Fuzzball now provisions compute on CoreWeave

The next version of Fuzzball, v3.2.0, includes preview support for the AI-focused cloud computing platform CoreWeave, making CoreWeave the second cloud environment and first “neo-cloud” that can be natively provisioned with Fuzzball.

Why CoreWeave matters for AI workloads

CoreWeave has built its business around GPU availability. While traditional hyperscalers treat GPUs as one compute option among many, CoreWeave designed its entire infrastructure around GPU-accelerated, high-performance AI workloads. Their Kubernetes-native architecture and early access to NVIDIA hardware (H100, H200, and now Blackwell GPUs) has made them a go-to option for organizations running large-scale model training and inference.

Access to CoreWeave means greater access to the latest GPU accelerators for Fuzzball users and their performance-intensive workflows. If your AI workflows need GPU density that's hard to get elsewhere, you can now orchestrate those workloads through Fuzzball while keeping your existing on-premises or AWS infrastructure in the same unified platform.

How it works

The CoreWeave provisioner follows the same model as Fuzzball's AWS provisioner. Fuzzball Orchestrate is deployed into the CoreWeave Kubernetes environment then provisions Fuzzball Substrate instances to execute jobs and services when your workflows need GPU resources. Federation between multiple Orchestrate deployments means you can split your compute between AWS, CoreWeave, and on-prem.

Because the Fuzzball platform enables workflow portability regardless of where it runs, you can develop workflows once (or select from our pre-populated workflow catalog) and run wherever the resources you need are available.

Running mixed on-prem and cloud GPU workloads? Let's discuss how Fuzzball can unify your infrastructure

What this enables

Consider an AI inference pipeline that needs to serve models at scale. With the CoreWeave provisioner, you can:

• Run Fuzzball on-prem, in AWS, and on CoreWeave
• Burst GPU-intensive inference workloads to CoreWeave
• Keep your workflow definition unchanged regardless of where compute runs
• Manage the entire lifecycle through Fuzzball's scheduler

This is the fulfillment of a long-term industry goal: hybrid, multi-cloud workload orchestration: one workflow, multiple execution environments, unified management.

Preview status

The CoreWeave provisioner ships as a preview feature in v3.2.0.If you're already running workloads on CoreWeave or evaluating it for GPU-intensive AI work, the provisioner gives you a path to orchestrate those workloads alongside your existing Fuzzball infrastructure.

What's next

CoreWeave support is one piece of the v3.2.0 release, which also includes a redesigned storage system and improvements to workflow observability. We'll cover those in upcoming posts.

For now, if you're interested in testing the CoreWeave provisioner, reach out to discuss your use case. We're particularly interested in feedback from teams running mixed on-prem and cloud GPU workloads.

---

Fuzzball is CIQ's workflow orchestration platform for HPC and AI. Learn more at ciq.com/products/fuzzball.