Fuzzball

Run and secure every AI and HPC workload

Fuzzball is the orchestration platform that runs and secures AI and HPC workloads. Describe the work once in provider-agnostic YAML and it runs the same way on one system or thousands of GPUs. Each job routes to wherever it runs best on cost, performance, and data locality, so models and data stay on infrastructure you control.

Fuzzball runs and secures AI and HPC workloads on infrastructure you control. Describe the work once in provider-agnostic YAML and it runs the same way on one system or on thousands of GPUs. Your own AI agents drive it over MCP, with every write and run waiting on your approval.

Get the solution brief

From the creators of Apptainer and the founding sponsor of Rocky Linux.

Runs anywhere you control

Your data center, your cloud accounts, or both inside one workflow. Priority, backfill, and preemption scheduling keep the hardware you already bought busy, and when a job needs GPUs you do not have, Fuzzball places it where they are. No second pipeline, no third-party AI service in the path.

Secures every job the same way

One identity model, one set of RBAC policies, one secrets posture across every environment, with static credentials eliminated. Every job runs in an isolated container, and policy expressions decide which groups reach which resources. Your security review happens once instead of once per environment.

Works on the first day and at every scale after

One command stands up a production-ready environment on a single system, and the identical workflows run across thousands of GPUs and federated clusters. Researchers work in a browser with no SSH and no cluster expertise, and your own AI agents drive the same workflows over MCP with every write and run waiting on your approval, so nobody waits on a platform team to get started.

Platform

Batch HPC and the full AI lifecycle, in one workflow file.

Environments

Your data center plus AWS, Google Cloud, OCI, Azure, and CoreWeave, federated as one.

Data leaves

Models, weights, and data stay on infrastructure you control.

Rip and replace

Slurm, PBS, Lustre, GPFS, and BeeGFS all stay exactly where they are.

Compare Fuzzball to a traditional HPC environment

Intersect360 Research on what a container-first, API-driven approach changes about running an HPC environment, and what it removes from your team’s week.

Get the white paper

Built for teams facing this

A sovereign-AI mandate with no plan for training and inference
Sensitive data leaving for third-party AI services
HPC and AI on separate platforms: queues on one side, idle capacity on the other
GPUs you already own sitting idle while you rent more

From the team that created Apptainer and Singularity at Lawrence Berkeley National Laboratory, and the founding sponsor of Rocky Linux. National labs, federal programs, and the largest enterprises run production on CIQ.

Compare Fuzzball to a traditional HPC environment

Intersect360 Research examined what a container-first, API-driven approach changes about running an HPC environment. The white paper covers where the model holds up, what it removes from your team’s week, and what it means for a fleet that now carries both batch and AI work.

Get the white paper

Built for teams facing this

A sovereign-AI mandate from above, and no plan for how training and inference will actually meet it
Researchers pushing sensitive data into third-party AI services before anyone checks whether the data-use agreements allow it
HPC and AI on separate platforms, so users queue on one side while capacity sits free on the other
Researchers who will not touch a command-line-only cluster, so the work leaks to laptops and shadow cloud accounts
Every new environment arrives with its own identity model, its own credentials, and its own security review
Capacity you already own sitting idle while the same work runs on rented GPUs
Jobs break the moment the infrastructure underneath them changes
An inference bill nobody can forecast for next quarter

One

Platform for batch HPC and the full AI lifecycle

One workflow definition covers train, fine-tune, validate, and serve, with inference endpoints declared alongside the training job.

Sovereign AI

The whole lifecycle on infrastructure you control

Train, fine-tune, validate, and serve in your own data center or your own cloud accounts. Models, weights, and data never move to a third-party AI service.

One

Security model across every environment

One identity model, one set of RBAC policies, one secrets posture. Static credentials eliminated, and your security review happens once instead of once per cloud.

Cloud and on-premises environments

On-premises, AWS, Google Cloud, OCI, Azure, and CoreWeave, federated into one virtual resource and routed automatically by cost, performance, and data locality.

Zero

Data migration, scheduler replacement, or SSH required

Runs against Lustre, GPFS, and BeeGFS in place, with PBS and Slurm staying on as backend provisioners. Researchers submit and monitor work from a browser.

See Fuzzball on your own workflow

Bring one workflow and the environments you have to run it in. A CIQ engineer will walk your team through the workflow file, a run on your own hardware, and the same run in a cloud, unchanged.

Get a demo

See Fuzzball on your own workflow

Bring one workflow. A CIQ engineer runs it on your hardware, then in a cloud, unchanged.

Get a demo