Anatomy of the Rack

Twenty-seven trays on a regular pitch, a spine of copper at the back, a bar of copper down one side and water through all of it. Drag the model, pull the explode slider, and click anything that looks interesting.

Reading the stack

Look at the model from the front and it is a stack of thin drawers. Most of them — eighteen — are compute trays, and those hold the chips that do the actual work. Nine more, sitting in the middle of the stack, are switch trays: they do no computing at all, they just connect the other trays to each other.

The rest of the rack exists to keep those twenty-seven drawers fed and cool. There is a bar carrying electricity down the side, pipes carrying water up the back, and a thick bundle of cables joining every compute tray to every switch tray.

The 18 × 1UCompute trays Split 8 and 10 around the switch band rather than evenly — Supermicro’s rack diagram lists ten above the NVLink switches and eight below. Supermicro · ServeTheHome compute trays and 9NVSwitch trays ServeTheHome switch trays are both 1U, on the same pitch, in the same rails. The ordering is what carries information: the switch trays sit in the middle of the stack, with compute trays above and below them.

That placement is not aesthetic. Every compute tray has to reach every switch tray over copper, and copper's reach at NVLink signalling rates is short and expensive. Putting the switches in the middle halves the worst-case cable run. The whole vertical arrangement of the rack is a cable-length optimisation.^{ServeTheHome,NVIDIA Technical Blog}

Group sizes above and below the switch band are not symmetric. Supermicro's rack diagram lists ten compute trays above the NVLink switches and eight below, with the eight power shelves in two banks of four rather than interleaved between tray groups, and the CDU at the base.^Supermicro The asymmetry follows from what the CDU and the shelf banks occupy at the ends of the rack, not from anything electrical about the fabric.

Service implications are worth stating plainly. Every tray is front-serviceable and blind-mates both its data and its liquid connections, so the mean time to replace a compute tray is a matter of minutes and does not involve breaking the coolant loop. The failure domain, however, is not the tray: losing one tray removes four GPUs from a 72-GPU NVLink domain, and the collective operations of any job spanning that domain must be reformed. The RAS engine exists to make that a scheduled event rather than a surprise.

Every part, in text

The same descriptions the model shows when you click a part. This list is the accessible path: nothing in the 3D view is stated only there.

Compute tray 18 × 1U

Two Grace CPUs and four Blackwell GPUs per tray — 36 CPUs and 72 GPUs across the rack.

Each 1U tray carries two GB200 "Bianca" boards. A board is one Grace CPU plus two Blackwell GPUs joined by NVLink-C2C at 900 GB/s. The tray is liquid-cooled through blind-mate quick disconnects at the rear, so it slides in from the front and makes both its data and coolant connections without anyone handling a hose.

NVSwitch tray 9 × 1U

Two NVSwitch5 ASICs each — 18 in total, one per NVLink port on every GPU.

Each ASIC moves 28.8 Tb/s across 36 + 36 ports and carries SHARP engines that perform reductions in-network. They sit in the middle of the rack for one reason: every one of the 5,000-plus copper cables in the spine has to reach them, and cable length is the budget being minimised.

NVLink spine 4 cartridges

Over 5,000 copper cables — around two miles of them — carrying 130 TB/s all-to-all.

Copper, not optical. Optical transceivers and retimers for this many lanes would have drawn roughly 20 kW on their own, about a sixth of the rack’s entire power budget, to move data a metre and a half. At this reach copper is both cheaper and dramatically more efficient.

Busbar 1, full height

A single shared conductor rated for 1,400 A, feeding every tray and replacing 27 pairs of redundant PSUs.

Rack-level power shelves rectify facility input once and drive the busbar; trays tap it directly. Converting power once at rack scale is more efficient than converting it 54 times at server scale, and it reclaims the volume those supplies would have occupied.

Power shelf 8 × 1U, in banks of 4

Eight 33 kW shelves — 132 kW of installed capacity — converting facility input for the whole rack.

Each shelf holds six 5.5 kW supplies, and the eight are arranged 4 + 4. Supermicro quotes an operating power of 125–135 kW for the loaded rack. For scale: the worldwide mean rack density is 7.6 kW, so this is roughly sixteen average racks in one footprint, and almost no existing hall is built for one of these — let alone a row.

Coolant manifold 1 pair, rear

Vertical supply and return pipes that every tray blind-mates into.

Supply runs up one side, return down the other. Each tray’s quick disconnects seal on removal, so a tray can be pulled while the rack is running without draining the loop.

CDU 1 (in-rack or sidecar)

Isolates the rack’s coolant loop from the building’s water.

The pump-and-heat-exchanger unit gives the rack a filtered loop at a controlled temperature and pressure, independent of facility water quality. Inlet is warm by design — 32–45 °C — because warm water can be cooled without a compressor for much of the year.

Rack frame 1

ORv3-inspired, 2,236 × 600 × 1,068 mm, about 1.36 tonnes loaded.

