Платформа Mirantis k0rdent AI получила сертификацию NVIDIA за виртуализацию GPU с производительностью на уровне физических серверов

Mirantis' k0rdent AI platform has been certified through NVIDIA's Certified Hypervisors programme, covering GPU virtualisation for enterprise AI factories and cloud service providers. The validation followed benchmarking tests designed to measure how closely virtualised GPU performance matched bare-metal deployments. In those tests, k0rdent AI delivered performance within 5% of comparable bare-metal systems across validation scenarios on NVIDIA GB200 NVL72 and NVIDIA HGX-based tracks. GPU virtualisation allows organisations to divide computing resources more precisely between workloads instead of assigning an entire GPU server to one user or project. The approach is becoming more relevant as companies and cloud operators look to share expensive AI infrastructure across multiple teams while keeping performance losses low. Mirantis is among the first independent software vendors to join the NVIDIA-Certified Hypervisors programme. The certification builds on the company's earlier involvement in NVIDIA's AI Cloud Ready validation work, which focused on AI infrastructure deployment. For companies building AI systems, the certification is intended to reduce the engineering work needed to design, test and run production GPU virtualisation. It also gives customers an external benchmark for software intended to support shared AI environments. Shared infrastructure The shift toward multi-tenant AI infrastructure has increased demand for software that can isolate workloads securely while using hardware efficiently. In practice, that means allowing different users or customers to run jobs on the same underlying systems without giving them full access to an entire server. In Mirantis' approach, virtual machines serve as the security boundary, with networking, GPUs and storage isolated for each customer or workload. This operating model mirrors what many large AI cloud environments require. The platform also manages both virtual machines and Kubernetes workloads through a single control plane. That allows organisations to run AI models, databases, traditional business applications and supporting services on the same management layer rather than maintaining separate stacks for virtualisation and container orchestration. A key part of the platform is its topology-aware scheduler, designed to understand how GPUs, CPUs, network interfaces and storage are connected inside a server and across a rack. Mirantis says this allows workloads to be placed with awareness of NVIDIA NVLink fabrics, NUMA domains and NVIDIA InfiniBand or RoCE networking, reducing bottlenecks that could affect AI training and inference. Market context The announcement reflects a wider effort across the AI infrastructure market to increase utilisation of GPU systems, which are costly and often in short supply. Cloud providers and enterprise operators are under pressure to allocate these resources more efficiently while preserving the speed needed for large-scale model training and inference. Near bare-metal performance has become a key metric in that effort because many AI workloads are sensitive to latency and bandwidth constraints. Any virtualisation layer that introduces too much overhead can undermine the economics of shared infrastructure. Mirantis says its software was built for AI factories and cloud providers that need flexible GPU allocation, workload isolation and operational efficiency across several teams or tenants. It argues those requirements are becoming more common as AI infrastructure shifts from dedicated systems to pooled environments. Kevin Kamel, Vice President of Product Management at Mirantis, said the certification showed the software could support that transition. "As AI infrastructure evolves into shared, multi-tenant AI factories, organizations need virtualization that delivers flexibility without sacrificing performance," said Kevin Kamel, Vice President of Product Management at Mirantis. "This certification demonstrates that k0rdent AI enables enterprises and cloud providers to securely virtualize NVIDIA accelerated infrastructure while maintaining the performance required for AI training and inference workloads." Product availability Mirantis said k0rdent AI Virtual Machine as a Service is available as a technical preview. It sits alongside other parts of the company's infrastructure software, including Bare Metal as a Service, Kubernetes orchestration and AI infrastructure management tools. The company sells Kubernetes-based infrastructure software for enterprise customers and has positioned k0rdent AI as part of a broader push into the market for shared GPU clouds and enterprise AI deployments. Its customer list includes Adobe, Ericsson, Inmarsat, MetLife, PayPal and Societe Generale. The latest certification gives Mirantis another credential as competition intensifies among software providers building management layers for GPU-heavy infrastructure, with performance, isolation and ease of operation emerging as the main points of comparison.