훈련에서 프로덕션까지, NVIDIA와 CoreWeave가 에이전틱 AI의 루프를 완성하다

Building on nearly a decade of co-engineering, CoreWeave has built NVIDIA compute, networking and software into a cloud purpose-built for AI that’s still returning on investment across multiple generations of deployment. Now, CoreWeave is bringing the next generation of NVIDIA infrastructure to production.

At CoreWeave Fully Connected, running this week in San Francisco, CoreWeave announced availability of [NVIDIA Vera Rubin NVL72](https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/) systems with Spectrum-X 102.4T Ethernet networking. Cognition, the applied AI lab behind the Devin AI software engineer, is the first customer running production workloads on Vera Rubin.

CoreWeave will also offer [NVIDIA Vera](https://www.nvidia.com/en-us/data-center/vera-cpu/), the first CPU built for AI agents. In addition, CoreWeave launched CoreWeave Forge, a connected environment for training, evaluating and improving models and agents on NVIDIA accelerated computing.

“NVIDIA accelerated computing delivers value across generations,” said Ian Buck, vice president of hyperscale and high-performance computing at NVIDIA. “CoreWeave’s NVIDIA V100 GPUs are still running customer workloads nearly a decade after Volta launched, even as CoreWeave brings Vera Rubin NVL72 into production. That’s the strength of the NVIDIA platform: infrastructure that keeps earning for years, and the flexibility to put the right GPU on the right workload.”

## **Cognition Runs on Vera Rubin NVL72 With 4.8x Higher Token Throughput**

Cognition runs training, reinforcement learning and production inference for Devin on CoreWeave. The company scaled to thousands of GPUs on CoreWeave in nine months, powering Cognition inference workloads.

Earlier this month, CoreWeave received its first Vera Rubin NVL72 production racks. Shortly after, Cognition benchmarked Vera Rubin’s inference performance against a GB200 NVL72 baseline using a real-world software engineering workload. To generate this workload, it sampled a subset of tasks from FrontierCode and deployed AI agents to solve them.

In its early tests, Cognition saw Vera Rubin NVL72 deliver up to a 4.8x increase in total token throughput for SWE-2 inference workloads over GB200 NVL72. For Devin, those gains mean faster real-time code generation and more responsive multistep reasoning.

“Agentic coding is a complex workload: long contexts, high concurrency and token volumes where cost per token decides what we can ship,” said Silas Alberti, founding team at Cognition. “Having all of it on one platform, with NVIDIA and CoreWeave engineers who work the hard problems alongside ours, matters more to us than any single spec.”

## **CoreWeave Announces NVIDIA Vera Rubin Availability on CoreWeave Cloud**

CoreWeave announced availability of NVIDIA Vera Rubin NVL72 on CoreWeave Cloud, making it one of the first cloud providers to deliver the platform in customers’ hands.

Early-access customers can put the performance of NVIDIA’s full-stack AI factory platform to work quickly on CoreWeave Cloud. In days, CoreWeave stood up a production Vera Rubin cluster for Cognition, achieved as a result of the codesign and collaboration between NVIDIA and CoreWeave up and down the stack, from infrastructure to tokens served.

Capacity can be operated through CoreWeave Kubernetes Service, SUNK, CoreWeave Mission Control, CoreWeave Sandboxes and CoreWeave Inference.

## **NVIDIA Vera CPU to Come to CoreWeave Cloud, Tests Show More Than 3x Faster Agentic Sandbox Startups**

Agentic AI puts pressure on infrastructure from two directions: serving agents demands low-latency compute at scale, while improving them through post-training requires thousands of isolated environments running at once.

NVIDIA Vera CPU is purpose-built for agentic workloads. For agentic AI, a key performance measure is how many isolated agent environments can run at once and how consistent and performant each one stays as that number grows.

CoreWeave’s deployment of Vera puts 128 CPUs and 11,264 cores in a single rack, enough for more than 11,000 concurrent environments at one core each. With CoreWeave Sandboxes, these environments are hardware-isolated and run alongside the training jobs they support, with Spectrum-X Ethernet switches and BlueField-4 DPUs ensuring secure, high-performance, secure agent communication at low latency.

