# UpStation design UPS-DSN-20261103-G9YQ — Budgetary design and estimate — not a quote

Prepared 2026-11-03 for **Example AI Lab** · Head of ML Infrastructure · prices as of 2026-11-03

## 4× HGX B300 training cluster (32 GPUs) with NVMe storage and a services cluster, Montana colo

4× 8x NVIDIA B300 (HGX B300 NVL8) (32 GPUs, HGX class), 5× 2U all-NVMe storage node storage (208.3 TB plannable, replica_3), 3 cluster services nodes. Networks: gpu fabric 800G on 1 switch; front end 400G on 2 switches; oob 1G on 1 switch. 3 racks at ≤ 40 kW, 82.51 kW provisioned, in Montana colo. Every figure is computed from the ai.upstation.io catalog with the arithmetic shown below.

| Hardware (CapEx) | Colocation | Year one | 3-year total | 5-year total |
|---|---|---|---|---|
| $3,307,252.84 | $21,453.00/mo (82.51 kW) | $3,564,688.84 | $4,079,560.84 | $4,594,432.84 |

**Delivery:** 20 weeks, set by 8x NVIDIA B300 (HGX B300 NVL8), Intel, 3TB (Aivres) — Aivres KR9288X3. **Colo:** Montana at $260/kW/month × 82.51 kW, 36-month term.

## Requirements

| Requirement | Value |
|---|---|
| Workload | training |
| GPUs | 32 |
| Storage | 200 TB, NVME |
| Rack limit | 40 kW |
| Colo | Montana |
| Timeline | production in Q2 2027 |

## How it was designed

- GPUs → nodes: ceil(32 ÷ 8) = 4 × 8x NVIDIA B300 (HGX B300 NVL8), Intel, 3TB (Aivres) (32 GPUs; nodes ship fully populated).
- Class: HGX — the GPUs must act as one (training, multi-node). Does the model at its precision fit one GPU's memory, and can each GPU work alone? Yes: PCIe. No: HGX. Distributed training is never served by PCIe servers.
- Storage (replica_3, Ceph across nodes, floor 3+1): plannable = 184 × N × (N−1)/N × 0.333 × 0.85; N = 5 gives 208.3 TB ≥ 204.5 TB and 5 × 20 = 100 GB/s ≥ 16.0 GB/s read.
- Storage demand: read = 32 × 0.5 = 16.0 GB/s; checkpoint burst = 18.7 GB/s; write = 18.7 GB/s; capacity = 204.5 TB.
- Cluster services (chosen with the buyer, optional): 3 × 1U CPU nodes for provisioning, scheduler and monitoring.
- GPU fabric: 32 × 800G ports (1 per GPU) ≤ 32 → one 32× 800G OSFP switch, 1:1 non-blocking (single point of failure; a leaf-spine removes it).
- Front-end: 2 ports per host, one to each switch of an MLAG pair of 32× 400G QSFP-DD switches; 7 ports per switch of 30 usable → 1 pair(s). Storage stays converged on this plane (well under 40 GB/s per node).
- OOB: 12 hosts + 3 switches = 15 × 1G ports → 1 × 48× 1G RJ45 + 4× 10G.
- GPU nodes per 40 kW rack = min(floor(40 U ÷ 10 U), floor(40 ÷ 16.96 kW)) = 2 (power binds).
- Compute 1: 22 U, 35.24 kW provisioned → 1 × 415V 3-ph 60A circuit pair (A+B), 82% of each feed; 8 power cords per feed.
- Compute 2: 20 U, 33.92 kW provisioned → 1 × 415V 3-ph 60A circuit pair (A+B), 79% of each feed; 6 power cords per feed.
- Services 1: 15 U, 13.35 kW provisioned → 1 × 208V 3-ph 50A circuit pair (A+B), 74% of each feed; 10 power cords per feed.

## Network topology

```mermaid
flowchart TB
  GF["GPU fabric: 1× 32× 800G OSFP<br/>800G 1:1, 32 links"]
  GN["4× 8x NVIDIA B300 (HGX B300 NVL8)"] -- "32× 800G" --> GF
  FE["Front-end MLAG pair: 2× 32× 400G QSFP-DD<br/>400G, storage converged"]
  GN -- "2× 400G per node" --> FE
  SN["5× 2U all-NVMe storage node"] -- "2× 100G per node" --> FE
  MG["3× cluster services node"] -- "2 ports per node" --> FE
  FE -. "uplinks to your core (not priced)" .-> CORE(("Your network"))
  OOB["OOB: 1× 1G switch"] -. "BMC, 15 ports" .- GN
```

