Nvidia released a preview of its Kubernetes Engine (NVKE), a management plane that platform teams install on their own GPU infrastructure to create and operate Kubernetes clusters, each with its own control plane, etcd and node pools. It is explicitly a preview not recommended for production, with APIs and behavior subject to change.
Control Plane
控制平面
Core layer managing, coordinating, and monitoring distributed system components
Curated Intelligence
Buoyant has released Kinference, an open-source control plane designed to build scalable inference platforms using vLLM on Kubernetes. It supports OpenAI/Anthropic-compatible APIs and features like token rate limiting, tenant cost tracking, and intelligent load balancing, achieving approximately 3.5x goodput improvement in multiturn workloads.
LangSmith launches its Bring Your Own Cloud (BYOC) mode, supporting deployment on AWS, Azure, and Google Cloud. This architecture separates the control plane from the data plane, allowing users to retain full ownership of traces, evals, and agent data while LangChain manages the underlying infrastructure operations.
Developer releases kube32, a Kubernetes-compatible control plane and WebAssembly pod runtime designed specifically for ESP32-S3 microcontrollers. The project enables managing hardware clusters via standard kubectl, executing WebAssembly modules as Pods using the wasm3 interpreter, thereby bringing cloud-native workload capabilities to resource-constrained IoT devices.
Evan Conrad of SF Compute outlined the 'Programmable Data Center' vision at AIE NYC, extending software-defined networking principles to GPU clusters. The platform enables programmatic control over the full stack from network to memory and facilitates instant buying and selling of GPU compute, lowering barriers to building supercomputing infrastructure.
SF Compute extends software-defined networking principles to GPU clusters, enabling programmatic control over resources from network to memory. This approach lowers the barrier to building supercomputers and facilitates instant trading of GPU capacity to optimize balance sheets.
Microsoft Research Asia released Agent Lightning v1.0, a lightweight agentic reinforcement learning framework comprising only ~3,500 lines of code. Its key innovation is the 'Harnessed Agentic RL' paradigm, which allows real deployment harnesses (e.g., mini-SWE-agent) to participate directly in training without reimplementation. The framework natively supports Kubernetes, eliminating dependencies on expensive commercial sandbox services and significantly lowering the barrier to agent RL training.