Run
Maximize GPU use, streamline AI workflows, enhance efficiency.. [Contact for Pricing]
Last verified:
What is Run?
NVIDIA Run:ai is an enterprise platform for AI workloads and GPU orchestration that accelerates AI operations with dynamic orchestration across the AI life cycle. It maximizes GPU efficiency, scales workloads, and integrates seamlessly into hybrid AI infrastructure with zero manual effort. The platform centralizes and automates AI workload execution across distributed environments, transforming fragmented infrastructure into a scalable AI factory.
Key features include AI-native workload orchestration, dynamic GPU allocation with fractional GPU sharing, policy-driven governance for centralized resource management, and open architecture with API-first design. Run:ai delivers 10x GPU availability, 20x more workloads running, and 5x GPU utilization with zero manual intervention. It supports public clouds, private clouds, hybrid environments, and on-premises data centers.
The platform is designed for enterprises running AI and machine learning operations, including data scientists, ML engineers, and IT teams who need to scale AI training and inference workloads. It is particularly valuable for organizations dealing with GPU contention, idle capacity, and the operational complexity of managing isolated environments for multiple teams. NVIDIA Run:ai is now included in NVIDIA AI Enterprise and NVIDIA Mission Control.
Open-source solutions from Run:ai include the KAI Scheduler for Kubernetes-based AI workload scheduling, Grove for topology-optimized serving, and Model Streamer to cut model loading times from minutes to seconds. The platform enables fractional inference, mitigates model cold start through GPU memory swapping, and maximizes token throughput by running diverse workloads concurrently.
Run pricing
Pricing model: Freemium
NVIDIA Run:ai is available only via private offer. There is no public pricing information available. Customers must contact their NVIDIA sales representative for pricing details. The platform is included in NVIDIA AI Enterprise subscription, which provides reliable, secure, and scalable production AI operations. NVIDIA AI Enterprise includes a five-year subscription for H200 NVL, H100 NVL, and H100 PCIe GPUs, and a three-year subscription for A800 40GB Active GPU. Free trial and free version options are listed but specific details are not publicly disclosed. Contact NVIDIA or visit the NVIDIA Partner Network Locator to find certified partners for purchasing.
Run pros
- 10x increase in GPU availability
- 20x more workloads running simultaneously
- 5x improvement in GPU utilization
- Zero manual intervention required
- Dynamic GPU allocation with fractional GPU sharing down to 0.125 GPU
- Policy-driven governance for fair resource access across teams
- Supports hybrid and multi-cloud environments seamlessly
- Open architecture with API-first design for easy integration
- Compatible with all major AI frameworks and ML tools
- Reduced operational costs through eliminated GPU waste
- Faster time to production for AI applications
- Centralized orchestration with end-to-end visibility
- GPU memory swap enables larger models on fewer GPUs
- KAI Scheduler provides simple YAML-based Kubernetes management
- Model Streamer cuts model loading times from minutes to seconds
- Includes priority preemption patterns tailored for AI workflows
- MIG support and automation for fine-grained GPU partitioning
Run cons
- No public pricing - enterprise private offer only
- Requires contact with NVIDIA sales representative
- Default multi-tenant model has weak tenant isolation
- Shared control plane creates no clear boundary between tenants
- Tenants cannot self-manage Run:ai projects in default model
- CRD restrictions make custom resource management difficult
- Complexity increases with multi-cluster separation approaches
- GPU fragmentation occurs when using cluster separation for isolation
- Steep learning curve for Kubernetes and AI infrastructure
- Primarily designed for enterprise-scale deployments
Frequently asked questions about Run
What is NVIDIA Run:ai?
NVIDIA Run:ai is an enterprise platform for AI workloads and GPU orchestration that accelerates AI operations with dynamic orchestration across the AI life cycle. It maximizes GPU efficiency, scales workloads, and integrates seamlessly into hybrid AI infrastructure with zero manual effort. The platform provides AI-native workload orchestration, dynamic GPU allocation, policy-driven governance, and open architecture with API-first design.
How does Run:ai improve GPU utilization?
Run:ai achieves 5x GPU utilization through dynamic GPU allocation with fractional GPU sharing down to 0.125 GPU, GPU time-slicing that proportionally shares compute, fair-share queueing with dynamic allocation, and intelligent queueing that prevents jobs from failing. It eliminates waste by dynamically pooling GPU resources across hybrid environments and aligning compute capacity with business priorities.
What deployment options does Run:ai support?
Run:ai supports flexible deployment across public clouds, private clouds, hybrid environments, and on-premises data centers. Documentation is available for SaaS (fully managed cloud-hosted platform), self-hosted (on-prem and private cloud deployments), and multi-tenant (on-prem/private cloud with centralized control plane for multiple isolated organizations) deployment models.
What is fractional GPU allocation in Run:ai?
Fractional GPU allocation allows multiple workloads to share a single GPU by dividing it into slices as small as 0.125 GPU. This enables notebooks and inference workloads to run on a portion of a GPU rather than requiring a whole GPU. The approach supports consistent throughput scaling, improved utilization with minimal idle capacity, and stable latency under mixed workloads and high concurrency.
Is Run:ai included in NVIDIA AI Enterprise?
Yes, NVIDIA Run:ai is now included in NVIDIA AI Enterprise, which accelerates and simplifies development and deployment of production AI applications. NVIDIA AI Enterprise reduces time to market and lowers infrastructure costs while ensuring reliable, secure, and scalable operations.
What open-source solutions come from Run:ai?
Run:ai offers three open-source solutions: KAI Scheduler for fair and efficient AI workload scheduling on Kubernetes using YAML files, Grove for topology-optimized serving on Kubernetes that bridges AI inference frameworks and scheduling, and Model Streamer (Python SDK with C++ backend) to cut model loading times from minutes to seconds by concurrently reading tensors and transferring directly to GPU memory.
How does Run:ai handle multi-tenancy?
Run:ai offers three approaches to multi-tenancy: default model with tenants in namespaces within a shared Kubernetes cluster (maximizes GPU utilization >90% but has weak tenant isolation), complete cluster separation with hosted control planes (strong isolation but fragmented node pools and underutilized hardware), or solutions combining Run:ai with vCluster for both optimal GPU utilization and strong tenant autonomy.
What performance benefits does Run:ai deliver?
Run:ai delivers 10x GPU availability, 20x more workloads running, 5x GPU utilization, and zero manual intervention. It accelerates AI throughput through dynamic scheduling and orchestration, delivers seamless scaling, and maximizes GPU utilization. The platform reduces bottlenecks, shortens development cycles, and scales AI solutions to production faster.
How does Run:ai mitigate model cold start?
Run:ai mitigates model cold start through GPU memory swap that dynamically swaps model memory between GPU and host. This approach keeps active parts of the model resident on GPU while transparently paging inactive portions, enabling larger models to run on fewer GPUs. This reduces infrastructure spend, lowers idle capacity, and supports cost-efficient inference for production deployments, especially for memory-intensive large language model workloads.
Who should use NVIDIA Run:ai?
Run:ai is designed for enterprises scaling AI workloads efficiently, including data scientists, ML engineers, and IT teams. It is ideal for organizations dealing with GPU contention, idle capacity, and operational complexity of managing isolated environments. It serves enterprises building production-scale agentic AI systems, telecom companies investing in AI, and any organization seeking to maximize compute efficiency while dynamically scaling AI training and inference.