Configuring Liquid-Cooled AI Server Racks in 2026: Rack Cooling for GB300, GB200 and H200
2026年11月19日
Procuring a liquid-cooled AI server rack typically involves two purchases. The server vendor supplies the GPU servers, and the rack cooling package is configured separately, including the cabinet, distribution manifolds, CDU and cold plates. For AI data halls that have already selected their GPU servers, Lianli offers a 48U liquid-cooled cabinet or ORV3-44OU rack, UQD04 distribution manifolds, 4U 65 kW or 6U 130 kW in-rack CDUs, and cold plates including a GB300 plate for the GB300 NVL72 platform and a 1,000 W GPU plate. Lianli supplies the rack and cooling equipment, and the server vendor supplies the servers.
What a Liquid-Cooled Server Rack for AI Includes
Lianli's rack cooling package has five layers. Each layer must match the next in connector specifications, flow distribution and temperature settings.
1. Rack or cabinet: houses the servers and supports the manifolds.
2. Distribution manifolds and quick connectors: distribute coolant to each server through UQD04 connectors.
3. Cold plates: mount directly on the GPUs and CPUs inside each server.
4. CDU: circulates the secondary loop and transfers heat to facility water or room air.
5. Primary-side heat rejection: an outdoor dry cooler or cooling tower.
The server hardware, including integrated rack-scale systems, comes from the server maker. Lianli supplies the surrounding cooling infrastructure across layers 1 to 5.
Liquid-Cooled AI Rack Packages by GPU Platform
GPU platform | Lianli cold plate | Lianli rack | Lianli CDU |
NVIDIA GB300 NVL72 | GB300 cold plate: 6,200 W total power, 45 °C inlet, DI water or PG25, UQD04 connectors | 48U liquid-cooled cabinet or ORV3-44OU | Liquid-to-liquid, sized to the rack's liquid-cooled load |
NVIDIA H100 / H200 | GPU cold plate: 1,000 W, 40 °C inlet, 1 L/min, 10 kPa | 48U liquid-cooled cabinet | 4U 65 kW or 6U 130 kW in-rack |
AMD Instinct MI300X/MI350X | GPU cold plate: 1,000 W, PG25, 40 °C inlet | 48U liquid-cooled cabinet | 4U 65 kW or 6U 130 kW in-rack |
RTX PRO 6000 servers | RTX PRO 6000 cold plate: 1,000 W, 40 °C, 1 L/min | 48U liquid-cooled cabinet | In-rack, by rack load |
Rack Options for AI: 48U Liquid-Cooled Cabinet and ORV3-44OU
Lianli offers two main rack formats for liquid-cooled AI servers:
- 48U liquid-cooled cabinet: Standard 19-inch format, support for a 50 kW liquid-cooled load, 20 UQD04 male connectors, two distribution manifolds and a flow rate of 12 m³/h. It fits standard 19-inch GPU servers equipped with cold plates.
- ORV3-RACK-44OU: 21-inch format, 44OU, 1,200 mm depth and 48 V power. The DLC-ready rack reserves space for a CDU and liquid cooling pipes. It accepts 19-inch equipment with an adapter frame and can ship fully assembled or in parts.
For retrofits of air-cooled data halls, Lianli offers 35 kW and 55 kW air-liquid hybrid cabinets designed for server platforms including NVIDIA HGX, DGX H100/H200, B100/B200 and GB200/GB300.
Pairing a CDU With Each Liquid-Cooled AI Rack
Lianli's in-rack liquid-to-liquid CDU mounts directly inside the AI rack.
- 4U 65 kW (6 m³/h) or 6U 130 kW (12 m³/h)
- rated temperatures of 30/40 °C on the primary side and 40/50 °C on the secondary side
- N+1 redundant pumps and dual power supplies
- a three-year factory warranty covering the pumps, heat exchanger and control electronics
When multiple AI racks share one cooling loop, use a Lianli in-row liquid-to-liquid CDU rated from 500 to 2,000 kW.
Cold Plates for GB300 and H-Series GPUs
Lianli's GB300 cold plate for the GB300 NVL72 platform lists a 1,400 W rating per GPU across four GPUs and 300 W per CPU across two CPUs, for 6,200 W in total. It has a 45 °C inlet, supports DI water or PG25, and uses UQD04 quick connectors and stainless steel liquid tubes.
