Global – November 10, 2026 -- CIQ's Fuzzball orchestration software lets operators describe an HPC or AI job in one YAML file and run it across five cloud platforms and on-premises clusters, so a new workload type no longer requires a separate stack, according to a session at AI Infrastructure Field Day.
One YAML file defines compute, containers, data and job order
A Fuzzball workflow specifies the compute each job needs, the container images it runs, the data it uses and the sequence of jobs. A training job can request GPUs, and a serving job can depend on it. Deployment targets are AWS, Google Cloud, Oracle Cloud, CoreWeave and Azure, alongside on-premises clusters. Where an operator runs more than one cluster, Fuzzball places each job by cost, performance and data locality.
Fuzzball runs on top of existing Slurm and PBS deployments
Fuzzball uses an existing Slurm or PBS deployment as its resource manager, so nothing has to move on day one. It is also a successor to that generation of resource manager, and operators can migrate fully onto it. Fuzzball deploys on a Kubernetes cluster, so teams not yet running Kubernetes must add it first.
Completed training workflows publish as autoscaling inference services
Fuzzball began with batch HPC work and now covers AI workloads, including long-running services. A finished training workflow can be published as an inference service from the same definition. The service runs as a pool that adds replicas as demand grows and releases its GPUs when idle. The control plane schedules NVIDIA and AMD GPUs and federates across clusters and clouds.
Workflow catalog and MCP server extend access to researchers and AI agents
Team members publish a workflow once, and researchers find it in a browser, fill in parameters and run it without knowing where the compute sits. The catalog also holds validated model configurations that start by name on an operator's own GPUs, with GPU counts and serving settings preconfigured.
An MCP server lets AI agents inspect the environment and draft and run workflows. Read access is open, while writes and runs wait for operator approval.
Fuzzball can be deployed without any other CIQ product, on existing clusters.