NVIDIA Vera Rubin (VR200 NVL72) Liquid Cooling: What Changes for Quick Disconnects
The next GPU generation after Blackwell roughly doubles the cooling flow per rack and moves to fully fanless trays. For procurement, that makes UQD quick disconnects an even more critical — and scarcer — line item.
From NVL144 to VR200 NVL72: what Vera Rubin is
Vera Rubin is NVIDIA's successor platform to Blackwell (GB200/GB300), presented at CES 2026 and in full production for shipment in H2 2026. The rack originally announced as "NVL144" (counting GPU dies) is now officially the VR200 NVL72: 72 GPU packages with two dies each, plus Vera CPUs, in a rack-scale NVLink domain — the same 72-package logic as GB200/GB300.
Thermally, the platform continues the direct-to-chip (D2C) approach of its predecessors, but at a significantly higher power density per rack position. Both the compute trays and the NVSwitch trays adopt a fanless design: virtually all heat leaves the rack through the warm-water liquid loop rather than the air path.
Cooling flow roughly doubles — and the fluid network with it
Supply-chain checks around CES 2026 indicate that the rack's TCS (technology cooling system) flow increases by roughly 100 % versus GB300 NVL72. More flow per rack means the entire fluid network is re-dimensioned: CDU capacity becomes a first-order design parameter, manifolds and cold plates are upgraded, and quick disconnects benefit from spec and/or count upgrades — larger effective flow paths, lower pressure drop per coupling, unchanged non-spill (dry-break) requirements.
| Aspect | GB200 (NVL72) | GB300 (NVL72) | VR200 NVL72 (Vera Rubin) |
|---|---|---|---|
| Cooling concept | large-area cold plate, D2C | individual cold plate per chip | fanless trays, warm-water D2C |
| Coupling pairs per rack (≈) | ~126 | ~250 | high count; spec/count upgrades expected |
| Rack cooling flow (TCS) | baseline | above GB200 | ~2× GB300 (supply-chain estimates) |
| Heat removal via air | residual fans | residual fans | virtually none (fanless) |
| Availability | shipping | shipping | H2 2026 |
Figures are industry and supply-chain estimates (CES 2026 reporting, analyst checks) and serve as orientation; exact quantities depend on the final tray/manifold design. NVIDIA has not published official coupling counts for VR200.
Why quick disconnects become the bottleneck item
Three effects stack up in the Vera Rubin ramp. First, volume: NVL72-class racks already consume large coupling quantities (GB300 ≈ 250 pairs per rack, per market reports), and gigawatt-scale AI factory buildouts multiply that across thousands of racks. Second, specification: double the flow at unchanged non-spill and blind-mate requirements tightens tolerances on valve geometry, seals and plating. Third, supply concentration: the cooling supply chain is consolidating around a few large vendors, so single-sourced couplings become a genuine schedule risk.
The practical answer is the same one hyperscalers apply across the BOM: qualify a second source early. Because UQD/UQDB is an open OCP standard and cross-vendor mate-compatible, a second supplier can be qualified against the same interface without redesigning trays or manifolds — see our overview of UQD vs. proprietary quick disconnects and the per-rack coupling math.
Netonx (via Green MeOH): OCP/UQD-compliant, non-spill, >5,000 mating cycles — European stock, samples and engineering support as your second source.
To Liquid Connectors & inquiry →Frequently asked questions
What is the NVIDIA Vera Rubin VR200 NVL72?
NVIDIA's GPU generation after Blackwell. The rack announced as VR200 NVL144 (counting dies) is now the VR200 NVL72 — 72 GPU packages of two dies each — fully liquid-cooled and shipping in H2 2026.
How does Vera Rubin change liquid cooling versus GB300?
Rack cooling flow (TCS) roughly doubles per supply-chain estimates, and compute/switch trays go fanless — nearly all heat is removed via warm-water direct-to-chip cooling. CDUs, manifolds, cold plates and quick disconnects are upgraded in spec and/or count.
What does this mean for UQD procurement?
Higher flow at unchanged non-spill requirements, high per-rack coupling counts, and a consolidating supplier base. Qualifying an OCP/UQD-compliant second source early reduces schedule risk for 2026/27 deployments.
Read more: GB200 & GB300 liquid cooling · UQD couplings per NVL72 rack · UQD vs. proprietary quick disconnects · UQD connectors explained · Liquid Connectors (UQD) – overview
Note: brands such as NVIDIA belong to their respective owners; references are for objective comparison. Flow and quantity figures are industry and supply-chain estimates as of 07/2026.