Upstream Kubernetes
Vanilla upstream Kubernetes, no forks. Every kubectl, every Helm chart, every operator you already use works out of the box — no proprietary lock-in.
/ KUBERNETES SERVICE
Managed Kubernetes built for AI
Tscale Kubernetes Service delivers a production-grade, GPU-aware managed K8s — with one-click GPU operators, intelligent auto-scaling, and multi-tenant isolation. Standard upstream Kubernetes, optimised for the workloads that matter.
GPU Native
NVIDIA Device Plugin, GPU Operator, and DCGM come pre-installed. Schedule pods onto specific GPUs, share devices with MIG, and monitor utilisation out of the box.
Production Hardened
HA control plane, encrypted secrets at rest, CIS-benchmarked nodes, and 24/7 SRE. Production-grade from day one, with a 99.95% uptime SLA.
/ GPU OPERATORS
One-click operators for the GPU stack
Tscale ships every operator your AI workloads need, pre-configured and tested together. No more wiring up NVIDIA drivers, runtime configs, and device plugins at 2am — just enable and ship.
- NVIDIA GPU Operator — drivers, container toolkit, device plugin, and DCGM exporter, all from a single Helm install.
- MIG manager — slice H100/A100 GPUs into isolated instances for multi-tenant inference workloads.
- KubeRay operator — first-class Ray cluster support for distributed training and serving.
- Volcano scheduler — gang scheduling, fair-share, and topology-aware placement for batch AI jobs.
- cert-manager + sealed-secrets — encrypted GitOps-friendly secret management, out of the box.
/ AUTO-SCALING
Right-size, automatically to your traffic
Tscale’s auto-scaler understands GPU workloads. Scale up fast when a request spike hits your inference endpoint, scale down to zero when notebooks go idle — and never over-provision a single GPU.
-
HPA + Cluster Autoscaler
Standard Kubernetes HPA scales pods. Cluster Autoscaler scales nodes. Both tuned for GPU workloads.
-
KEDA event-driven scaling
Scale from queue depth, request rate, Kafka lag, or any custom Prometheus metric.
-
Scale to zero
Dev environments and bursty workloads scale to zero — and warm back up in seconds when a request arrives.
/ PLATFORM
A complete cloud-native stack
Tscale’s Kubernetes Service ships with the entire ecosystem of CNCF projects you’d otherwise have to install, configure, and maintain yourself. Production-grade from the moment you create a cluster.
Core Platform
- Kubernetes 1.30+ (latest stable)
- HA control plane (3 etcd replicas)
- Cilium CNI with eBPF
- CoreDNS with split-horizon
- Gateway API ingress
GPU & Acceleration
- NVIDIA GPU Operator
- MIG slicing support
- DCGM + Prometheus exporter
- AMD ROCm operator
- Time-slicing & MPS
Storage
- High-performance CSI driver
- ReadWriteMany (RWX) volumes
- S3-compatible object storage
- Lustre CSI for HPC
- Snapshots & clones
Operators
- cert-manager
- sealed-secrets
- Volcano (gang scheduling)
- KubeRay
- Knative (serverless)
Observability
- Prometheus + Grafana
- Loki log aggregation
- Jaeger distributed tracing
- DCGM GPU dashboards
Developer Tooling
- kubectl + Helm + Argo CD
- Tscale Radar (REST + CLI)
- Terraform provider
- VS Code extension
Performance
Production control plane
Multi-AZ control plane with automatic failover and 24/7 SRE on call.
Pod to ready in 42s
From CPU spike to new pod scheduled and serving traffic in under a minute.
Rolling upgrades
K8s version upgrades happen without dropping a single connection.
Hardened by default
Every node ships hardened to CIS Kubernetes Benchmark Level 1 — no extra work for you.
Built for AI-native teams
Multi-tenant Isolation
Namespaces, RBAC, network policies, and resource quotas out of the box. Run dozens of teams on one cluster without stepping on each other.
Learn MoreSecure & Compliant
SOC 2 Type II controls, encrypted secrets, audit logging, and network policies — for teams shipping AI in regulated industries.
Learn MorePair with the rest of the stack
Kubernetes Service is the runtime for everything Tscale builds. Combine it with managed Slurm, inference, and fine-tuning to run the full AI lifecycle on one platform.