UpStation
Design UPS-DSN-20261103-BKCU
Prepared 2026-11-03 · prices as of 2026-11-03
Prepared for Example SaaS Co · —
Budgetary design and estimate — not a quote

4× 8-GPU RTX PRO 6000 inference farm (32 GPUs), Texas colo

4× 8x NVIDIA RTX PRO 6000 Blackwell Server Edition (32 GPUs, PCIe class). Networks: front end 25G on 2 switches; oob 1G on 1 switch. 1 rack at ≤ 40 kW, 34.33 kW provisioned, in Texas colo. Every figure is computed from the ai.upstation.io catalog with the arithmetic shown below.

$724,604.56hardware (CapEx)
$7,724.00/mocolocation, 34.33 kW
$817,292.56year one
$1,002,668.563-year total
$1,188,044.565-year total

Delivery: 5 weeks, set by 8x NVIDIA RTX PRO 6000 Blackwell Server Edition, Intel — SYS-422GA-NRT. Colo: Texas at $225/kW/month × 34.33 kW, 36-month term.

Requirements

Workload
inference
GPUs
32
Rack limit
40 kW
Colo
Texas
Timeline
within 6 weeks

How it was designed

  1. GPUs → nodes: ceil(32 ÷ 8) = 4 × 8x NVIDIA RTX PRO 6000 Blackwell Server Edition, Intel (32 GPUs; nodes ship fully populated).
  2. Class: PCIe — each GPU works alone (inference). 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.
  3. Front-end: 2 ports per host, one to each switch of an MLAG pair of 48× 25G + 8× 100G switches; 4 ports per switch of 48 usable → 1 pair(s).
  4. OOB: 4 hosts + 2 switches = 6 × 1G ports → 1 × 48× 1G RJ45 + 4× 10G.
  5. GPU nodes per 40 kW rack = min(floor(40 U ÷ 4 U), floor(40 ÷ 8.47 kW)) = 4 (power binds).
  6. Compute 1: 19 U, 34.33 kW provisioned → 1 × 415V 3-ph 60A circuit pair (A+B), 80% of each feed; 19 power cords per feed.

Network topology

flowchart TB
  FE["Front-end MLAG pair: 2× 48× 25G + 8× 100G
25G, storage converged"] GN -- "2× 25G per node" --> FE FE -. "uplinks to your core (not priced)" .-> CORE(("Your network")) OOB["OOB: 1× 1G switch"] -. "BMC, 6 ports" .- GN

Rack elevation and power

Compute 119/40 U · 34.33 kW · 1× 415V 3-ph 60A A+B A 80% B 80% 1 6 11 16 21 26 31 36 gpu-01 8x NVIDIA RTX PRO 6000…8.47 kW gpu-02 8x NVIDIA RTX PRO 6000…8.47 kW gpu-03 8x NVIDIA RTX PRO 6000…8.47 kW gpu-04 8x NVIDIA RTX PRO 6000…8.47 kW fe-1a 48× 25G + 8× 100G switc…0.19 kW fe-1b 48× 25G + 8× 100G switc…0.19 kW oob-1 48× 1G RJ45 + 4× 10G sw…0.07 kW gpu node storage node services node gpu fabric front end oob
RackSpaceProvisionedCircuitsDevices
Compute 119/40 U34.33 kW1× 415V 3-ph 60A (A+B)4× 8x NVIDIA RTX PRO 6000 Blackwell Server Edition, 2× 48× 25G + 8× 100G switch — front-end pair, MLAG, 1× 48× 1G RJ45 + 4× 10G switch — OOB

Bill of materials

ItemQtyUnit (USD)ExtendedLead timeSource
GPU servers
8x NVIDIA RTX PRO 6000 Blackwell Server Edition, Intel — SYS-422GA-NRT4177,430.31709,721.24In stock (barebone and RTX PRO 6000 cards, 2026-11-03); conf/p/UPS-GPU-RTXPRO-8
Front-end network
48× 25G + 8× 100G switch (front-end pair, MLAG) — FS S5470-48Y, PicOS26,049.0012,098.0097 in stock (global warehouse, 2026-11-03). ~1 week/p/UPS-NET-25G-48
+1x 25G 2x SFP28 CX6 NIC (AOC-631102AN-ADAT) (listed option, one per GPU node)4344.081,376.32/p/UPS-GPU-RTXPRO-8
Front-end network cabling
25G 2 m passive DAC — GPU nodes to the pair, 2 per node839.00312.00/switches
100G 2 m passive DAC — peer link254.00108.00/switches
Out-of-band
48× 1G RJ45 + 4× 10G switch (OOB) — FS S3410-48TS, PicOS1989.00989.00593 in stock (global warehouse, 2026-11-03). ~1 week/p/UPS-NET-1G-48
Hardware total$724,604.56

Assumptions

Not included