CIQ’s Fuzzball Adds Azure to Its Five-Cloud AI Orchestration Layer

CIQ has added Microsoft Azure as the fifth cloud its Fuzzball platform runs AI and HPC jobs on, alongside AWS, Google Cloud, Oracle Cloud Infrastructure and CoreWeave. Enterprises write one workflow file, and Fuzzball decides where each job actually runs, based on cost, GPU availability and data residency rules.

The rollout puts Azure inside a routing engine that ranks it against four rivals on price and GPU access, job by job. That is a different position than simply being wherever a team happened to start building.

One Workflow File Now Runs on Five Different Clouds

Fuzzball’s core idea is a single definition file describing compute jobs, data movement, container images and resource requirements. That file carries no cloud-specific code. CIQ’s orchestration layer translates it into whatever sits underneath, whether that is a hyperscale cloud, CoreWeave’s GPU fleet or a company’s own servers.

CIQ describes what that looks like in practice with three scenarios:

• A genomics team validates a sequencing pipeline on AWS, then moves it to Azure or Oracle Cloud Infrastructure without changing a line of the workflow definition.
• A training run that needs dense H100 graphics processing unit (GPU) access routes automatically to CoreWeave.
• A data-sensitive simulation stays on premises, kept there by policy instead of a person manually blocking it from leaving.

CIQ’s own product page also points to a live example: FYR’s EV-Omics platform, which processes multi-omics data to identify precision medicine targets, is listed as a workload already running on Fuzzball’s routing engine that places jobs by cost, performance and data locality.

In its announcement, CIQ said the update completely levels the playing field for teams that previously rebuilt pipelines and rewrote deployment scripts every time they added a provider.

Why Enterprises Are Spreading AI Jobs Across Five Providers

Multi-cloud strategies have become more common among AI and HPC teams chasing scarce GPU supply, comparing prices between providers, and keeping some workloads inside specific jurisdictions or their own data centers. CoreWeave in particular has grown into a major supplier of GPU-dense capacity that hyperscalers themselves cannot always match on availability.

CIQ is betting that pressure only grows. Back when Fuzzball first launched, the company cited International Data Corporation (IDC) projections that the performance-intensive-computing-as-a-service market would grow from $22.3 billion in 2021 to $103.1 billion by 2027, a compound annual growth rate near 28 percent, in a 2023 launch announcement introducing Fuzzball.

AI teams today are asked to ship faster, control costs and maintain sovereignty over their data, simultaneously, across infrastructure that was never designed to work together. We built Fuzzball to solve that problem at the architectural level. When your workflow definition abstracts its requirements properly, you get portable access to every GPU environment the market offers and the freedom to route to wherever the best price, performance and data policy lives.

Gregory Kurtzer, chief executive officer and founder of CIQ, made that case in the announcement, adding that controlling infrastructure and workloads is what enterprise AI requires for production.

Azure Joins the Rotation Alongside Four Rivals

Microsoft has been building its own reasons for enterprise AI teams to stay put on Azure. The company’s agentic AI push has already signed KPMG and Atos as partners, even as some investors have questioned the spending required to get there. Microsoft has also built a shared memory layer for enterprise AI agents so tools built on Azure keep context across sessions.

None of that changes what Fuzzball does to Azure’s position once a workload is defined in its format. Azure competes for that job the same way Oracle Cloud Infrastructure or CoreWeave do, on whatever price and GPU capacity each can offer that day. Bjorn Hovland, president of CIQ, framed the shift from the customer’s side.

“Fuzzball turns multi-cloud from a liability into a competitive advantage,” Hovland said. “Five clouds used to mean five identity and access management (IAM) models, five deployment pipelines and five sets of operational overhead, with complexity and risk multiplied.”

One Identity Model Replaces Five

Each cloud deployment is provisioned through a two-phase automated process that stands up a complete, production-ready cluster without manual setup. Underneath, Fuzzball keeps one IAM model, one set of role-based access control (RBAC) policies and one approach to secrets management across every environment it touches.

Static credentials disappear in favor of each cloud’s own native identity system.