In testing, CoreWeave achieved more than 3x faster agent sandbox startup times on NVIDIA Vera CPUs, accelerating and scaling its sandboxes, an execution layer for reinforcement learning (RL), agent tool use and model evaluation that let AI teams run code in isolated environments on CoreWeave. On Terminal-Bench, CoreWeave saw a 1.7x performance gain on Vera CPU across all passing tasks.

## **CoreWeave Forge: Closing the AI Loop From Production Back to Training**

Models and agents improve by running a loop: production behavior informs the next training run, and each evaluation sharpens the next version. This loop has historically been split across tools from different vendors, with signal lost at every handoff.

CoreWeave Forge unifies Weights & Biases, post-training expertise from OpenPipe and the open source marimo notebook project in one connected environment built for continuous model and agent improvement. It stays open across models, frameworks and clouds.

New and expanded capabilities available include:

* **CoreWeave ARIA**— now generally available — helps users learn, research, code and iterate across the AI loop, analyzing runs, proposing experiments, recommending code changes and storing them in GitHub, and bringing back actionable evidence that analyzes experiment data, surfaces what drove a change and proposes the next experiments to run.
* **CoreWeave Agent Lens**— a new service — turns production agent observability into continuous improvement and understandable insights. It improves failure detection by 20% and fixes issues at half of the cost, which turns tens of millions of production agent traces into insights that drive fixes.
* **CoreWeave Sandboxes** — now generally available — let users run agents, tool calls, RL and evaluations in isolated CPU or GPU execution environments, on serverless infrastructure or on the infrastructure they already train on, providing a fresh, isolated environment for every agent tool call, RL run or evaluation.
* **Post-training improves model quality and cuts latency and costs** harnessing users’ own production signals, with no training cluster required.[Serverless supervised fine-tuning](https://www.coreweave.com/products/coreweave-forge/serverless-sft) and[serverless RL](https://www.coreweave.com/products/coreweave-forge/serverless-rl) let users experiment with their own training recipes. Serverless RL trains 1.4x faster at 40% lower cost than a self-managed setup.

[NVIDIA Dynamo](https://www.nvidia.com/en-us/ai/dynamo/), an open source inference framework for AI factories, powers CoreWeave’s managed inference service as well as RL Rollouts, now in private preview. RL Rollouts loads new checkpoints into a live deployment while it’s running, so reinforcement learning continues without redeploys, and post-training gets the same inference efficiency as production.

Canva, Capital One and MasterClass are among the first companies building on Forge.

[NVIDIA Nemotron](https://www.nvidia.com/en-us/ai-data-science/foundation-models/nemotron/) open models give teams on Forge a direct path to customizing and deploying reasoning and multimodal models for agentic workflows.

## **Proven Impact From Startups to Global Enterprises**

AI labs, AI-natives and global enterprises are using the co-engineered NVIDIA and CoreWeave platform to move from prototype to production faster.

In healthcare, Ennoble Care, a home-based care provider serving about 50,000 high-need Medicare patients across 15 states, selected CoreWeave to run clinical AI inference. It will use reserved NVIDIA RTX PRO 6000 GPU capacity on CoreWeave Kubernetes Service to scale AI agents for clinical documentation, decision support and back-office automation.

CoreWeave has delivered record [MLPerf results](https://coreweave.com/blog/coreweave-leads-cloud-providers-in-mlperf-r-inference-v6-1-performance-with-nvidia-blackwell-ultra) in every round of training and inference. It’s the only cloud provider who holds the Platinum ranking in SemiAnalysis ClusterMAX 1.0, 2.0 and 3.0, and serves nine of the 10 leading AI labs.

Together, NVIDIA and CoreWeave are giving customers a platform to turn experimental agents into production systems that write software, support clinicians and do useful work in the real world.

_Learn more by attending_[_NVIDIA sessions, demos and workshops at CoreWeave Fully Connected_](https://www.nvidia.com/en-us/events/coreweave-fully-connected/)_._