Your AI infrastructure,
frontier velocity.
One platform for all your AI compute — Kubernetes, Slurm, VMs, on-prem, and 20+ clouds.
No servers to configure, no environments to build.
- k8s / prod-us-westH100 × 512
- slurm / hpc-legacyA100 × 256
- vm / burst-awsH200 × 128
- k8s / onprem-dc-aL40S × 96
- vm / gpucloud-euH100 × 64
AI Compute Platform
Turn fragmented compute into frontier intelligence.
Kernpool brings in all your AI compute — Kubernetes, Slurm, VMs, on-prem — and runs the entire AI lifecycle on top of it, at frontier velocity.
| CLUSTER | TYPE | ACCELERATORS | UTILIZATION |
|---|---|---|---|
| prod-us-west | k8s | H100 × 512 | 91% |
| hpc-legacy | slurm | A100 × 256 | 78% |
| burst-aws | vm | H200 × 128 | 64% |
| onprem-dc-a | k8s | L40S × 96 | 88% |
| gpucloud-eu | vm | H100 × 64 | 73% |
Capabilities
One platform, frontier velocity
Manage any AI compute
Any cluster, any cloud, any Kubernetes or Slurm — under one interface. Onboarding a new cluster takes no workload migration and no changes to your training code.
| NODE | ACCELERATORS | WORKLOAD | STATE |
|---|---|---|---|
| node-a01 | 8 × H100 | pretrain-7b | Running |
| node-a02 | 8 × H100 | pretrain-7b | Running |
| node-a03 | 8 × H100 | rl-rollout | Scheduling |
| node-b11 | 8 × H100 | — | Idle |
| node-b12 | 8 × H100 | batch-infer | Running |
Capabilities for frontier AI teams
From CLI to intelligent scheduler, GPU monitoring, and quotas, Kernpool gives your infrastructure frontier velocity. Environment and job as code — one config runs on any compute.
Runs your existing GPU, TPU, and CPU workloads with no code changes.
# one config, runs on any computeresources:accelerators: H100:8any_of: [prod-us-west, burst-aws, gpucloud-eu]num_nodes: 64priority: highsetup: pip install -r requirements.txtrun: torchrun --nnodes=$KP_NUM_NODES train.py
Coverage
Trusted by leading compute providers
Hyperscalers, AI neoclouds, GPU marketplaces, and the machines already in your rack — all under one control plane.
Orchestration & your own machines
8StableBring existing clusters and hardware in as they are.
Hyperscalers
6StableReserved capacity, on-demand, and spot across regions.
AI neoclouds & GPU rental
13ElasticGPU-first providers and marketplaces, onboarded the same way.
Onboard a new provider in under 15 minutes — no changes to the workloads above it. Kernpool is BYOC: everything launches inside your own accounts, VPCs, and clusters.
The AI lifecycle
Fast-moving AI teams, faster.
Kernpool gives AI teams a simple interface to run the entire AI lifecycle, so everyone moves faster.
Development
Spin up in seconds. Connect over SSH or your IDE, or run fleets of coding agents. Idle devboxes stop themselves, so nobody has to remember.
Pre-training
Scale to thousands of nodes. When a GPU fails, it is swapped out and the run continues — one bad card no longer costs you the whole job.
Post-training & RL
Co-schedule RL components on heterogeneous hardware. Trainers, samplers, and environments each land on the hardware that suits them.
Batch inference
Run cost-efficient, fault-tolerant batch workloads. Failed tasks retry on their own, and preempted instances resume where they stopped.
Sandboxes
Sub-second sandboxes for agents and RL, on your own infrastructure. Spawn them in bulk, tear them down instantly, hold no standing capacity.
Endpoints
Production-ready inference on every cluster you own. Cross-cluster routing and canary releases work out of the box.
For infra teams
Supercharge your AI infra
Infra teams orchestrate every cluster and cloud seamlessly. Being able to run on any compute future-proofs your infrastructure.
Easily add GPUs, maximize utilization
Add a new cluster or GPU cloud in minutes. Kernpool pools all your compute providers into one resource pool and removes fragmentation.
Scalable control plane
Onboard new users, teams, or workloads with ease. Kernpool's control plane scales with you.
AI abstractions for Kubernetes
Kernpool makes K8s AI-native: multi-cluster support, topology-aware scheduling, quotas, and preemption.
Bring machines you already have
A list of IP addresses you can SSH into is enough. On-prem racks and leftover reserved capacity join the same pool as everything else.
Scaling out
Multi-cloud GPU infrastructure
When you need to, scale to new providers with confidence.
Cost and capacity optimizer
Kernpool searches every provider you have enabled in cheapest-and-available order, and fails over automatically when capacity runs out.
GPU health checks
Proactive and reactive health checks across the fleet, with auto-remediation of GPU and NCCL-related faults.
BYOC by default
Every cluster, job, and endpoint launches inside your own accounts, VPCs, and Kubernetes clusters. Your data never leaves them.
Editions
Start open source. Scale on the platform.
The orchestration layer is open source and self-hosted. The platform adds the fleet operations, governance, and uptime a growing team needs.
Open source
Free, self-hosted, Apache 2.0. Install it this afternoon.
$ pip install kernpool- Run on Kubernetes, Slurm, VMs, and 20+ clouds
- Environment and job as code
- Cost and capacity optimizer with auto-failover
- SSH and IDE access, managed jobs, endpoints
- Community support
Kernpool Platform
ManagedFor teams operating large GPU fleets across several providers.
- Everything in open source, plus:
- High-availability managed control plane
- Multi-cluster fleet management and GPU health checks
- Priority queueing, quotas, and workspaces
- SSO, RBAC, and policy enforcement
- Cost reporting and chargeback
- Dedicated support
Governance
Enterprise ready
Run the platform inside your own perimeter, with the controls your security and finance teams ask for.
Secure, on your premises
- BYOC and BYOK
- Private VPCs and airgapped deployments
- Use open source or the enterprise platform
Increase utilization, control AI spend
- Auto-stop idle compute
- Advanced quota management
- Cost management and reporting
Team controls and uptime
- High-availability control plane
- Fast onboarding with SSO
- Policy enforcement, RBAC, workspaces