On Oracle’s cloud specifically, CIQ’s own engineering documentation describes how a deployment uses Dynamic Groups and IAM policies to eliminate static credentials, the same no-keys approach applied on every other cloud it supports.

From a 2023 Launch to Five Clouds in Three Years

The five-cloud release did not appear overnight. It caps a steady sequence of releases stretching back to Fuzzball’s original launch.

1. May 18, 2023: CIQ launches Fuzzball, calling it “HPC 2.0” and pitching one interface across cloud and on-premises systems.
2. March 27, 2025: CIQ ships Federation, letting Fuzzball route jobs between on-premises clusters and the cloud, and lists Fuzzball on the AWS Marketplace.
3. June 4, 2026: CIQ announces full multi-cloud support across CoreWeave, AWS, Google Cloud, Oracle Cloud Infrastructure and Microsoft Azure, the release at the center of this story.
4. June 11, 2026: Fuzzball 4.0 ships, aimed at national laboratories and HPC centers.
5. June 30, 2026: Fuzzball extends the same portability down to a single NVIDIA DGX Spark desk-side system.

That June 11 release mattered for an audience beyond hyperscaler customers. Fuzzball 4.0 connects directly to existing multi-petabyte Lustre, GPFS and BeeGFS storage systems without requiring data migration, a detail aimed at labs that were never going to move their archives for a new orchestration layer.

Fuzzball’s Next Stop Is a Single Desktop Box

CIQ did not stop at clouds. Its newest move takes the same portability down to a single machine: NVIDIA’s DGX Spark.

Hovland said DGX Spark “is the first platform this runs on, and it will not be the last.” With hundreds of built-in workflow templates, CIQ says the same workflows that run on one Spark scale unchanged to thousands of GPUs and full data-center infrastructure.

NVIDIA’s Spark branding has also reached into other corners of computing, including giving Windows PCs another shot at Arm. CIQ’s next public test of its own roadmap comes July 16, when it walks customers through taking a self-hosted large language model to production on Fuzzball, one more claim the five-cloud pitch will need to keep proving, one workload at a time.

Frequently Asked Questions

What is CIQ’s Fuzzball platform?

Fuzzball is CIQ’s orchestration platform for AI training, inference and HPC workloads, built around two internal layers. Substrate handles the container runtime and per-node resources on each machine, while Orchestrate schedules multi-step workflows and the data those workflows need, across whatever infrastructure is available.

Which clouds and systems does Fuzzball support today?

Fuzzball currently runs on five public clouds, CoreWeave, AWS, Google Cloud, Oracle Cloud Infrastructure and Microsoft Azure, plus on-premises clusters built on Warewulf, VMware or bare metal. CIQ describes on-premises support as a first-class target rather than a secondary option bolted on for compliance.

How does Fuzzball decide where a specific job runs?

Fuzzball’s Federate layer evaluates cost, performance and data locality at runtime, then places each job automatically without an engineer choosing a destination by hand. CIQ says placement policies can also pin regulated workloads to infrastructure a company directly controls, overriding whatever the cost calculation alone would otherwise recommend.

Does Fuzzball replace schedulers like Slurm or PBS?

No. Fuzzball integrates with schedulers instead of replacing them. Teams can submit batch jobs through PBS or Slurm, and that existing scheduler keeps acting as the backend provisioner underneath Fuzzball’s orchestration layer, so nothing already in production has to be ripped out.

How can a company evaluate Fuzzball before committing to it?

CIQ points prospective customers toward requesting a live demo inside their own cloud environment rather than a generic sales pitch. Its own Oracle Cloud Infrastructure documentation, for example, walks through standing up a full deployment inside a customer’s actual tenancy step by step.

Is Fuzzball CIQ’s only product?

No. CIQ is also the founding support and services partner for Rocky Linux and sells hardened operating system variants alongside it. Arthur Tyde, the company’s senior vice president of global business development, has argued publicly that AI infrastructure security has to come before performance, pointing to hardened builds like RLC-H as the alternative to training models on unhardened Linux distributions.