Everythingadatahallprovides,insideonesealedcabinet.

Compute, cooling, power continuity, fire protection and monitoring are not selected independently and assembled — they are a single coupled system, and the sealed-envelope requirement drives all of them.

20U
Primary rack + 3U coupled: 8×
Blackwell GPUs, baseline: 1.5 TB
DDR5-6400 memory: 128
PCIe lanes per socket

Anatomy of the node

20U, IP55, sealed.

A 20U primary rack plus a coupled 3U power-distribution rack, with a top-mounted outdoor condenser. Every subsystem the node needs to survive alone is inside the envelope.

Top-mounted condenser+55 °C rated

Coupled 3U rack · 3U

2 × 3φ 32 A horizontal RPDU-B + 1U spare

Total 20U + 3U coupledIP55

Spare U-space

Reserved for future expansion

One rack unit held back so a field change — an extra switch, a second DPU, additional instrumentation — does not require re-cutting the enclosure.

Fire suppression

Rack-mount clean-agent detection & discharge

Detection, control and agent in a single rack-mount unit. Optical smoke and heat sensors sample the internal return-air path, where all circulating air passes.

AI compute node

8 × NVIDIA Blackwell GPUs · dual AMD EPYC

MSI CG480-S6053 chassis. Eight RTX PRO 6000 Blackwell Server Edition at 600 W each — roughly 4.8 kW of GPU power in 4U, front-to-rear airflow inside the sealed loop.

Online UPS

Double conversion · ~3:45 autonomy at load

The node is always fed from the UPS inverter, so the ATS break — tens of milliseconds — is never seen by the load. Output stays a regulated sine wave through sags, surges and frequency drift.

Precision DX cooling

Closed-loop evaporator · 7.5 kW net sensible @ 50 °C

An inverter-driven compressor with EC fans modulates to load rather than cycling, holding 18–25 °C at the equipment intake as GPU utilisation swings.

Coupled 3U rack

2 × 3φ 32 A horizontal RPDU-B + 1U spare

Power distribution is carried in its own coupled enclosure so that a distribution change never breaks the seal on the compute cabinet.

Compute platform

Eight Blackwell GPUs on an MSI CG480-S6053.

AMD EPYC selected for memory topology and PCIe lane budget at this GPU count — 128 lanes per socket feed eight accelerators, the DPU and the NVMe pool without contention.

Cutaway of the 20U sealed cabinet showing the fire suppression unit, the 4U eight-GPU MSI server, the 10 kW online UPS, the rack-mount DX cooling unit and the coupled 3U distribution rack

Coupled rack module · UAE Dubai · render for illustration only

~5.8 kW

Typical AC draw

~7.3 kW

Peak AC draw

~5.2 kW

GPUs capped to 400 W

Platform specification

GPU

8 × NVIDIA RTX PRO 6000 Blackwell Server Edition

600 W per GPU · ~4.8 kW total

CPU

2 × AMD EPYC 9575F

64 cores @ 3.3 GHz per socket

Memory

1.5 TB DDR5-6400

24 × 64 GB · one DIMM per channel

Storage

8 × 7.68 TB U.2 NVMe SED

+ 2 × 960 GB RAID-1 boot

Network board

NVIDIA BlueField-3 DPU B3220

on-board

On-board network

2 × 10GbE RJ45 LAN

I/O budget

128 PCIe lanes per socket

feeds 8 GPUs, DPU and NVMe

Security

On-board TPM · self-encrypting drives

Form factor

4U rack-mount

front-to-rear airflow

Note · 600 W per GPU — roughly 4.8 kW of GPU power in a single 4U chassis, cooled by a sealed closed-loop DX system with no air exchanged with ambient.

Configuration range

Four to sixteen GPUs, in one architecture.

GPU count is not an isolated choice — cooling capacity, UPS rating and rack height scale with it.

Peak node AC draw~7.3 kW

Cooling, net sensible @ 50 °C7.5 kW

Online UPS rating10 kW

20U + 3U coupled

Rack height

At sixteen accelerators the condenser and the incoming three-phase feed become the binding constraints rather than the compute. 8× is the proof-of-concept baseline; the 4× and 16× figures are indicative, pending design freeze.

Accelerator options

Not tied to one GPU.

The enclosure, cooling circuit and power chain are specified against watts and form factor — not a part number.

Accelerator classes and their effect on the node
Accelerator class	Board power	Effect on the node	Status
RTX PRO 6000 Blackwell SE	~600 W	Baseline — eight fit within the 7.5 kW sealed air-cooled envelope.	Baseline
H200 — PCIe / NVL class	~600 W	Comparable envelope to baseline; count and plant ratings broadly unchanged.	Within envelope
H200 — SXM class	~700 W	Reduced count at baseline cooling, or an uprated cooling and UPS package.	Plant up / count down
B200 class	~1,000 W	Eight parts alone exceed the baseline cooling duty. Count down, or plant up.	Plant up / count down
B300 / Blackwell Ultra class	~1,200–1,400 W	Beyond sealed air cooling at 50 °C in this footprint — direct liquid becomes the path.	Liquid required

The binding constraint is heat rejection, not compatibility. Above roughly one kilowatt per part the cooling envelope — not the chassis or the slot count — decides how many the node carries. Board power is indicative, confirmed against current datasheets at design freeze.

Appendix A

Specification summary.

The complete envelope in one table — enclosure, environment, thermal, power, protection and monitoring.

Enclosure

Rating: IP55, sealed — zero air exchange with ambient
Primary rack: 20U outdoor cabinet + coupled 3U distribution rack
Shell: PIR sandwich panel over mineral insulation, UV-stable C4+ coating
Mounting: Prepared slab or 600 mm plinth, top-mounted condenser

Environment

Design ambient: 50 °C continuous, excursions to 55 °C, direct sun
Internal band: 18–25 °C at equipment intake under varying utilisation
Ingress: Dust, sand, salt aerosol and humidity excluded by construction

Thermal

Type: Precision DX, closed loop, inverter-driven compressor + EC fans
Capacity: 7.5 kW net sensible at 50 °C ambient (baseline 8× config)
Condenser: High-ambient, rated +55 °C, vertical upward discharge
Fail-safe: Alarm → staged GPU capping → graceful shutdown → door release

Power

Sources: Solar PV (priority) + grid, via automatic transfer switch
Inverter: 15 kW
UPS: 10 kW rack-mount online, double conversion, ~3:45 min at load
Distribution: 400 V 3φ+N 50 Hz, TN-S, 63 A main, five outgoing ways
Aggregate design load: ~11 kW (node ~7.3 kW peak + cooling ~2.5–3.5 kW)

Protection

Detection: Optical smoke + heat on the internal return-air path
Suppression: Self-contained rack-mount clean-agent unit
Interlocks: Fan/compressor stop, damper close, node power off, door secure

Monitoring

Local: 7-inch door-mounted HMI acting as local DCIM head-end
Northbound: Modbus TCP, SNMP v3, MQTT to central DCIM or NOC
Functions: Trending, event logging, remote alerting, setpoint control

Appendix C · Component specification policy

Thermal, power and protection components are specified by function and measured performance rather than by manufacturer. Compute silicon and the server platform are named, because platform behaviour under sustained thermal load is the central subject of the proof of concept.