Virtualized infrastructure is how raw GPU capacity becomes a business. It is what lets a GPU fleet serve many teams and customers at once, with the isolation, quotas, and as-a-service delivery that make compute consumable and monetizable. The open question has long been performance. A virtualization layer sits between the workload and the accelerator, and if it does not expose the hardware correctly, throughput and utilization quietly erode. That tradeoff kept many GPU cloud and enterprise teams on bare metal, giving up operational flexibility to protect performance.
NVIDIA built the NVIDIA-Certified Hypervisors program to settle that question. The program evaluates representative performance-critical behaviors across compute, memory, data-path efficiency, and large language model (LLM) inference, helping organizations identify virtualization platforms and infrastructure software designed to support demanding AI and accelerated computing workloads.
Rafay Systems has announced that its Virtual Machines-as-a-Service (VMaaS) offering, available to Rafay customers through the Rafay Platform, has been certified by NVIDIA for use with NVIDIA HGX systems, as well as NVL72 rack-scale systems.
The NVIDIA-Certified Hypervisors program targets NVIDIA accelerated computing platforms including NVIDIA HGX H200 and HGX H100 systems, as well as GB200 NVL72 and GB300 NVL72 rack-scale systems. HGX systems are the workhorse of most production AI deployments running today. The NVL72 rack-scale systems are the architecture defining the next generation of AI factories, so certification here signals that customers can plan those buildouts on virtualized infrastructure without treating performance as an unknown. Certification applies specifically to the product version and NVIDIA platform tested.
Together, these platforms cover what customers are deploying now and what they are designing for next.
NVIDIA-Certified Hypervisors are virtualization-based solutions that have proven to deliver near bare-metal performance for representative AI and accelerated computing workloads by accurately exposing hardware topology and implementing key performance optimizations. Rather than checking specifications on paper, the program validates representative performance-critical behaviors across compute, memory, data-path efficiency, and LLM inference. Passing it means the virtualized stack behaves close to bare metal on the specific NVIDIA platforms tested, so the operational benefits of virtualization no longer come at the cost of GPU performance.
For a control plane like Rafay, this matters because virtualization is foundational to how the platform delivers multi-tenant compute. The Rafay Platform lets many teams and tenants share the same GPU fleet safely, each in an isolated environment with its own quotas and policies, governed from a single control plane. Virtualization is also one of several consumption models the platform delivers from that same control plane, alongside bare metal, Kubernetes, SLURM, and AI services, so operators can offer virtual machines as part of one operating model rather than as a separate stack. Certified virtualization performance means operators no longer choose between secure multi-tenancy and full GPU throughput. They get both.
Neoclouds and GPU cloud providers. Virtualization is what turns a GPU fleet into sellable, metered products. Certified near bare-metal performance lets providers offer virtual machine and multi-tenant SKUs with confidence, raising utilization and margin without a performance penalty that customers would notice and price against.
Enterprises. Handing a dedicated GPU server to a single team is simple, but it is also wasteful, since most teams do not keep those GPUs busy and utilization stays low while the hardware stays expensive. Certified virtualized performance changes the math. Enterprises can pack multiple GPU VMs onto one bare metal server, serving several teams and users from the same hardware, driving utilization up while each VM still delivers near bare-metal performance.
Telecommunications providers. Telcos entering AI infrastructure are effectively becoming metered-services operators, serving many enterprise and internal tenants on shared capacity. Certified virtualization lets them carve one GPU fleet into isolated, governed environments and deliver them as a service, extending the operating model they already run at national scale rather than standing up a separate stack for each tenant.
Sovereign AI clouds. Sovereign operators serve many public and enterprise tenants on shared national infrastructure, where isolation and governance are non-negotiable. Validated virtualization performance lets them deliver that isolation at scale while keeping the utilization economics of a shared fleet intact.
Virtualization no longer has to be the layer where GPU performance goes to hide. With NVIDIA-Certified Hypervisors validation across NVIDIA HGX systems and NVL72 rack-scale systems, the Rafay Platform gives operators a path to run production AI on virtualized infrastructure and keep the performance that justified the GPU investment in the first place, from HGX systems today through GB200 NVL72 and GB300 NVL72 systems next.
Certified virtualization is one thread in a wider body of work between Rafay and NVIDIA to make accelerated infrastructure production-ready. Rafay, a member of NVIDIA Inception, has collaborated with NVIDIA spanning GPU Platform-as-a-Service reference architectures, NVIDIA AI Cloud Ready, NVIDIA Cloud Partners (NCPs), AI factories, NVIDIA DSX OS, NVIDIA Infra Controller (NICo), NVIDIA AI Enterprise software, and AI services delivered through Rafay Token Factory. The through-line across all of it is a single requirement the market keeps returning to: making accelerated infrastructure not just available, but securely consumable and operationally ready for production.
Rafay CEO Haseeb Budhani joins theCUBE to discuss sovereign AI, multi-tenant cloud platforms, GPU monetization, and the shift from infrastructure to AI services.
Read Now
Rafay MKS has achieved NVIDIA GPU Operator partner validation, providing platform teams with a standardized, governed approach to deploying GPU-accelerated Kubernetes. Learn how to move beyond manual, inconsistent configurations to repeatable, version-controlled AI infrastructure using Rafay Cluster Blueprints.
Read Now
See how Rafay transforms the NVIDIA Omniverse DSX Blueprint into a one-click, self-service digital twin offering with governed GPU access, multi-tenancy, and automated session management.
Read Now