The mass is a real constraint: 1.36 tonnes over a 0.64 m² footprint is more than two tonnes per square metre, which exceeds the floor loading many raised-floor halls were designed for. NVIDIA’s OCP contribution describes over 100 lb of added reinforcement steel in the frame alone.

Mechanical figures

Rack
Rack height	2,236 mm	Roughly a 42U-class footprint; ORv3-inspired / NVIDIA MGX reference rack.
Rack width	600 mm
Rack depth	1,068 mm
Rack weight	~1.36 t	About 3,000 lb — a floor-loading problem for most existing halls. NVIDIA’s OCP contribution describes over 100 lb of added reinforcement steel in the frame alone.
Compute trays	18 × 1U	Split 8 and 10 around the switch band rather than evenly — Supermicro’s rack diagram lists ten above the NVLink switches and eight below.
NVSwitch trays	9
NVSwitch5 ASICs	18	2 per switch tray.
GB200 Superchips per rack	36
Power shelves	8 × 1U	Arranged 4 + 4, each 33 kW from six 5.5 kW supplies — 132 kW of installed shelf capacity feeding the busbar.
Busbar current capacity	1,400 A
Rack power	~120 kW	± sources disagree: 120 kW nominal · 125–135 kW operating (Supermicro) · 132 kW fully loaded (Schneider Electric). Supermicro’s datasheet states an operating power of 125–135 kW and 132 kW of installed power-shelf capacity. Steven Carlini, writing for Schneider Electric: "When fully loaded into a rack, the latest NVIDIA-based GPU servers require 132 kW of power." The commonly quoted ~120 kW is the nominal design figure, not a measured ceiling.

Why 1.36 tonnes is a bigger problem than it sounds

Spread over the rack's 0.64 m² footprint — 600 mmRack width Supermicro by 1,068 mmRack depth Supermicro — ~1.36 tRack weight About 3,000 lb — a floor-loading problem for most existing halls. NVIDIA’s OCP contribution describes over 100 lb of added reinforcement steel in the frame alone. ServeTheHome · NVIDIA Technical Blog is over two tonnes per square metre, before you account for the fact that raised-floor tiles are rated for distributed load, not for four castors. Many existing halls are rated around 700–1,200 kg/m², so an NVL72 is at or past the limit of the floor it would sit on, and the delivery path to get it there has its own limits.

Combined with the ~$2M retrofitCost of liquid cooling per MWSources disagree: ~$2M per MW to retrofit · upwards of $11M per MW for a new greenfield liquid-cooled build STL Partners, May 2026. A widely repeated "$5–10M per MW" retrofit figure is often attributed to Schneider Electric; it does not appear in the Schneider article this site cites, and no primary source for it could be found — so it is not used here. STL Partners (supported by Airedale) quoted to add liquid cooling to existing space,^{STL Partners (supported by Airedale)} this is a large part of why these racks concentrate in new build rather than spreading into existing colocation — though the same research puts a new greenfield liquid-cooled build at roughly five times the per-megawatt cost of a retrofit, so the trade is not one-sided.

The spine deserves its own paragraph

At the back of the rack, four cable cartridges carry >5,000Copper cables in the NVLink spineSources disagree: 5,000 – 5,184 NVIDIA’s OCP contribution says "over 5,000"; other sources cite 5,184. Jensen Huang described it as "5,000 NVLink cables. In total, 2 miles." NVIDIA Technical Blog · ServeTheHome copper conductors between the compute trays and the switch trays. Jensen Huang's description — "5,000 NVLink cables, in total, 2 miles" — is close to NVIDIA's own OCP filing, which specifies four cartridges delivering 260 TB/sAllReduce bandwidth over the spine NVIDIA Technical Blog of AllReduce bandwidth.^{NVIDIA Technical Blog}

The decision to use copper rather than optics is the single most consequential engineering choice in the rack, and it was made on a power budget. NVIDIA's stated position is that the transceivers and retimers for an optical spine would have drawn about ~20 kWPower an all-optical spine would have cost NVIDIA’s stated rationale for copper. SemiAnalysis independently computed ~19.4 kW from 648 × 1.6T transceivers at ~30 W. SemiAnalysis ; SemiAnalysis arrived at ~19.4 kW independently by counting 648 1.6T transceivers at roughly 30 W each.^SemiAnalysis

Against a ~120 kWRack powerSources disagree: 120 kW nominal · 125–135 kW operating (Supermicro) · 132 kW fully loaded (Schneider Electric) Supermicro’s datasheet states an operating power of 125–135 kW and 132 kW of installed power-shelf capacity. Steven Carlini, writing for Schneider Electric: "When fully loaded into a rack, the latest NVIDIA-based GPU servers require 132 kW of power." The commonly quoted ~120 kW is the nominal design figure, not a measured ceiling. Supermicro · Steven Carlini, Schneider Electric — Forbes Technology Council · ServeTheHome budget, that is roughly one sixth of the rack's entire power draw spent on moving data a metre and a half. Copper at that reach costs almost nothing by comparison. The rack's shape — switches centred, trays close-packed, spine short — falls out of the requirement that copper remain viable.