Zesty introduced a Kubernetes agent for NVIDIA GPUs on Oct. 15, saying it can rebalance resources as artificial intelligence workloads shift and help operators make more use of hardware they already own.
The feature, named Automated Fractional GPUs, is offered through Zesty’s platform. It is built to work alongside NVIDIA’s GPU Fractioning Operator and the Kubernetes scheduler. Zesty says its agent decides which workloads share a GPU and how to distribute memory among them.
NVIDIA made Run:ai open-source on Oct. 13. The release includes the GPU Fractioning Operator, which brings CUDA-native GPU fractions into Kubernetes. NVIDIA says the CUDA driver sets a firm memory ceiling for each workload. The operator is scheduler-agnostic, NVIDIA said.
Zesty says its agent weighs memory consumption and GPU utilization, while factoring in workload behavior over time. The company says it assigns sharing modes and allocations under customer policies for latency, resource use and availability. Supported modes include time-slicing and NVIDIA Multi-Process Service (MPS).
The company says teams can install the tool without changing application or model code, and installation does not restart running workloads. Zesty also says the agent can adjust allocations and placement as demand changes.
Zesty’s business case is that changing demand can leave GPU capacity idle under static allocations, while GPUs are expensive and difficult to obtain. The company says improved sharing could let teams run more workloads on existing hardware and avoid buying extra capacity. Its launch announcement did not include utilization measurements, savings estimates, latency results or customer deployments.
The Cloud Native Computing Foundation’s 2025 survey found that 66% of organizations that host generative AI models use Kubernetes for at least some inference workloads.
CUDA-native fractioning is part of CUDA 13.4. NVIDIA’s release notes say new features in that version require driver branch R615 or later. Zesty’s launch announcement did not list supported GPU models or a full driver compatibility range.
