Published: July 3, 2026 | Category: Product Spotlight | QSCompute
GPU power density has crossed the line where air cooling stops being practical. An 8-GPU H100 node dissipates 5.6 kW. A single rack of four such nodes hits 22.4 kW — beyond what most data center air handlers can manage. And at the edge, fanless enclosures with L40S or RTX 6000 Ada GPUs run into thermal throttling within 15 minutes of sustained inference.
Liquid cooling isn't optional anymore — it's infrastructure. QSCompute stocks the full liquid cooling stack: GPU cold plates, Coolant Distribution Units (CDUs), quick disconnects, and immersion cooling fluids. In stock, pre-tested, ready to integrate with your existing chassis.
| Metric | Air Cooling | Direct-to-Chip Liquid | Immersion Cooling |
|---|---|---|---|
| Heat removal capacity (per rack) | 10–15 kW | 50–80 kW | 100+ kW |
| GPU operating temperature | 75–85°C (throttle zone) | 55–65°C | 50–60°C |
| PUE (Power Usage Effectiveness) | 1.4–1.6 | 1.05–1.15 | 1.02–1.05 |
| GPU density (per rack unit) | 1–2 GPUs / 1U | 4 GPUs / 1U | 8+ GPUs / 1U |
| Noise (dB at 1m) | 70–85 dB (jet engine) | 45–55 dB (office ambient) | 40–50 dB |
| Infrastructure cost (per kW cooled) | $2,000–$3,000 | $4,000–$6,000 | $8,000–$12,000 |
For edge deployments, the killer metric is noise. A fanless IPC with direct-to-chip liquid cooling is silent enough to sit on a factory floor without hearing protection. The GPU runs 20°C cooler, which translates to longer boost-clock duration and higher sustained throughput.
Direct-to-chip cold plates are the core of any liquid cooling loop. We stock cold plates for every GPU we sell — tested for thermal performance, pressure drop, and material compatibility.
| Cold Plate Model | Compatible GPUs | Thermal Resistance | Material | Unit Price |
|---|---|---|---|---|
| CP-L40S-01 | NVIDIA L40S | 0.035 °C/W | Copper + nickel-plated | $320 |
| CP-RTX6000-01 | RTX 6000 Ada, A6000 | 0.038 °C/W | Copper + nickel-plated | $310 |
| CP-H100-01 | NVIDIA H100 PCIe | 0.028 °C/W | Copper micro-fin, nickel-plated | $480 |
| CP-H200-01 | NVIDIA H200 | 0.026 °C/W | Copper micro-fin, nickel-plated | $520 |
| CP-4090-01 | RTX 4090, 4080 Super | 0.042 °C/W | Copper + nickel-plated | $280 |
All cold plates ship with thermal interface material pre-applied (Thermal Grizzly Kryonaut Extreme, 14.2 W/mK). Nickel plating prevents galvanic corrosion in mixed-metal loops.
| CDU Model | Cooling Capacity | Flow Rate | Form Factor | Unit Price | Status |
|---|---|---|---|---|---|
| CDU-5KW | 5 kW | 15 L/min | 1U rackmount | $6,500 | In Stock |
| CDU-15KW | 15 kW | 40 L/min | 2U rackmount | $12,800 | In Stock |
| CDU-50KW | 50 kW | 120 L/min | 4U rackmount | $28,000 | Build-to-Order |
CDUs handle the secondary side: pump, heat exchanger, reservoir, and control electronics. The facility water (primary side) connects to the CDU; the CDU delivers temperature-controlled coolant to the cold plates (secondary side). N+1 pump redundancy on all models. Modbus TCP for integration with BMS/DCIM.
| QD Model | Type | Flow Rate | Drip-Free | Unit Price (pair) |
|---|---|---|---|---|
| QD-NS4-01 | No-spill, 1/4" barb | 8 L/min | Yes | $45 |
| QD-NS6-01 | No-spill, 3/8" barb | 20 L/min | Yes | $65 |
| QD-BLIND-01 | Blind-mate, panel mount | 25 L/min | Yes | $180 |
Quick disconnects let you hot-swap GPUs without draining the loop. Blind-mate QDs are for blade-style chassis where the GPU sled automatically mates with the cooling manifold on insertion. All QDs are rated for 10,000+ connect/disconnect cycles and use EPDM seals compatible with propylene glycol coolants.
| Fluid | Type | Dielectric Strength | Boiling Point | Price (20L) |
|---|---|---|---|---|
| IC-DIELECTRIC-01 | Synthetic hydrocarbon | >40 kV | 175°C (single-phase) | $980 |
| IC-FLUORO-01 | Fluorocarbon-based | >50 kV | 56°C (two-phase) | $2,400 |
Single-phase fluid for tank-based immersion cooling. Two-phase fluid for high-density deployments where boiling heat transfer at the GPU surface provides maximum thermal performance. Both are non-conductive, non-toxic, and non-flammable.
| Scenario | Configuration | Total Cost | Cooling Capacity |
|---|---|---|---|
| Single L40S edge node | 1 × CP-L40S-01 + 1 × CDU-5KW + 2 × QD-NS4-01 | $6,910 | Up to 350W (single GPU) |
| 4-GPU inference server | 4 × CP-H100-01 + 1 × CDU-15KW + 8 × QD-NS6-01 | $15,240 | Up to 1.4 kW (4 GPUs) |
| 8-GPU training node | 8 × CP-H200-01 + 1 × CDU-50KW + 16 × QD-BLIND-01 | $35,040 | Up to 5.6 kW (8 GPUs) |
Every configuration is pre-assembled, pressure-tested at 2× operating pressure for 8 hours, and shipped with coolant pre-fill. You connect facility water and power — it's running the same day.
Liquid cooling components in stock — tested, assembled, and ready to ship.
Not sure what you need? Send us your GPU configuration and chassis model. We'll spec the right cold plates, CDU, and plumbing — with CAD drawings for your mechanical team.
Contact: +86 189-9192-7716 | info@qscompute.com