# Control Plane vs. Data Plane for AI Infrastructure > Control plane vs. data plane separates AI governance decisions from execution, making model, agent, tool, cost, and audit controls manageable. Source: https://hyperlake.cloud/blog/control-plane-vs-data-plane Published 2026-10-07 · by Hyperlake Team · Hyperlake Video: [Watch: Control Plane vs Data Plane (2:39)](https://www.youtube.com/watch?v=ZGZSu8A-ELg) A control plane decides what should happen, while a data plane performs the work. In AI infrastructure, the control plane sets access, routing, policy, cost attribution, and audit rules; the data plane runs models, agents, GPU workloads, and tool calls according to those decisions. Separating them makes complex systems governable. This distinction matters because AI execution can appear healthy even while unanswered questions about access, ownership, policy, and accountability accumulate. The video above walks through the core ideas. ## What is the difference between a control plane and a data plane? The control plane makes decisions and defines desired state. The data plane handles traffic, computation, and actions according to those decisions. Keeping these responsibilities separate makes large systems easier to manage. A control plane can apply consistent rules across many execution environments without becoming the path through which every packet, inference request, or agent action must flow. The distinction does not mean that enforcement happens only inside a central control service. Enforcement points may sit close to workloads, gateways, data systems, or tools. The control plane defines and distributes the policy, while the data plane applies it during execution and reports relevant events back for observation and audit. This separation also limits responsibility. The data plane should not independently decide which policies govern a request, and the control plane should not become the system doing every unit of application work. ![Diagram: The control plane defines decisions and policy while the data plane executes work and reports events.](https://hyperlake.cloud/blog/img/production/4db69cf376670dad6fce6ad4516d5e0a244dd950-1200x750.png?w=1600&fit=max&auto=format) *The two planes stay separate but coordinate through policies, enforcement, and operational feedback.* ## How do control planes and data planes work in networking and Kubernetes? Networking provides the classic example: the control plane determines routes and defines policies, while the data plane forwards packets according to those decisions. The control plane does not carry ordinary network traffic, and the data plane does not independently create routing policy. Kubernetes applies the same architecture to compute. Its control plane includes components such as the API server, scheduler, and controller manager. These components accept desired state, select where workloads should run, and continually reconcile the cluster toward that state. Worker nodes form the execution layer. They run containers and provide the CPU, memory, networking, and other resources the workloads consume. The Kubernetes control plane governs placement and state; the nodes perform the actual work. AI infrastructure follows this established pattern rather than introducing a completely new architectural concept. The difference is the scope of decisions: AI control planes must account for models, agents, tools, enterprise data, identity, policies, costs, and audit evidence. ## What belongs in an AI control plane and data plane? The AI data plane contains the systems that execute requests and actions. The AI control plane determines what those systems are permitted and expected to do. Common data-plane components include: - GPU pools performing model inference or training workloads. - Model deployments receiving and serving requests. - Agents executing workflows and maintaining task state. - MCP tool integrations carrying out searches, updates, or other actions. - Data and knowledge services responding to authorized queries. The control plane defines which models and tools can be accessed, how requests should be routed, which policies apply to each agent or workflow, how costs should be attributed, and what must be recorded in an audit trail. These concerns are broader than model deployment alone, as explained in [agentic control plane architecture](https://hyperlake.cloud/blog/what-is-an-agentic-control-plane). Identity must also flow into these decisions. A request associated with a user, service, or agent needs a scoped identity so policy systems can decide what it may access. Cost records similarly need enough context to connect consumption with the responsible workload, product, customer, or team; [LLM cost attribution](https://hyperlake.cloud/blog/llm-cost-attribution-who-owns-which-part-of-the-ai-bill) examines that problem in more detail. ## When should teams add an AI control plane? Teams should add an AI control plane when execution expands beyond a small, easily inspected deployment. Building the data plane first is a natural adoption path, but waiting too long allows governance and accountability gaps to accumulate. The warning signs often appear as questions the organization cannot answer reliably: - Which agent accessed a particular sensitive data source? - Which team or product is responsible for a model cost? - Which policy governed a specific workflow or tool call? - What record can investigators inspect after an incorrect action? These gaps may remain hidden because the data plane continues operating. Agents still complete tasks, models still return responses, and GPU workloads still run. The architecture looks functional until a security review, cost investigation, incident, or audit requires evidence that was never captured. A practical control-plane rollout starts by identifying workloads, identities, data boundaries, model and tool permissions, routing rules, cost ownership, and required audit events. Teams can then place enforcement and logging at the relevant execution points rather than relying on shared credentials or application-specific rules. ![Diagram: Four warning signs indicate that an AI deployment needs a control plane for governance and accountability.](https://hyperlake.cloud/blog/img/production/0485cbc66d216a189874182084072f6163883ec3-1200x750.png?w=1600&fit=max&auto=format) *Unanswered access, cost, policy, and audit questions signal a growing control-plane gap.* ## Key takeaways - The control plane decides and governs, while the data plane executes requests and actions. - Networking and Kubernetes demonstrate why separating these responsibilities improves manageability. - AI data planes include models, GPUs, agents, tools, and data services. - An AI control plane governs access, routing, policy, cost attribution, and audit records. - Building execution first is natural, but governance should arrive before unanswered questions become operational risks. ## How Hyperlake helps Hyperlake provides a sovereign control surface for assembling, deploying, governing, observing, and maintaining AI data, models, applications, and tools in infrastructure the customer controls. It supports identity-aware access, network isolation, scoped secrets, policy decisions, audit, lineage, and reusable deployment patterns, with operational details depending on the engine and deployment. To discuss how these controls apply to your environment, [talk to our team](https://hyperlake.cloud/contact). ## Frequently asked questions ### Can an AI control plane be added after models and agents are deployed? Yes. Building execution capabilities before centralized governance is a common and natural order of adoption. The control plane can be introduced incrementally by inventorying workloads, assigning identities, defining access and routing policies, establishing cost ownership, and capturing audit events at the relevant enforcement points. ### Does an AI control plane replace the Kubernetes control plane? No. Kubernetes governs cluster resources, workload placement, and desired state, while an AI control plane addresses higher-level concerns such as model and tool access, agent policy, request routing, cost attribution, and AI audit records. An AI control plane can use Kubernetes as its infrastructure foundation while adding controls specific to AI systems. ### Why can a missing AI control plane go unnoticed? The data plane can continue functioning without centralized governance: models answer requests, agents execute tasks, and tools perform actions. The gap becomes visible when someone needs to determine who accessed sensitive data, which policy applied, who owns a cost, or what happened during an incident and finds that no authoritative record exists.