The GPU cold plate has a TDP of 1,000 W and uses PG25 coolant. Its specifications include a 40 °C inlet, a flow rate of 1 L/min and a flow resistance of 10 kPa. It is made of pure copper, has a G1/4" threaded connection, operates at 3 to 5 bar and withstands up to 2.5 MPa. Supported models include NVIDIA H100, H200, A100, GB200, GB300, B200 and B100, AMD Instinct MI300X, MI300A, MI325X and MI350X, as well as Huawei Ascend and MetaX series.
A GB200 NVL2 Liquid-Cooled Rack Example
For an unnamed data center project in Shenzhen using NVIDIA GB200 NVL2, a two-GPU configuration, Lianli supplied copper microchannel cold plates and a room-level CDU. The target delivery timeline was seven days.
For a larger unnamed project in Xinjiang, Lianli provided liquid cooling for 128 NVIDIA H200 GPU servers with a total IT load of about 3 MW. The scope included cold plates, rack distribution manifolds, row-level and room-level CDUs, outdoor dry coolers, piping and a DCIM system. The target delivery timeline was 60 days.
Before requesting a quote from Lianli, confirm the server model and quantity, redundancy requirements (N+1 or 2N), local temperature and humidity, three-phase voltage and frequency, UPS and diesel generator requirements, and the project's current stage and budget. The minimum order is one unit.
Summary
Lianli suits AI teams that have selected their GPU servers and need a complete rack cooling package. Its range includes 48U liquid-cooled cabinets or ORV3-44OU racks, UQD04 distribution manifolds, 65 kW and 130 kW in-rack CDUs with N+1 redundant pumps, and GB300 and 1,000 W GPU cold plates. All components come from one factory and share a warm-water design. Size each component to the liquid-cooled load per rack specified by the server vendor.
Send the GPU platform, server count and rack format through the Get a Quote form on lianliwork.com or WhatsApp +86-18002936305.
FAQ
Does Lianli sell liquid-cooled AI servers?
Lianli supplies the rack, manifolds, CDU, cold plates and primary-side heat rejection equipment. The server vendor supplies the servers. Send the server model so Lianli can match the cooling package to it.
Which rack suits GB300 NVL72 liquid cooling?
Use a rack with UQD04 distribution manifolds and a liquid-to-liquid CDU sized to the rack load. Lianli's GB300 cold plate is rated at 6,200 W total with a 45 °C inlet and UQD04 connectors. Its 48U liquid-cooled cabinet includes manifolds, and its ORV3-44OU rack reserves space for a CDU.
What CDU goes with a liquid-cooled H200 rack?
Lianli's 4U 65 kW and 6U 130 kW in-rack CDUs fit one H200 rack on facility water, with a 30/40 °C primary loop. For a row of H200 racks on one loop, use an in-row unit from 500 kW. Size the CDU according to the server vendor's liquid-cooled figure.
How long does a liquid-cooled AI rack package take to deliver?
Production and shipment take about 4 to 8 weeks after signing, with another 2 to 5 weeks for sea freight. Timing depends on the model, quantity and customization requirements.
Which Lianli parts fit an NVIDIA GB200 rack?
Lianli lists NVIDIA GB200 as compatible with its GPU cold plate and identifies GB200/GB300 server racks as target applications for its 4U 65 kW and 6U 130 kW in-rack CDUs. For an unnamed GB200 NVL2 data center project in Shenzhen, Lianli supplied copper microchannel cold plates and a room-level CDU.
Can server OEMs and integrators buy Lianli cold plates or CDUs on their own?
Yes. Lianli supplies cold plates and CDUs to server OEMs and integrators, with a minimum order of one unit. In March 2026, Lianli delivered NVIDIA H200 and AMD SP5 liquid cold plates in volume to a customer in Taiwan.
Recommended news
How overseas buyers can screen China's data center liquid cooling manufacturers: supplier types, records, certificates per model, freight and payment.
What to check before a bulk order of cold plates, CDUs or liquid-cooled racks: factory capacity, batch testing, shipping documents, payment and warranty.