Cognition is first customer on CoreWeave's NVIDIA Vera Rubin NVL72

Limited availability and first production customer

CoreWeave says hundreds of NVIDIA Rubin GPUs are deployed across multiple regions, and that it has started onboarding customer workloads in production. Cognition is the first customer to bring production agentic AI workloads to the platform.

Access is tied to CoreWeave's wider stack. According to the company, customers can use Vera Rubin NVL72 capacity across CoreWeave Kubernetes Service (CKS), CoreWeave Inference, CoreWeave SUNK, CoreWeave AI Object Storage and CoreWeave Mission Control. CoreWeave frames the offer around compute-heavy training, high-throughput agentic inference with low latency, and managing end-to-end AI pipelines.

The company also says Cognition's engineers ran production workloads within days of rack handover without needing custom bring-up work of their own. CoreWeave presents this as evidence of production readiness.

How Cognition uses the platform

Devin is described as an autonomous AI software engineer that works alongside development teams on complex, end-to-end assignments. Its work can span entire code repositories and involve long chains of reasoning, code generation, debugging, execution, evaluation and iteration.

Cognition runs both model training and production inference on the same bare-metal CoreWeave Cloud platform. CoreWeave lists four areas:

• Production inference serving: real-time agentic queries served globally through CoreWeave Inference.
• Inference optimization: work with CoreWeave performance engineers on KV cache management, serving parameters and runtime configurations.
• Experiment tracking: Weights & Biases Models used to monitor and evaluate experiments across training runs.
• Training and reinforcement learning: large-scale pre-training, fine-tuning, RL post-training and test-time scaling on Devin models.

CoreWeave adds that Cognition has scaled from initial bridge capacity to thousands of GPUs for training and dedicated inference on its cloud.

Benchmark results and hardware specifications

Cognition's engineering team ran independent benchmarks on CoreWeave, comparing Vera Rubin NVL72 against previous-generation GB200 NVL72 baselines, using SWE-2, which the source describes as the company's latest generation of autonomous AI software engineer.

• Inference: a 4.8x increase in total token throughput. CoreWeave states that Vera Rubin NVL72 delivers up to 4.8x the total token throughput per GPU of GB200 NVL72 at matched interactivity.
• Reinforcement learning: a 3.8x boost in output token throughput, which CoreWeave says is per GPU at matched interactivity. Cognition generates large volumes of trial-and-error reasoning trajectories during post-training, so faster output speeds up iteration on new model weights.

Silas Alberti, SVP Research and Founding Team at Cognition, said that for an agentic workload where every step waits on the last one, the gain "compounds into real work Devin gets done."

On hardware, CoreWeave describes Vera Rubin NVL72 as a unified, liquid-cooled platform engineered as a single logical GPU. It integrates 72 Rubin GPUs connected with NVIDIA NVLink 6, plus 36 Vera CPUs. The source cites 1,400 TB/s of HBM4 memory bandwidth and 216 TB/s of NVLink bandwidth, and says the platform can handle trillion-parameter models while scaling rollouts at a 45x lower token cost versus Blackwell.

What it means for customers

For teams building agents, the main claim is that faster processing at each step of a reason, retrieve, execute and retry loop compounds across the whole workflow, so agents finish more work in less time and at lower cost. The figures come from Cognition's own SWE-2 workloads and from CoreWeave, so results for other workloads may differ.

Availability is limited and aimed at select customers. CoreWeave invites interested teams to request a briefing on capacity planning, onboarding timeline and workload fit, or to request an ARENA Pass to run their own workload on Vera Rubin NVL72.