AI & Data Blueprint
AI racks supplied, powered and tested.
AI & Data Blueprint
AI racks supplied, powered and tested.
What to ask every vendor.
What to specify for an AI compute block on NVIDIA DGX B300, how it recovers from each failure, and how to test it.
16
DGX B300 systems per block
4
DGX B300 per rack
> 50 kW
Per compute rack
800 Gbps
XDR InfiniBand compute fabric
30 kW
Per rack, the hall's rating
≤ 220 kW
Per rack on the Rubin path
All three documents, in one place.
Ask AI about this blueprint
All three documents, in one place.
Ask AI about this blueprint
Who it is for.
For the head of research computing or the CTO buying an on-premises training or inference block in 2026, who has a data hall with a power number and needs the block to arrive built, burned in and cabled to a plan, with a path to the platform NVIDIA ships next. Sized for 16 DGX B300 systems, inference plus fine-tuning, a hall rated at 30 kW per rack on air.
What is in, what is out.
In: the DGX B300 block, its compute and storage fabrics, management nodes, racks, power and cooling plan, staging and burn-in, acceptance, and the Rubin NVL8 upgrade path. Out: models and data pipelines, the enterprise network and the second scalable unit.
Tested at the first build.
Five principal architects write the designs. Each name is on the document.
DGX B300
for 2026, Rubin NVL8 as the path
DC busbar
for rack power
Rear-door heat exchangers
on compute racks
Quantum-X800 InfiniBand
compute fabric
SN5610 pair
for the storage fabric
Full SU cabled now,
Rubin in a new hall
Where it applies.
Questions about this design.
Talk to an expertNVIDIA DGX B300 systems, Quantum-X800 InfiniBand, Spectrum-4 SN5610 storage switches, and NVIDIA Mission Control with UFM 3.5.
Models and data pipelines, the enterprise network (Networking Blueprint) and the second scalable unit.
Diksha Shah, Principal AI Infrastructure Architect.
No. Symbeeo's industry research of September 2026 links no industry rule to this design's requirements, so none is named.