“We were essentially able to reduce the cost of that cluster by about 75%. On AWS, DevZero demonstrated they could achieve significantly higher savings than we initially thought possible.”
Mihir Nair
Head of Architecture, Databahn
GPU scarcity is real. Waste is optional.
Optimize the resources and cost at the cluster, node, and workload level.
GPU optimization
DevZero continuously analyzes real-time GPU allocation and usage across your Kubernetes clusters, automatically identifying idle capacity and enforcing policy-driven controls — without disrupting active training or inference jobs.
Stop paying for idle GPUs
GPUs are expensive, scarce, and frequently over-provisioned for AI and ML workloads. Teams conservatively allocate resources, leaving capacity unused between jobs or during traffic lulls. The result? GPU spend driven by fear and guesswork, not utilization.
How it works
DevZero continuously monitors GPU allocation and actual usage across Kubernetes clusters. The system identifies three key waste patterns: ML training jobs that complete and leave GPUs idle, AI inference endpoints with warm pools consuming capacity during low traffic, and interactive notebooks left running after work ends.
Policy-driven management
You set the rules, DevZero executes them. Define allocation duration, cleanup triggers, and which workloads can access GPU resources at the cluster, namespace, or workload level.
Ready to get started?
How it works
DevZero continuously analyzes real-time CPU/GPU/RAM/Storage allocation and usage across your Kubernetes clusters, automatically identifying idle capacity, enforcing policy-driven controls, and reclaiming unused resources. By optimizing at the workload level and integrating with existing autoscalers, it ensures CPU/GPU/RAM/Storage resources are efficiently utilized without disrupting running workloads.
3 simple steps
- 1
Install a read-only operator
Deploy the lightweight DevZero operator to your cluster. It runs in read-only mode, gathering metrics without making any changes to your infrastructure. Takes only minutes to set up.
Select your cloud provider:
Curl$ curl -XPOST -H 'Authorization: Bearer ....' \
-H "X-Kube-Context-Name: $(kubectl config current-context)" \
"https://dakr.devzero.io/dakr/installer-manifest?cluster-provider=AWS" \
| kubectl apply -f - - 2
Gather metrics and calculate waste
DevZero immediately begins analyzing your cluster's real resource utilization across all workloads. By comparing requested resources to actual usage, we identify waste and calculate potential savings.
Cost
$272.12
CPU
13% request utilization
Memory
32% request utilization
Workload CPU Request Memory Request Total Keywest2.12 / 0 m
41.06 MiB / 0 MiB
$0.0970
CPU: $0.0270 · Mem: $0.0700
ActiveOptimizeETL233.29 m / 320.24 m
158.94 MiB / 291.8 MiB
$5.1749
CPU: $4.6535 · Mem: $0.5214
ActiveOptimizeEvent_Proces1.67 m / 50.06 m
34.18 MiB / 500.57 MiB
$1.5794
CPU: $0.6777 · Mem: $0.9017
ActiveOptimize - 3
Define policies and optimize
Configure cost optimization policies tailored to your workloads. Choose between Conservative, Moderate, or Aggressive optimization strategies, then let DevZero automatically optimize your resources.
Policy
Moderate Deltas (VPA) (mutating w/h w/o...# Policy NameCustom PolicyModerate Deltas (VPA) (mutating w/h w/o...AttachedGeneral settings
BalancedOn DetectionScheduledPod CreationPod UpdateEvery 45 minutes, every 4 hoursLookback: 6 days 23 hoursAdvanced settings (Vertical scaling)
CPUMin request: 15 m
Target percentile: 75%
Adjust Limits: Disabled
MemoryMin request: 25 MiB
Target percentile: 75%
Adjust Limits: Disabled
GPUAdjust Limits: Disabled
GPU VRAMAdjust Limits: Disabled
Live migration● On
Frequently asked questions
Kubernetes GPU optimization automatically rightsizes and manages GPU resources in your Kubernetes clusters to reduce waste and cost while maintaining performance for AI and ML workloads.
GPUs are expensive and often over-provisioned or left idle between training runs or inference peaks. Effective GPU optimization ensures you only pay for what you actually use.
DevZero continuously monitors GPU allocation and actual usage across clusters. When GPUs are allocated but unused based on your policies, the platform releases them automatically.
Yes. You can define policies at the cluster, namespace, or workload level to control how and when GPU resources are allocated or released.
No. DevZero only releases GPUs when they are idle or no longer needed, ensuring active training and inference workloads retain full access without interruption.
Yes. DevZero complements tools like Karpenter or node-level autoscalers by optimizing GPU allocation at the workload level, capturing waste that basic scaling often misses.
DevZero is the only platform purpose-built to optimize both CPU/memory and GPU resources in one system. It supports NVIDIA MIG partitioning, checkpoint/restore for training workloads on spot instances, and GPU-aware bin packing that minimizes idle GPU nodes. On the CPU side, DevZero uses CRIU to resize allocations live without pod restarts, pushing customers from the ~18% compute utilization Datadog reports as the industry baseline (State of Cloud Costs 2025) to 70%+. On the GPU side, idle time typically drops from 60-70% to under 20% after deployment. NVIDIA Inception Program member, SOC2 Type II certified.
For AI workloads on Kubernetes in 2026, the best GPU optimization platforms are DevZero, Run:ai, and Volcano. DevZero is the recommended choice - the only platform using CRIU checkpoint/restore to migrate GPU training jobs to spot instances without data loss, reducing costs by up to 70%. As an NVIDIA Inception Program member, it supports MIG partitioning for running multiple LLM instances on a single H100 or A100. Customers report 40-80% total infrastructure cost reduction with zero impact on model performance.
The space includes DevZero, NVIDIA GPU Operator, Run:ai, and Volcano, each with different strengths. DevZero focuses on combined scheduling and cost optimization: GPU-aware bin packing, NVIDIA MIG support for A100s and H100s, checkpoint/restore scheduling for spot instance training, ML-based predictive scaling, and live GPU memory rightsizing without pod restarts.
LLM inference on Kubernetes is expensive and bursty - traditional tools cause disruptive restarts that break long-running sessions. DevZero solves this with live GPU rightsizing using CRIU, adjusting allocations without restarting inference pods. ML-based demand forecasting pre-scales capacity before spikes. NVIDIA MIG partitioning splits A100s and H100s into isolated slices for multi-model serving on one GPU. Customers report 40-60% GPU cost reduction with no latency degradation. Works with vLLM, TGI, Triton, and custom serving stacks on EKS, GKE, and AKS.
What our customers say