CIQ adds Fuzzball for NVIDIA DGX Spark AI workloads

CIQ has added a new Fuzzball feature for NVIDIA DGX Spark that creates a ready-to-run environment for AI development and inference.

The update is aimed at teams that want to start with private local inference on a single system, then move the same workloads to larger GPU infrastructure without rebuilding the software environment.

The latest version of Fuzzball provides what CIQ describes as a production-ready compute and inference setup for the DGX Spark stack. Users can develop, tune and deploy AI workloads in one environment across hardware they control, from a single machine to larger GPU clusters and data-centre infrastructure.

NVIDIA DGX Spark is the first supported platform for this approach. More platforms are planned, though CIQ did not identify them.

Private inference

CIQ is targeting a problem familiar to AI infrastructure teams: building the stack needed to move a model from experimentation into production. That often means assembling storage, container registries, schedulers, inference servers and deployment pipelines, then repeating parts of the process when the underlying compute changes.

Fuzzball is intended to remove that rebuild cycle. CIQ said organisations can tune models privately on DGX Spark, then move the same containers, model assets and workflow definitions to larger NVIDIA GPU deployments without changing the application or deployment model.

That argument is likely to resonate with organisations in regulated sectors and others pursuing sovereign AI strategies, where data must remain on infrastructure under direct control. CIQ said the platform is designed specifically for private, local inference on DGX Spark and can shorten the path from model development to a running inference service from months to days.

Gregory Kurtzer, Chief Executive Officer and Founder of CIQ and Founder of Rocky Linux, described the product as a way for teams to retain operational control as projects scale.

"Fuzzball is the Kubernetes of performance-intensive computing, and it is what AI teams have needed to truly own their infrastructure. The hard part of AI has never been the model. It has been operating that model at scale without rebuilding everything underneath it every time compute changes, and most teams spend months on that problem before a single workload reaches production. Fuzzball ends that. With hundreds of built-in workflow templates, a single DGX Spark becomes a complete environment for AI development, testing and validation from day one, and those same workflows run unchanged across thousands of systems and GPUs. CIQ stands for controlling intelligence. Fuzzball is how teams actually do it," Kurtzer said.

Scaling path

CIQ said a single DGX Spark can serve as a complete development and deployment environment from the outset, with a path to larger installations as demand rises. Multiple DGX Spark systems can also operate together under one Fuzzball environment, allowing smaller teams to expand local compute resources in stages.

When workloads outgrow local hardware, users can shift them to larger NVIDIA deployments, including NVIDIA GB300 NVL72 systems, without altering the application, orchestration model or deployment process, according to CIQ. That continuity between a desktop-scale or local setup and a larger cluster is central to the company's pitch.

CIQ also framed the release as a bridge between AI and high-performance computing teams, which have traditionally worked with separate software toolchains. By putting job-based orchestration, workflow portability, containerised execution and inference services under a single operational model, Fuzzball can connect local DGX Spark systems, existing HPC clusters, cloud resources and larger GPU estates, the company said.

Bjorn Hovland, President of CIQ, linked the release directly to concerns over data control in regulated industries.

"Organisations in regulated industries have had to choose between moving AI to production and keeping data on infrastructure they control. That has never been an acceptable tradeoff. Fuzzball removes it. Teams can tune models privately on DGX Spark, operationalize those models as production inference services and expand onto larger controlled infrastructure without changing the environment underneath their work. For sovereign AI to be practical it has to run the same way at every compute tier, and that is exactly what Fuzzball on DGX Spark delivers. DGX Spark is the first platform this runs on, and it will not be the last," Hovland said.

CIQ is best known as the company behind Rocky Linux. Its portfolio also includes products for Linux infrastructure, IT automation, cluster provisioning and container orchestration for high-performance computing. Adding DGX Spark support extends that portfolio further into AI operations for organisations that want to build and run models on systems they manage themselves.

As many companies weigh whether to keep AI workloads on premises, in dedicated facilities or in the cloud, the ability to preserve the same containers, model assets and workflows across different compute tiers is becoming a more prominent issue for infrastructure buyers.

For CIQ, the immediate significance is giving DGX Spark users a packaged route from local model work to production inference on larger controlled systems, without having to reassemble the software foundation each time the hardware changes.