MODULE 04 · COMPUTE ARCHITECTURE · GB300 NVL72

2,592 racks. 186,624 GPUs. One machine.

The Alyth fleet is built from a single repeating element — the NVIDIA GB300 NVL72 rack — composed into 576-GPU scalable units, 384-rack halls and a 405 MW campus, exactly as the DSX reference design composes an AI factory. Platform figures below are NVIDIA datasheet values carried in the DSX Blueprint.
NVL72 Racks
2,592
@ 135 kW each
Blackwell Ultra GPUs
186,624
72 per NVLink domain
Grace CPUs
93,312
36 per rack
Scalable Units
324
8 racks · 576 GPUs each
FP4 Compute
3.73ZF
sparse · 1.44 EF per rack
Fast Memory
95.9PB
37 TB per rack

Rack → Scalable Unit → Hall

GB300 NVL72 RACK 72 GPU · 36 Grace · 135 kW · liquid POWER SHELVES · 800 VDC-READY BUSBAR PSU / E-CAP RIDE-THROUGH COOLANT SUPPLY ~32 °C COOLANT RETURN ~50 °C 18 compute trays (4 GPU + 2 Grace) · 9 NVLink switch trays
SCALABLE UNIT — 8 RACKS 576 GPUs · 1.08 MW · one leaf pair LEAF SWITCH PAIR — 800G Quantum-X800 InfiniBand / Spectrum-X800 Ethernet rail-optimised: every GPU one hop from its rail leaf
60 MW HALL — 384 RACKS 48 SU · 12 rows × 32 racks · 27,648 GPUs CDU gallery on the cooling-plant wall · power rooms both ends hot-aisle contained · residual air via rear-door HX

Interactive Capacity Model

The whole campus derives from three numbers. Move them and watch the fleet re-size — the same what-if loop the DSX configurator runs against its digital twin.

Baseline: 60 MW hall = 384 racks (51.84 MW) + 4.5 MW fabric + 2.5 MW storage + 1.2 MW management.

NVL72 racks
2,596
GPUs
186,912
Grace CPUs
93,456
Scalable units
324
FP4 sparse
3.74 ZF
Heat to liquid loop
326 MW
Annual energy @PUE 1.12
3.97 TWh
≈ Seagreen output share
~79%

Hall-by-Hall Build-Out

HallPhaseIT MWNVL72 RacksSUsGPUsGrace CPUsFP4 (EF)Compute / Overhead MW

Network Fabric

SCALE-UP — INSIDE THE RACK NVLink 5 domain — 72 GPUs 130 TB/s all-to-all 9 NVSwitch trays · copper backplane 1.8 TB/s per GPU, non-blocking every rack behaves as one coherent accelerator — the unit the fabric below stitches together RAIL DESIGN 8 rails per scalable unit — every GPU reaches its rail leaf in one hop ConnectX-8 SuperNICs · GPUDirect RDMA adaptive routing + SHARP in-network compute SCALE-OUT — ACROSS THE HALL & CAMPUS Quantum-X800 InfiniBand / Spectrum-X800 Ethernet · 800G everywhere

Compute fabric

NVIDIA Quantum-X800 InfiniBand (or Spectrum-X800 Ethernet) in a rail-optimised fat tree: 8 rails per SU, leaf per rail, non-blocking to the hall spine, 2:1 tapered hall-to-campus core. 800G everywhere; ConnectX-8 SuperNICs per GPU pair.

Storage & front-end

Separate 400G Ethernet fabric for storage (≈2.5 MW per hall of NVMe-over-fabrics flash), front-end, and management. BlueField-3 DPUs isolate tenants and offload storage/telemetry from the training fabric.

Campus interconnect

Diverse dark-fibre pairs between hall meet-me rooms and dual campus POPs; hall-to-hall latency <10 µs enables multi-hall training jobs spanning the full 186k-GPU fleet.

Platform Specifications (DSX Blueprint data)

NVIDIA GB300 NVL72 — selected platform

NVIDIA GB200 NVL72 — comparison

Why GB300. ~2× FP4 inference and 1.5× HBM capacity per rack versus GB200 at only ~12% more power — fewer racks, fewer network ports and less floor for the same 405 MW envelope. Halls are provisioned to re-tray to Rubin-generation NVL systems without electrical or cooling rework.
● FIU Portfolio