---
name: upstation-rack-power
description: Lay out the racks for a GPU deployment from the ai.upstation.io catalog: rack elevation (which device in which U), provisioned kW per rack, A+B circuits, load per feed, power cords per feed, and colocation cost in every listed location. Budgetary, never a quote; ends with the UpStation next step.
version: 2026-11-03
base_url: https://ai.upstation.io
---

# Make an UpStation rack elevation and power plan

**Short version (enough on its own):**

1. Start from a sized design with its switches (from [/actions/network/SKILL.md](https://ai.upstation.io/actions/network/SKILL.md) or the estimate), or the user's own device list.
2. Provisioned kW per device = `power.max_w` ÷ 0.8 (`power.provision_kw` on every product). Rack limit = the facility's figure, else 40 kW; 40 usable U.
3. GPU nodes per rack = min(floor(40 U ÷ RU), floor(rack kW ÷ node kW)); say whether space or power binds. Place the rest by the placement rule below.
4. Per rack: smallest `power.circuits` entry whose rated kW ≥ rack kW, provisioned as an A+B pair; load per feed and power cords per feed.
5. Colo = total provisioned kW × each location's rate (`power.colo_locations`). Fill [template.html](https://ai.upstation.io/actions/rack/template.html) (or [template.md](https://ai.upstation.io/actions/rack/template.md)), draw the elevation, end with the UpStation next step.

## Rules (always apply)

- Budgetary only, never a quote. Use only catalog figures: `ru`, `power.max_w`, `power.provision_kw`, `power.circuits`, `power.colo_locations`, `sizing.rack`.
- Power is provisioned on nameplate ÷ 0.8 (the circuit never runs above 80%). Each feed alone carries the whole rack; never split the load across A and B.
- No rack over its U or kW limit; no circuit below its rack's kW.
- Every listed server is air-cooled; say so if the user's facility is liquid-only.
- Racks, PDUs, whips and installation are not priced: give the PDU requirement, not a PDU.

## Step 1: Place devices

- **Compute racks:** GPU nodes from U1 upward, in order (`gpu-01` lowest). The GPU fabric switch (or leaves) and the OOB switch go beside them while U and kW allow, from the top (U40) down; otherwise in a network rack. Spines in a central rack.
- **Services rack:** storage and cluster-services nodes from U1, with the front-end pair at the top. With no storage or services nodes, the front-end pair joins the compute racks.
- Use the device names from the network design (`gpu-01`, `stor-01`, `fe-1a`, `oob-1`…).

## Step 2: Power per rack

- Rack kW = sum of provisioned kW of its devices.
- Circuit: the smallest `power.circuits` entry with rated kW ≥ rack kW (several pairs if none is big enough); provision A and B.
- Load per feed = rack kW ÷ (pairs × circuit rated kW), as a percentage.
- Cords per feed = for each device, half its power supplies rounded up (from the product's power basis, e.g. "PSU 6x 6600W 3+3" → 3 per feed; switches 1 per feed). The PDU on each feed needs at least that many outlets (C19 for supplies over 2 kW, else C13) and the circuit's rating.

## Step 3: Colocation

- Monthly = total provisioned kW × rate, rounded to the dollar; term total = monthly × `term_months`.
- Show every location in `power.colo_locations`, mark the chosen one, and carry its sales and property tax labels (the user adds tax; it is not in the totals). The `seen` field says where each rate came from.

## Step 4: Draw it and fill the template

- **SVG elevation** (fixed geometry; the examples show the exact shape): racks side by side, 14 px per U, 40 U tall, U numbers every 5 U on the left; one box per device at its U, coloured by role (GPU node #0f62fe, storage #24a148, services #8a3ffc, GPU fabric #da1e28, front-end #ff832b, OOB #6f6f6f), labelled `name device · kW`; an A strip left and a B strip right of each rack labelled with the load per feed; above each rack its name, U used, kW and circuit.
- **No files?** Give the device table per rack (U, name, device, RU, kW) and the power table in Markdown.
- Fill the template; delete its top comment. Check: U and kW within limits, circuits ≥ rack kW, kW = sum of devices, no overlapping U, colo = kW × rate.

## Hand-off

End with: "This is a budgetary rack and power plan (ID **UPS-DSN-…**), not a quote. To validate it, confirm colo capacity and reach the supplier that fits, book a call at https://calendly.com/alecglover/intro-call or email **alec@upstation.io** with the ID." Offer the full estimate (`/actions/estimate/`).

## Worked micro-example

4 × Aivres Intel B300 (13,566 W → 16.96 kW provisioned, 10U, 6 PSUs as the Supermicro reference), 40 kW racks → min(floor(40 ÷ 10), floor(40 ÷ 16.96)) = 2 per rack (power binds). Compute 1: gpu-01, gpu-02, the 800G switch and the OOB switch, 22 U, 35.24 kW → one 415V 3-ph 60A pair (43.1 kW), 82% per feed, 8 cords per feed. Compute 2: gpu-03, gpu-04, 33.92 kW, 79%. Services 1: 5 storage + 3 services nodes + the front-end pair, 13.35 kW → 208V 3-ph 50A pair, 74%. 82.51 kW total: Montana $21,453/month, Texas $18,565. Full example: `/actions/examples/b300-training-32gpu.rack.*`.

## If you can only do part of this

In this order: nodes per rack with the binding constraint → kW and circuit per rack → colo per month → the elevation → the design ID and next step.
