Step 3 · Build plan · Enterprise
HGX B300 build plan
8 × Blackwell Ultra in one server. Every parameter below links to an NVIDIA, OEM or government source; what we could not confirm is listed at the end.
Server configuration (8-GPU node)
⚠ NVIDIA’s HGX page says “2.1 TB” vs 2.30TB in the reference architecture (likely TiB vs TB) – to be verified
Systems you can buy (official pages)
| Vendor | Model | Form / cooling | CPU | Source |
|---|---|---|---|---|
| NVIDIA | DGX B300 | 10U air (busbar version available) | 2 × Xeon 6776P | nvidia.com |
| Supermicro | SYS-822GS-NB3RT | 8U air | 2 × Xeon 6700P | supermicro.com |
| Supermicro | SYS-422GS-NB3RT-LCC | 4U DLC (rack-level delivery) | 2 × Xeon 6700P | supermicro.com |
| Supermicro | AS-4126GS-NB3RT-LCC / AS-8126GS-NB3RT | 4U liquid / 8U air | 2 × AMD EPYC 9005 | supermicro.com |
| Dell | PowerEdge XE9780 | 10U air⚠ Dell lists “B300 NVL8 270GB 1100W”, vs 288GB from NVIDIA – to be verified | 2 × Xeon 6 | dell.com |
| Dell | PowerEdge XE9785L | Liquid-cooled, ORv3 rack | 2 × AMD EPYC 9005 | dell.com |
| HPE | Compute XD690 | 10U air⚠ QuickSpecs warns of 128GB DIMM supply constraints | 2 × Xeon 6 | hpe.com |
| HPE | ProLiant Compute XD685 | 5U DLC / 6U air | 2 × AMD EPYC | buy.hpe.com |
| Lenovo | ThinkSystem SR680a V4 | 8U air | 2 × Xeon 6 | lenovopress.lenovo.com |
| 浪潮 / IEIT Systems | To be verified | No HGX B300 model found on the official site (NF5688G8 is Hopper) | — | ieisystems.com |
Cisco also resells the Supermicro 8U system; configurations and availability vary by region.
From one node to a cluster
- 1. Single node (8 GPUs)
8 GPUs over NVLink. For pure inference the RA says a compute fabric may be skipped – but adding one later for training means downtime. North-south via BlueField-3 to storage and the business networkdocs.nvidia.com
- 2. Small cluster (2–16 nodes)
Each node’s 8 × 800G ports go rail-aligned to leaf switches (GPU n → rail n). At small sizes the RA merges compute and converged fabrics to save cost. Switches: Spectrum-X Ethernet or Quantum-X800 InfiniBanddocs.nvidia.comnvidia.comnvidia.com
- 3. Enterprise RA (32 / 64 / 128 nodes)
Spectrum-X Ethernet, dual-plane fat tree: compute core SN5600 × 12 / 24 / 48, converged SN5600 × 12 / 24 / 48, OOB SN2201 × 4 / 8 / 16; a cheaper single plane is also supporteddocs.nvidia.comdocs.nvidia.com
- 4. DGX SuperPOD (72 nodes per SU)
Quantum-X800 (Q3400-RA, 800Gb/s) InfiniBand, rail-optimised, non-blocking, twin-plane fat tree; 1 SU = 72 nodes / 576 GPUs, reference 8 SU = 576 nodes / 4,608 GPUs, scalable to 72 SU and 2,000+ nodes. Plus storage (IB NDR or Spectrum SN5610 Ethernet), in-band and OOB fabrics – 4 networks in total; ≥2 × 100GbE customer uplinksdocs.nvidia.comdocs.nvidia.com
How big a model fits
- NVIDIA’s reference architecture: each B300 supports a model of roughly 120B parameters at most; larger models need model parallelism, which still fits inside one 8-GPU node.docs.nvidia.com
- Our formula estimate (weights only, no KV cache): 2,304GB per node holds ~1T parameters in FP8 with 1TB+ left for KV cache; FP4 halves the weights.
- Real model size and concurrency depend on the engine, quantisation format and context length – benchmark with vLLM / SGLang tooling before going live.
Price
Quote requiredQuote required. NVIDIA, Supermicro, Dell, HPE and Lenovo publish no HGX B300 system prices; we do not quote channel rumours. The “$0.24 per million DeepSeek-R1 tokens” figure (SemiAnalysis) on the DGX B300 page is a vendor claim.nvidia.com
Delivery & compliance
- Export controls: the US BIS rule of 2026-01-15 moves only chips with TPP < 21,000 and total DRAM bandwidth < 6,500GB/s (e.g. H200, MI325X) to case-by-case review for China and Macau; more capable products stay under a presumption of denial. B300 offers up to 8TB/s per GPU, above that threshold – exports to China / Macau need a licence and are likely to be denied. Rely on BIS and your supplier’s compliance classification.
- Entities headquartered in Country Group D:5 or Macau (including overseas subsidiaries) need a BIS licence for advanced computing items anywhere in the world (BIS guidance, 2026-05-31).
- Delivery: DGX B300 ships without cables (user guide); liquid-cooled models need rack-level delivery and on-site service; HPE warns of 128GB DIMM shortages; confirm lead times with the vendor.
- Facility: ≈14.5–15kW per node; the SuperPOD reference puts 4 nodes (≈56kW) in a rack – plan high-density power (12 × C19/C20 per node or a 54VDC busbar) and floor loading (120–170kg per node).
- Offering public generative-AI services in China also requires regulatory filing – see the “Use AI compliantly in China” scenario.
Sources:federalregister.govbis.govdocs.nvidia.comsupermicro.comhpe.comdocs.nvidia.com · Use AI compliantly in China
This is a summary of public documents, not legal advice.
To be verified
- Total GPU memory: NVIDIA’s HGX page says 2.1 TB, the reference architecture and DGX user guide say 2.3TB (2,304GB) – probably TiB vs TB.
- Dell XE9780 configuration lists “B300 NVL8 270GB 1100W” per GPU, vs 288GB from NVIDIA.
- Per-GPU power: only Dell’s configuration (1,100 W) found; NVIDIA does not list a per-GPU TDP on the pages we checked.
- Inspur / IEIT Systems: no HGX B300 model on the official site.
- Prices: no official public price for any HGX B300 system.
- Lead times: vendor-specific, not published.