Upbound Submits Modelplane for Contribution to CNCF to Advance Open Source Inference Across Clouds

Modelplane gives platform and AI teams one control plane for managing models across changing GPU infrastructure, serving technologies and providers without tying applications to the infrastructure underneath them

SALT LAKE CITY, Nov. 10, 2026 (GLOBE NEWSWIRE) -- Upbound, the company behind Crossplane, today submitted Modelplane, its open source control plane for AI inference, to the Cloud Native Computing Foundation (CNCF) Sandbox. If accepted, Modelplane will bring a fleet-wide control plane for managing models and GPU infrastructure into a neutral community where providers, maintainers and enterprise users can help shape its development.

Teams are increasingly running and experimenting with different models, serving technologies and GPU infrastructure across multiple clusters and providers. Platform engineers have to make those pieces work together, find enough GPU capacity and keep applications running when a model, serving engine or provider changes.

Modelplane applies the open, declarative control plane approach Upbound brought to cloud infrastructure with Crossplane. Platform and AI teams specify which models to run and what infrastructure those models need. Modelplane coordinates deployments, serving engines and GPU capacity across a fleet of Kubernetes clusters and cloud providers. Developers get a consistent way to access inference without having to track where a model runs or how it’s served.

“Getting a model running is only the beginning. Keeping it reliable as workloads, models and GPU infrastructure change is where platform teams are breaking new ground,” said Bassam Tabbara, founder and CEO of Upbound. “We are leading an open effort to advance that work by building Modelplane, using it ourselves and sharing what works and what breaks. The CNCF Sandbox gives the community a place to work together.”

An open ecosystem for inference

AI inference is evolving into a horizontal ecosystem, with innovation happening independently across models, serving engines, accelerators and infrastructure providers. No single vendor will own every layer, and teams increasingly need to bring these technologies together rather than build around a single stack. Modelplane provides an open control plane across that ecosystem, allowing organizations to take advantage of advances at each layer while maintaining a consistent way to deploy and operate inference across their infrastructure.

At KubeCon + CloudNativeCon North America, Upbound founder and CEO Bassam Tabbara demonstrated that approach using Modelplane across infrastructure from AWS and Vultr. The demonstration showed how an application can maintain the same inference endpoint while the models, serving technologies and infrastructure underneath it change. It's a practical example of why portability matters as organizations manage increasingly diverse GPU environments.

“Modelplane makes running inference on Vultr a configuration choice, not a separate deployment model. Teams can use the same model manifests across providers and choose GPU infrastructure based on what each workload needs,” said Mayank Debnath, Director of Developer Relations, Vultr.

That interoperability extends beyond infrastructure providers to the technologies that make inference possible. Modelplane is designed to work alongside the specialized serving, scheduling and acceleration projects advancing independently across the ecosystem. NVIDIA Dynamo is one example of how those capabilities can be integrated into a common fleet wide operating model.

Modelplane integrates with NVIDIA Dynamo to bring distributed inference capabilities to fleets of GPU clusters. The integration uses Grove and the KAI Scheduler to coordinate multi-node workloads, and ModelExpress to transfer model weights directly between GPUs, reducing repeated reads from storage as replicas scale.

“Scaling AI inference requires coordination across serving engines, GPU resources and the software connecting them. NVIDIA Dynamo provides an open, modular foundation for distributed inference. Modelplane builds on that ecosystem at the fleet level, helping platform teams deploy models across environments while continuing to benefit from advances in accelerated computing,” said Itay Neeman, Senior Director, Dynamo Engineering NVIDIA.

These integrations demonstrate Modelplane's approach to bringing diverse models, serving technologies and GPU infrastructure together as one heterogeneous inference fleet. By submitting Modelplane to CNCF, Upbound is inviting the broader cloud native and AI infrastructure communities to help shape this open, fleet-wide control plane.

Enterprise customers are shaping what comes next

Enterprise infrastructure teams are already moving toward a unified, API-first, cloud-native control plane model: one platform to govern infrastructure across environments, with consistent interfaces for engineers, software and increasingly AI agents. Inference is a natural extension of that model. Modelplane brings the same declarative control plane pattern to models and GPU infrastructure, while its open foundation gives the ecosystem a neutral place to build that next layer together.

“We’re seeing large global enterprises including some of the world’s largest financial institutions future-proof their infrastructure for an AI-driven operating model,” said Pablo Luna, VP of Product Management, Upbound. “They’re standardizing on unified, API-first control planes so humans can set policy and intent while software and agents consume infrastructure through the same governed platform. Bringing inference into that model is the next step: it gives enterprises a consistent way to scale AI without creating a separate infrastructure stack.”

What Upbound Learned Running Inference on Modelplane

Upbound uses Modelplane to run its own chat, coding assistant and security agent workloads. That work has surfaced problems that aren’t apparent when a model first starts: GPU quota doesn’t guarantee available capacity, and an apparently successful request can still return an incomplete response. Upbound is sharing those lessons and using them to guide improvements to the open source project.

The six-part Drinking Our Own Champagne series documents how Upbound evaluated models for a build security agent, diagnosed failures in the serving stack and tested what it would take to expand a self-hosted coding assistant beyond a pilot.

• ⁠Drinking Our Own Champagne: What It Takes to Own Your Intelligence
• ⁠Drinking Our Own Champagne, Part 2: Finding the Corks
• ⁠Drinking Our Own Champagne, Part 3: Use Case #1, the Build Security Agent
• ⁠Drinking Our Own Champagne, Part 4: Candling the Bottles
• ⁠Drinking Our Own Champagne, Part 5: What Does It Take to Grow a Self-Hosted Coding Model from Pilot to Team?
• ⁠Drinking Our Own Champagne, Part 6: The Model Was Rarely the Problem

Availability and participation

Modelplane is available under the Apache 2.0 license at https://github.com/modelplaneai/modelplane, with documentation at https://docs.modelplane.ai.

Developers and platform teams can explore the project, share operating experience and contribute improvements.

About Upbound

Upbound is pioneering infrastructure platforms for the agentic AI era, serving Fortune 500 companies and platform engineers across more than 100 countries. The company empowers infrastructure and platform teams with intelligent control planes, based on Kubernetes and Crossplane, that provision, operate, and adapt so platforms are ready for both humans and AI agents. Upbound is the creator and primary maintainer of Crossplane, the popular open-source framework for building cloud-native control planes, with over 100 million downloads and adoption by more than 1,000 organizations in production worldwide. Upbound is also the creator of Modelplane, the open-source control plane for AI inference built on Crossplane. Founded in 2017 and headquartered in San Mateo, CA, a Series B startup backed by GV (formerly Google Ventures), Altimeter Capital, and Intel Capital, Upbound has raised $69M to date. For more information, visit www.upbound.io.

Media Contact
upbound@orogroup.com