Built, burned in, cabled.¹
For the head of IT told to bring AI in-house: we design, supply, integrate and support NVIDIA DGX, Dell or HPE racks, tested before they ship, with a path to NVIDIA's next platform.²
Power and cooling first,
fabric second, compute third.
800 Gb/s
compute fabric⁴
64
systems per scalable unit⁵
8
GPUs per system⁶
Will it fit our hall's power?
- Rear-door cooling. Four B300 systems per rack exceed what 30 kW of air can cool.⁷
- Power to the rack. A DC busbar where the site allows; AC PDUs where the survey finds no busbar path.⁸
- Michigan State University, 2025. Cooling failed July 9; a temporary chiller restored it by July 21.¹⁴
- We write into your requirement: what shuts down first, how long racks can run, who brings replacement cooling, in how many days.
- We check spare-part and temporary-cooling lead times for your hall, from a named supplier.
One block, eight racks.¹
Four compute, two fabric, one storage and one management rack.¹ Every position is wired on day one.⁵
Where do your prompts and answers stay?
A US court ordered OpenAI on May 13, 2025, to preserve output logs it would otherwise delete; that ended September 26, 2025, with exceptions.¹⁵ Your requirement should state where prompts and answers stay, who can reach them and for how long. Hosted or on your own racks, we check each option against that line.
Cut a power feed. The rack runs on.⁹
Led by Diksha Shah
Diksha Shah, Principal AI Infrastructure Architect · NVIDIA Certified Professional: AI Operations · NVIDIA Certified Associate: AI Infrastructure and Operations · Google Cloud Professional Cloud Architect · AWS Certified Data Engineer · credential IDs on request
~14 kW
per system⁶
Up to 800 Gb/s
per port, eight per system⁶
3 years
NVIDIA hardware and software support⁶
2 planes
in the compute fabric⁴
Every compute rack, tested before it ships.¹⁰
- Built.
- Cabled.
- Burned in, for the hours set at your site survey.
- Powered on.
- GPU diagnostics on every node.
- NCCL test inside the rack.
- Shipped assembled.
Design studies
Pharma research cluster: 32 GPUs, 2 PB of flash, InfiniBand fabric
Why DGX B300 now?
It fits an air-cooled hall with rear doors.¹¹
- HPE XD685. Confirm its cooling with HPE for an air-cooled hall.¹²
- AMD. The fabric and management tools the design names are NVIDIA's.¹²
- Dell XE9780. The choice where you standardize on Dell servers.¹²
- Rubin NVL8. Up to 220 kW per rack, so a new hall.¹¹
Questions about this design? You get a reply the same day.
Send what you need to buy and by when. You get a reply the same day, then a quote.
