NVIDIA Vera Rubin Goes Live on CoreWeave as Devin Sees a Major AI Throughput Jump
NVIDIA’s next-generation Vera Rubin AI infrastructure has moved from testing into production on CoreWeave, with Cognition becoming the first customer to run production workloads on the new platform.
The development was announced September 30 at CoreWeave Fully Connected in San Francisco, where CoreWeave said NVIDIA Vera Rubin NVL72 is now in limited availability on CoreWeave Cloud. Cognition, the company behind the Devin autonomous AI software engineer, is already using the system for production agentic AI workloads.
The headline performance result is significant: Cognition reported up to a 4.8x increase in total token throughput for SWE-2 inference workloads compared with an NVIDIA GB200 NVL72 baseline. CoreWeave says the testing used a real-world software engineering workload designed around the demands of autonomous coding agents.
NVIDIA Vera Rubin moves into production
Vera Rubin NVL72 is designed as a rack-scale AI system rather than simply a standalone GPU upgrade. A single system combines 72 NVIDIA Rubin GPUs with 36 Vera CPUs and NVIDIA’s NVLink 6 interconnect, alongside networking and other infrastructure intended for large-scale AI workloads.
CoreWeave said it received its first Vera Rubin NVL72 production racks earlier in September and was able to bring Cognition’s production workloads online within days of the rack handover.
That is an important shift for the platform. CoreWeave had previously completed the first industry bring-up and validation of Vera Rubin NVL72 in June, followed by measured performance testing in July.
The company has now moved beyond validating the infrastructure and is making the system available to selected customers through its cloud platform.
Cognition puts Vera Rubin to work on Devin
Cognition is using CoreWeave infrastructure for training, reinforcement learning and inference associated with Devin.
According to CoreWeave, Cognition scaled to thousands of GPUs on the platform within nine months. Its engineers then used Vera Rubin NVL72 to test production-style agentic software engineering workloads against a GB200 NVL72 baseline.
The result was up to 4.8x higher total token throughput for SWE-2 inference at matched interactivity, according to the company’s benchmark.
CoreWeave also reported a 3.8x increase in output-token throughput for reinforcement learning in the same comparison.
For an AI coding system such as Devin, throughput can affect how quickly the system processes context, generates code, executes tasks and moves through repeated reasoning and tool-use steps.
Cognition’s Silas Alberti said agentic coding requires long contexts, high concurrency and rapid reasoning, making performance at the infrastructure level particularly important to the company’s workloads.
CoreWeave is adding NVIDIA Vera CPU
The rollout is not limited to Rubin GPUs.
CoreWeave also announced plans to offer NVIDIA Vera, a CPU designed specifically for AI-agent workloads, on CoreWeave Cloud.
CoreWeave says its Vera deployment places 128 CPUs and 11,264 cores in a single rack. In its testing, the company reported more than 3x faster agent sandbox startup times compared with its previous setup, while Terminal-Bench testing showed a 1.7x performance gain across passing tasks.
These capabilities target a different bottleneck in agentic AI: running large numbers of isolated environments for reinforcement learning, evaluation and tool execution.
CoreWeave launches Forge for the AI development loop
Alongside the hardware rollout, CoreWeave introduced CoreWeave Forge, a connected environment intended to bring model training, evaluation, post-training and production feedback into a more continuous workflow.
The platform brings together technologies including Weights & Biases, OpenPipe’s post-training capabilities and the open-source marimo notebook project. CoreWeave says Forge is designed to remain open across models, frameworks and clouds.
New capabilities include CoreWeave ARIA, Agent Lens and generally available Sandboxes. CoreWeave also said its serverless reinforcement-learning system can train 1.4x faster at 40% lower cost than a self-managed setup, based on its testing.
Companies including Canva, Capital One and MasterClass are among the early organizations building on Forge, according to CoreWeave.
NVIDIA’s open Nemotron models are also being made available within the environment for teams working on reasoning and multimodal agentic applications.
The bigger AI infrastructure shift
The Vera Rubin rollout comes as AI companies increasingly focus on systems that do more than generate individual responses.
Agentic AI workloads repeatedly call models, retrieve information, use tools, execute code and evaluate results. That makes latency, throughput, networking and the ability to run large numbers of concurrent environments increasingly important.
CoreWeave and NVIDIA have been working together on this infrastructure for years. CoreWeave previously brought up and validated Vera Rubin NVL72 in June and reported a separate benchmark in July showing up to 10x more tokens per megawatt than GB200 NVL72 on a DeepSeek R1 workload at matched interactivity.
The latest announcement takes that progression another step: Vera Rubin is no longer only being benchmarked. It is now running production workloads for customers on CoreWeave Cloud.
CoreWeave said Vera Rubin NVL72 is initially available on a limited basis, with customers able to access the infrastructure through services including CoreWeave Kubernetes Service, CoreWeave Inference, SUNK, Mission Control and other parts of its cloud platform.
