IREN is building a large GB300 AI cluster at its Childress, Texas campus. One Horizon is three data halls; this is Horizon 1 = halls 011, 012, 013.
Top-down view straight from the fabric's rack addressing: each hall is 11 aisles (rows) × 8 GPU positions. Each tile is a GB300 NVL72 rack (72 GPUs), sitting at positions 05–12 with network/switch racks at the aisle ends. Each pod (SU) spans 2 adjacent aisles (half-pods, 1). Colour = cabling-validation status.
Built entirely from publicly observable network data — no private access. Rows in halls 012/013 carry a documented +2 label offset (in the source), normalised here to the clean hall-011 pattern.
Pods = SU1…SU18. Racks = RCK-<hall>-<row>-<pos>
— e.g. RCK-011-04-05 = hall 011, row 4, position 5. Hover any GPU rack for its ID + switches.
Switches: SU3-EW-SP-1-1 → pod SU3, EW = East-West InfiniBand GPU fabric
(NS = North-South Ethernet); SP = spine, LF = leaf, CN = core-node; then group + index.
OOB = out-of-band management (OOB-GPU-LF = 2 per GPU rack).
It's a top-down view of each hall — like looking down at the floor. Every hall is 11 aisles of racks; one aisle looks like this:
A pod = 2 aisles (16 racks / 1,152 GPUs); three "half" pods are 1 aisle. Green = wiring checked · amber = still being wired · a red dot = a rack needing a fix.