English 
Upbound the company that created the open-source infrastructure framework called Crossplane has sent Modelplane, its control plane for AI inference to the Cloud Native Computing Foundation Sandbox. The goal of this submission is to put the project into an open-source community. In that community cloud providers, developers, maintainers and enterprise users can all work together to build the project.
This announcement was made on November 10 2026. Modelplane is meant to help teams that work on platforms and AI engineering manage models, serving technologies and GPU infrastructure across Kubernetes clusters and cloud providers. The main aim of Modelplane is to give organizations a way to deploy and run AI inference workloads consistently. This way applications are not locked into one infrastructure provider or one serving technology.
The submission is a proposal for participation in the CNCF Sandbox. It does not mean the project has been accepted into the foundation yet. Modelplane is available now, under the Apache 2.0 open-source license.
Managing AI Inference Across Multiple Clouds
As organizations grow their use of intelligence infrastructure teams face growing challenges in managing different models, inference engines, accelerators and cloud environments. Trying to handle each of these pieces on their own can lead to mess. This becomes especially difficult when workloads shift between cloud providers or when more GPU power is needed.
Modelplane solves this by offering a control plane. It builds on the approach that Upbound created with Crossplane. Teams simply describe which models they want to run and what infrastructure those models need. Modelplane then takes care of coordinating deployments managing serving engines and allocating GPU resources across Kubernetes clusters and cloud providers.
This setup gives developers a way to access inference. They no longer have to worry about where a model’s running or which serving system is handling their requests. For platform teams Modelplane provides a place to manage both infrastructure and AI workloads. It acts as a layer that brings everything together.
This difference matters a lot. Businesses today often use models and rely on more, than one infrastructure provider. Of locking themselves into one cloud or one GPU setup organizations can stay flexible. They can adapt easily when their needs change or when new technologies become available.
Integration With NVIDIA Dynamo and Other Infrastructure Technologies
Modelplane is built to work beside technologies across the AI inference ecosystem instead of replacing them. The announcement points out Modelplane’s integration with NVIDIA Dynamo, a framework for distributed inference.
Modelplane’s integration brings distributed inference capabilities to fleets of GPU clusters. Modelplane uses Grove and the KAI Scheduler to coordinate -node workloads while ModelExpress transfers model weights directly between GPUs. Modelplane aims to reduce repeated reads from storage as model replicas grow.
Upbound also showed Modelplane across infrastructure from Amazon Web Services (AWS) and Vultr, at KubeCon + CloudNativeCon North America. The demonstration revealed how an application could keep using the inference endpoint while the underlying models, serving technologies and infrastructure changed.
This provider flexibility can help organizations pick GPU infrastructure based on workload requirements, availability and cost considerations. It also shows how Modelplane connects developed technologies through a common fleet-wide operating model.
Enterprise AI Infrastructure and Operational Experience
Upbound says that enterprise infrastructure teams are starting to use control planes that are built around APIs and are designed for the cloud. These control planes help organizations manage their resources in a way across different environments. The company thinks that using AI inference makes sense as part of this approach.
When there is a shared control plane engineering teams can set rules. Define what infrastructure is needed. Software and AI agents then communicate with resources through these controlled interfaces. This could allow companies to grow their AI workloads without needing a system for managing infrastructure for each model or application.
Upbound is also using Modelplane inside the company to handle chat functions, coding assistants and security agents. The company says that this experience has shown problems with how things are set up. For example there is a difference, between the GPU space that is promised and the space that is actually available. Also sometimes requests seem to work but don’t give results.
These real-life problems are helping shape the project. Upbound is writing a six-part series. The series covers model evaluation making sure inference works reliably handling security agent workloads and expanding self-hosted coding assistants.
What Modelplane Could Mean for Japan’s AI Infrastructure Market
The development could be very useful for Japan’s growing AI infrastructure ecosystem. Japanese companies, cloud providers and research groups are looking for ways to grow AI use while still keeping control over data, infrastructure and costs.
For firms that run sovereign AI environments a vendor‑neutral control plane might let them manage inference workloads across their own data centres and outside cloud services. This idea is especially important for fields such as finance, healthcare and manufacturing where rules about data, safety and infrastructure shape how technology is put in place.
こちらもお読みください： アドバンスト・テクノロジー・リサイクル、IT資産の安全な廃棄サービスを拡大
Modelplane’s method could also help Japanese businesses avoid being too dependent on one provider of infrastructure. Companies may be able to test GPU setups and serving tools while still using the same application interface.
Still people will only adopt the system if the project is mature enough fits with business setups works reliably and supports the right infrastructure options. The CNCF Sandbox proposal could open the door to community involvement but it does not mean the solution is ready for production.
For Japan’s technology providers the project also shows how much open‑source orchestration is becoming important in the AI infrastructure field. As inference tasks spread out managing the links, between models, accelerators, serving engines and cloud settings will become an engineering problem.
日本語