References archiectures are validated blueprints that tell us exactly what to buy connect and configure it to get a working AI cluster.
NVIDIA DGX BasePOD Models
The DGX BasePOD is NVIDIA’s entry-to-mid-range reference architecture so you can pick any DGX compute and it now spans four generations
- DGX H100 – Hopper gen
- DGX B200 – Blackwell
- DGX GB200 – Grace CPU + Blackwell.
- DGX Rubin – Current generation
The components of BasePOD include
- DGX Compute Options : DGX B200 – current Blackwell & DGX H100 – previous Hopper gen
- NVIDIA Networking – InfiniBand or Ethernet fabric for GPU-to-GPU communication
- Base Command Software – NVIDIA’s cluster management platform for provisioning, job scheduling, and monitoring
- Partner Storage Appliance- Storage needs to be purchased from HPe, IBM etc
- NVIDIA AI Enterprise – Full suite of enterprise AI software, frameworks, and support — sits above the hardware
DGX SuperPOD
SuperPOD is the AI factory and these can be deployed in customer data centres and cloud service providers. Basepod can be scaled out to 16 nodes and superpod can scale up to 32 nodes.
CES 2026 and detailed at GTC 2026 Announcements
Rubin-generation SuperPOD is a fundamentally different beast from earlier versions and comes in two distinct configurations depending on whether you’re going all-in on NVIDIA’s rack-scale NVL72 or staying in the x86 world with NVL8 nodes.
Configuration 1 — SuperPOD with DGX Vera Rubin NVL72
Configuration 2 — SuperPOD with DGX Rubin NVL8 (x86)
NVIDIA Enterprise RA
Instead of using DGX hardware, customers that want NVIDIA’s software stack can also use the RA to you turn a consumer server from Dell, HPE, Lenovo, or Supermicro into an AI powerhouse and these will all be NVIDIA-Certified Systems as well.
