Series✓ Complete
Tested onNKP 2.19Nutanix Enterprise AI 2.8.1Envoy AI GatewayLiteLLMVelero 1.18NDK 2.3.0
Customers want to run AI models in their own datacenter, with governance and without depending on a public API. The series follows, step by step and with the real screens, the enablement of Nutanix Enterprise AI (NAI) on NKP: what to read first, the storage it requires, installation and licensing, and how to bring CPU, GPU and agents together behind the same API.
· 11 min read · Updated on
Before clicking Enable: what Nutanix Enterprise AI is, how it is licensed (vCPU or GPU memory, with a 2-day trial), what the 2.8.1 User Guide requires, how much the control plane consumes, what changes without a GPU and what the two Nutanix Validated Designs recommend. And the lab's decision: a dedicated cluster, with the memory math done first.
· 19 min read
The RWX storage NAI requires, from scratch: Nutanix Files 5.4 deployed from Prism Central in 27 minutes, NFS v3 and v4, the NKP CSI in Prism Central mode talking to the Files v4 API, and a ReadWriteMany volume ready in 6 seconds, with two pods on different nodes writing to the same file.
· 18 min read
A dedicated NKP 2.19 cluster, the seven catalog dependencies, NAI 2.8.1 with a certificate from the internal CA, the license applied on day one, a validated model and a custom one from Hugging Face, and the first endpoint answering on CPU, with the prerequisites that make the install go smoothly and the real numbers.
· 17 min read
The same OpenAI call going to NAI on CPU and to a GB10 GPU just by changing the model, +40% tokens/s on CPU with one variable, vLLM metrics in the NKP Grafana, the Agentic Gateway with a unified endpoint, token-based rate limiting and MCP, and NAI protected with Velero and NDK, with the best practices and the real numbers.