How Many UQD Couplings per NVL72 Rack? GB200 vs GB300
"How many quick disconnects are in a rack?" is now a procurement question, not just an engineering one. This is the per-tray math behind the headline figures — ~126 UQD pairs for GB200 and ~250 for GB300 per NVIDIA NVL72 rack — and what the jump means for sourcing.
The NVL72 rack, tray by tray
An NVIDIA NVL72 rack is built from 18 compute trays and 9 switch trays. Every tray that carries heat needs coolant delivered and returned, so the coupling count is driven mainly by how the compute trays are cooled. Coolant reaches the trays through a rack manifold; each tray connects to that manifold with quick disconnects — increasingly blind-mate (UQDB) couplings so a tray connects automatically on insertion.
We count in pairs because coolant always needs a supply and a return path. One "pair" is therefore one supply coupling plus one return coupling.
GB200: why ~126 pairs
In the GB200 class a compute tray uses large-area cold plates. In practice a tray needs about six coupling pairs — cold-plate supply and return, plus the connection to the manifold. Doing the arithmetic across the rack:
| GB200 build-up | Count |
|---|---|
| Coupling pairs per compute tray | ~6 |
| Compute trays per NVL72 rack | 18 |
| Compute-tray pairs (18 × ~6) | ~108 |
| Plus switch trays & manifold connections | remainder |
| Total per rack (≈) | ~126 |
The 18 compute trays account for roughly 108 pairs; the 9 switch trays and manifold connections bring the practical total to about 126 pairs per rack. The exact number depends on the specific tray and manifold design.
GB300: why ~250 pairs
The GB300 generation switches from large-area cooling to an individual cold plate per chip. Each chip's cold plate needs its own supply and return, so the pairs per compute tray roughly double: about six chips ≈ 12 pairs, and with the manifold connection roughly 14 pairs per compute tray.
| GB300 build-up | Count |
|---|---|
| Coupling pairs per compute tray | ~12–14 |
| Compute trays per NVL72 rack | 18 |
| Compute-tray pairs (18 × ~13) | ~234 |
| Plus switch trays & manifold connections | remainder |
| Total per rack (≈) | ~250 |
Market reports estimate around 250 pairs per NVL72 rack for GB300 — almost a doubling versus GB200. At the same time the couplings get smaller (a more compact UQD generation), which raises the bar on manufacturing tolerance, sealing and blind-mate alignment even as the count climbs.
GB200 vs GB300 at a glance
| Feature | GB200 (NVL72) | GB300 (NVL72) |
|---|---|---|
| Cooling concept | large-area cold plate | individual cold plate per chip |
| Coupling pairs per compute tray | ~6 | ~12–14 |
| Coupling pairs per rack (≈) | ~126 | ~250 |
| UQD form factor | UQD / UQDB (OCP) | more compact UQD generation |
Figures are industry estimates from market reports and serve as orientation; the exact quantity depends on the specific tray/manifold design.
What the jump means for procurement and second sourcing
Multiply ~250 pairs per rack by the number of racks in a build-out and UQD becomes one of the fastest-growing bill-of-material lines in an AI data center — and one that has already shown supply tightness across the established makers. Two consequences follow:
- Volume risk. A single-source coupling supply is a single point of failure for an entire deployment schedule. Doubling the count from GB200 to GB300 magnifies that exposure.
- Standardization pays off. Because the couplings are built to the open OCP UQD interface, a qualified alternative can drop in without redesigning the manifold — provided it matches the interface, sealing (non-spill) and mating-cycle life.
Netonx (via Green MeOH): OCP/UQD-compliant, leak-free, from NVIDIA-qualified production — an alternative or second source to CPC, Stäubli and Parker, with a European point of contact.
To Liquid Connectors & inquiry →Frequently asked questions
How many UQD couplings does a GB200 NVL72 rack use?
Roughly ~126 coupling pairs. A GB200 compute tray needs about six pairs; across 18 compute trays plus 9 switch trays and manifold connections this comes to about 126, depending on the build.
Why does GB300 need about 250 UQD pairs?
GB300 uses an individual cold plate per chip, so pairs per compute tray roughly double (about 12–14). Market reports estimate ~250 pairs per NVL72 rack — nearly double GB200.
What does the higher count mean for procurement?
More couplings per rack across more racks makes UQD a fast-growing, supply-constrained line item. Buyers de-risk delivery by qualifying an OCP-compatible second source with proven sealing and volume.
Read more: GB200 / GB300 liquid cooling overview · UQD connectors explained · Liquid Connectors (UQD) – overview · UQD vs. proprietary quick disconnects · Data-center liquid cooling · Vera Rubin (VR200 NVL72) liquid cooling
Note: brands such as NVIDIA, CPC, Stäubli and Parker belong to their respective owners; references are for objective comparison. Quantities are industry estimates.