## Racks and power

(rack elevation: the SVG drawing in the HTML/PDF version of this document)

| Rack | Space | Provisioned | Circuits | Devices |
|---|---|---|---|---|
| Compute 1 | 22/40 U | 35.24 kW | 1× 415V 3-ph 60A (A+B) | 2× 8x NVIDIA B300 (HGX B300 NVL8), 1× 32× 800G OSFP switch — GPU fabric, 1× 48× 1G RJ45 + 4× 10G switch — OOB |
| Compute 2 | 20/40 U | 33.92 kW | 1× 415V 3-ph 60A (A+B) | 2× 8x NVIDIA B300 (HGX B300 NVL8) |
| Services 1 | 15/40 U | 13.35 kW | 1× 208V 3-ph 50A (A+B) | 2× 32× 400G QSFP-DD switch — front-end pair, MLAG, 5× 2U all-NVMe storage node, 3× 1U compute node |

## Bill of materials

| Section | Item | Qty | Unit (USD) | Extended | Lead time | Source |
|---|---|---|---|---|---|---|
| GPU servers | 8x NVIDIA B300 (HGX B300 NVL8), Intel, 3TB (Aivres) — Aivres KR9288X3 | 4 | $537,196.40 | $2,148,785.60 | Configure-to-order on Aivres B300 allocation (October 2026 s | https://ai.upstation.io/p/UPS-GPU-B300-8-T2-AIV |
| Storage | 2U all-NVMe storage node, dual socket, 24x 7.68TB NVMe (184TB raw) — SYS-222C-TN | 5 | $177,958.91 | $889,794.55 | Barebone in stock (2026-11-03); configure-to-order. Est. 3-5 | https://ai.upstation.io/p/UPS-SN-INT-NVME |
| Cluster services | 1U compute node, dual socket, 64 cores, 512GB, diskless NVMe (virtualisation host) — SYS-122C-TN | 3 | $34,523.31 | $103,569.93 | Barebone in stock (2026-11-03); configure-to-order. Est. 3-5 | https://ai.upstation.io/p/UPS-CN-INT-1U |
| GPU fabric | 32× 800G OSFP switch (GPU fabric) — FS N8650-32OD, PicOS | 1 | $37,500.00 | $37,500.00 | 26 in stock (global warehouse, 2026-11-03). ~1 week | https://ai.upstation.io/p/UPS-NET-800G-32 |
| GPU fabric cabling | 800G link: 2× SR8 + 2× MTP-12 OM4 10 m — node to leaf, 1 per GPU port (estimated optical) | 32 | $2,058.00 | $65,856.00 |  | https://ai.upstation.io/switches |
| Front-end network | 32× 400G QSFP-DD switch (front-end pair, MLAG) — FS N8610-32D, PicOS | 2 | $20,000.00 | $40,000.00 | 33 in stock (global warehouse, 2026-11-03). ~1 week | https://ai.upstation.io/p/UPS-NET-400G-32 |
| Front-end network | +2× 400G north-south (budgeted at the Supermicro equivalent option; Aivres fits its own 400G NICs and confirms them on its quote) (listed option, one per GPU node) | 4 | $2,693.94 | $10,775.76 |  | https://ai.upstation.io/p/UPS-GPU-B300-8-T2-AIV |
| Front-end network cabling | 400G AOC 10 m — GPU nodes to the pair, 2 per node | 8 | $1,055.00 | $8,440.00 |  | https://ai.upstation.io/switches |
| Front-end network cabling | 400G→4×100G breakout 2 m — storage nodes to the pair (4 nodes per cable) | 4 | $199.00 | $796.00 |  | https://ai.upstation.io/switches |
| Front-end network cabling | 400G→4×100G breakout 2 m — cluster services nodes to the pair (4 nodes per cable) | 2 | $199.00 | $398.00 |  | https://ai.upstation.io/switches |
| Front-end network cabling | 400G 2 m passive DAC — peer link | 2 | $174.00 | $348.00 |  | https://ai.upstation.io/switches |
| Out-of-band | 48× 1G RJ45 + 4× 10G switch (OOB) — FS S3410-48TS, PicOS | 1 | $989.00 | $989.00 | 593 in stock (global warehouse, 2026-11-03). ~1 week | https://ai.upstation.io/p/UPS-NET-1G-48 |
| | **Hardware total** | | | **$3,307,252.84** | | |

## Assumptions

- GPU server: 8x NVIDIA B300 (HGX B300 NVL8), Intel, 3TB (Aivres) (UPS-GPU-B300-8-T2-AIV), the default HGX line (proposed)
- Second source: 8x NVIDIA B300 (HGX B300 NVL8), Intel (UPS-GPU-B300-8), same price and design, 30+ weeks on NVIDIA allocation. Next generation: NVIDIA Rubin (HGX Rubin NVL8, liquid-cooled racks) follows B300, not priced

## Not included

- Racks, PDUs, whips and cabling labour
- Uplinks to the buyer's core network
- Software (cluster manager, storage software), unless listed
- Taxes, shipping and duties
- Installation and on-site services

**Next step.** Book the intro (https://calendly.com/alecglover/intro-call) or email alec@upstation.io with the design ID; UpStation validates the design and introduces the supplier that fits.

Prepared by UpStation from public supplier prices at ai.upstation.io. Budgetary only; the supplier confirms price, availability and lead time on order. UpStation does not resell. alec@upstation.io
