Reference architectures
How Klique is deployed, drawn for IT architects and sales engineers.
Meta-Scheduler Reference Architecture
One control plane schedules GPU work across Kubernetes clusters, Slurm, PBS and LSF clusters, bare metal and VMs, public cloud providers and neoclouds. Each environment keeps its own scheduler, and governance layers sit above it: identity, tenants, services, tokens and billing.
Multi-Tenant Reference Architecture
A sovereign cloud provider, or an enterprise serving its own business units, runs one white-labelled control plane and serves many isolated tenants from a shared GPU estate. Tenant isolation, allocations, tokens as a service, and metering through to the bill.
GPU as a Service Reference Architecture
An enterprise platform team offers GPUs as a service internally through one control plane. Kubernetes, bare-metal and HPC estates, interactive and non-interactive work, and every provisioning path governed the same way.
Cloud Autoscaler Reference Architecture
The control plane runs on-prem and GPU work bursts to a cloud region when on-prem capacity is full, through VM autoscaling or managed Kubernetes. One LLM gateway on-prem routes and meters every model call.
LLM Gateway and Tokenomics Reference Architecture
An enterprise IT department provides LLMs as a service through one LLM gateway in front of open-weight models on owned GPUs and public model providers. One alias can send each request to a different model by rule, with guardrails, budgets and cost attribution.
