Reported by NVIDIA on 30 September 2026: From Training to Production, NVIDIA and CoreWeave Close the Loop on Agentic AI.