CIQ Fuzzball

Fuzzball is CIQ's AI and HPC workload orchestration platform for running model training, fine-tuning and inference on infrastructure an organisation controls, on premises or across major clouds. It is built for research-computing, platform and ML-infrastructure teams that want reproducible, portable workflows and governed AI-agent access without needing deep HPC expertise.

Fuzzball is a workload orchestration platform from CIQ, the Reno, Nevada company founded in April 2020 by Gregory Kurtzer, the original creator of CentOS, who also built Warewulf and developed Singularity (now Apptainer) at Lawrence Berkeley National Laboratory. CIQ is best known as the founding commercial sponsor of Rocky Linux. Fuzzball layers on top of existing schedulers, storage and clusters so that training, fine-tuning, validation and inference run as portable, reproducible workflows that users can submit, monitor and reproduce from a browser interface, the CLI or the API. The same workflow definition runs on premises (via Warewulf, VMware or bare metal) and on AWS, Google Cloud, Oracle Cloud, Azure and CoreWeave. That portability underpins CIQ's sovereign-AI pitch: models, weights and data stay on infrastructure the organisation owns and governs, and every action is auditable. In January 2026 CIQ added service endpoints, which put persistent inference services in the same pipeline as batch jobs. Fuzzball 4.2, released on 3 September 2026, adds an MCP server so an existing AI agent can inspect an environment, draft and submit workflows and monitor them, within operator-set privilege limits on writes, workload execution and destructive operations. Every job and service container now receives a workflow-scoped API credential automatically, so a running workflow can submit, track and stop further workflows without long-lived shared secrets. The release also brings AMD ROCm GPU support alongside NVIDIA, host and GPU health signals for scheduling with automatic job restart on node faults, organisation-level storage isolation, rootless workload execution and per-workflow accounting for compute, storage and network egress. CIQ's own engineers describe the accounting and node monitoring as initial capabilities that are still maturing. Fuzzball is sold through CIQ and on AWS Marketplace, where it deploys as a Helm chart on Amazon EKS, and CIQ runs a startup programme that includes six months of free access.

Category

Infrastructure & Cloud

Pricing

Usage-based, Contact for pricing

Target Market

CTOs, Heads of AI, HPC and Research Computing Teams, ML Platform Engineers, Government and National Labs

Deployment

Hybrid, Multi-cloud, Self-hosted

Founded

2020

Headquarters

Reno, United States

• ✓
  MCP server for AI agents

  Lets an existing AI agent inspect a Fuzzball environment and draft, submit and monitor workflows within operator-defined privilege limits.

• ✓
  Workflow-scoped credentials

  Every job and service container receives a scoped API credential automatically, so workflows can launch further workflows without long-lived shared secrets.

• ✓
  Service endpoints

  Persistent inference services run in the same portable pipeline as batch training and fine-tuning jobs, covering the full model lifecycle.

• ✓
  Multi-cloud and on-prem portability

  The same workflow definition runs on premises and on AWS, Google Cloud, Oracle Cloud, Azure and CoreWeave without modification.

• ✓
  Health-aware scheduling

  Host and GPU health signals feed reliability scoring, and jobs restart automatically after node faults while bad hardware is cordoned and replaced.

• ✓
  Multi-tenant isolation

  Organisation-level storage isolation, compute-policy grants and rootless, unprivileged execution let multiple teams share the same clusters safely.

• ✓
  NVIDIA and AMD GPU support

  Fuzzball 4.2 adds AMD ROCm support alongside NVIDIA, giving buyers more flexibility when sourcing accelerators.

• •
  Sovereign model training and serving

  Train, fine-tune and serve models on owned hardware so that proprietary data and model weights never leave the organisation's control.

• •
  Cloud bursting for HPC and AI

  Keep jobs on premises by default and burst the same workflows to major clouds or CoreWeave when local capacity runs short.

• •
  Agent-driven compute operations

  Point an AI agent at the MCP server to draft, submit and monitor training or simulation jobs while operators block destructive actions.

• •
  Broadening access to research computing

  Give scientists and analysts browser-based job submission, monitoring and reproduction instead of requiring command-line scheduler expertise.

CIQ does not publish a general list price for Fuzzball on its product page; direct deals go through CIQ sales. On AWS Marketplace it is metered per minute: $0.20 for Fuzzball Orchestrate and $0.025 for each Substrate CPU or GPU node, with no free trial listed. VC-backed AI startups (seed to Series D) get six months free, then 80% off for months 7-12 and 50% off through month 24, before moving to standard pricing.

This page was written from 10 sources, 4 on domains other than ciq.com.

1. 1.ciq.com — fuzzballvendor
2. 2.ciq.com — fuzzball 4 2 ai agents that drive fuzzball and workflows thavendor
3. 3.ciq.com — ciq delivers turnkey sovereign ai with service endpoints capvendor
4. 4.ciq.com — startupsvendor
5. 5.ciq.com — aboutvendor
6. 6.ciq.com — software infrastructure leader ciq closes 26 million series vendor
7. 7.globenewswire.com — fuzzball 4 2 gives enterprises one platform to build train a
8. 8.techstrong.ai — ciq fuzzball 4 2 brings ai agents into compute orchestration
9. 9.hpcwire.com — ciq launches fuzzball 4 2 with agentic ai and amd gpu suppor
10. 10.aws.amazon.com — prodview 2fgl7v635pm4y