AI Platform Design

AI infrastructure architecture guides

Vendor-neutral explanations of the runtime and infrastructure layers behind production LLM inference on Kubernetes.

01Architecture overviewAI Platform Architecture

Explore a vendor-neutral AI platform architecture for production LLM inference, including control plane, data plane, runtime dependencies and GPUs.

Read guide →
02Traffic entry and policyAI Gateway Architecture on Kubernetes

Learn how an AI gateway uses Gateway API, HTTPRoute, external processors, Service and InferencePool backends for production inference traffic.

Read guide →
03Model serving runtimeLLM Inference Architecture

Understand LLM inference architecture from request queue and scheduler through prefill, decode, KV-cache management and token streaming.

Read guide →
04Runtime endpoint selectionInference-Aware Routing on Kubernetes

Learn why LLM workloads need inference-aware routing and how InferencePool and an Endpoint Picker select serving replicas using runtime signals.

Read guide →
05Accelerator resourcesGPU Management on Kubernetes

See how Kubernetes discovers, configures, partitions, shares and allocates GPU resources to model inference workloads.

Read guide →
06Hardware executionHow GPU Inference Works

Learn how GPU inference uses high-bandwidth memory and parallel tensor compute, and why compute, bandwidth and memory capacity limit performance.

Read guide →
07Platform operating modelControl Plane vs Data Plane in an AI Platform

Understand how an AI platform control plane configures and reconciles infrastructure while the data plane handles live model inference traffic.

Read guide →