Quick Take
要点速览： Cooling is the line item that strands GPU crates. This checklist names every liquid-cooling entry a Blackwell-class RFQ must carry, from CDU kilowatts and redundancy policy to hose length matrices and leak interlocks, and explains why a 30-76 kW rack cannot be approved on a compute-only quote.
CDU Sizing Lines Buyers Must Spec
速答： Spec each CDU with thermal capacity (kW), N+1 policy, primary/secondary loop interfaces, design facility-water temperature and flow, control/monitoring ports, and spare pumps or filters. “One CDU for the row” without redundancy is a single point of failure on day one of full-cluster training.
Search interest for liquid-cooled AI racks and CDUs keeps rising because buyers who already locked H100, H200, B200, or B300 allocation discover the hall still cannot reject the heat. Power planning alone is not enough — see our rack power and cooling architecture guide. This page is the procurement checklist: which cooling lines must appear on the RFQ so GPU crates are not idle capital.
Related reading: QM9700 / ConnectX-7 fabric BOM, B300 deployment SOP, data center power crisis 2026. Shanghai TaoHan Industrial Co., Ltd. supplies NVIDIA GPU systems and related liquid-cooling infrastructure with configuration support — request a free cooling BOM risk assessment.
| BOM line | What to lock | Why it matters |
|---|---|---|
| Thermal capacity | kW at stated water temps | Undersize = throttle under sync load |
| Redundancy | N+1 or 2N policy | Service without draining the row |
| Loop interface | TCS / FWS connections | Mismatched flanges stop install |
| Controls | Modbus / SNMP / BMS hooks | Cooling events need the same ops board as power |
| Consumables | Filters, coolant, spare pumps | Lead time on spares can exceed GPU ATP |
Row math: sum peak rack loads, apply diversity only with written engineering approval, then pick CDU count so a single unit failure still clears the remaining heat. Related power context stays in the facility planning guide.
DLC vs Active RDHx — Choose Before You RFQ
速答： Direct-to-chip DLC moves heat at the cold plate; active RDHx rejects rack exhaust into a coolant loop at the door. Dense B300-class fills usually need DLC (often with facility-backed CDUs). Do not let a partner substitute “rear door” language without stating active vs passive and the kW rating that matches your fill.
Direct-to-chip (DLC)
Cold plates on GPUs (and often CPUs) connect through rack manifolds to a CDU. Highest density path; hose/QDC quality and leak detection are non-negotiable BOM lines.
Active rear-door heat exchanger (RDHx)
Rejects heat at the rack rear into a liquid loop. Useful in some hybrid halls; still needs CDU capacity and must match the rack’s peak heat — passive doors are a different product class.
Hybrid / platform-mandated kits
OEM reference architectures may require specific manifolds, CDUs, or coolant chemistry. Treat those kit SKUs as performance-critical — not commodity substitutes.
Why Liquid Cooling BOMs Fail Before Go-Live
速答： Most failures are missing lines, not wrong GPUs: no CDU redundancy, wrong DLC vs RDHx assumption, manifolds ordered after racks land, or leak detection treated as optional. Freeze peak rack kW and cooling architecture first, then size CDUs and rack kits as one package with the compute RFQ.
| Failure pattern | What happens | BOM fix |
|---|---|---|
| Compute-only RFQ | Nodes arrive; loop not ordered | Require CDU + manifold lines on every GPU quote |
| Passive RDHx default | Dense Blackwell heat exceeds door capacity | Lock DLC or active RDHx per platform guide |
| CDU at average only | Training peaks throttle or trip protection | Size to peak + N+1, document diversity rules |
| No leak / spare kit | First hose fault stops a row | Detection, QDCs, filters, and hot spares on BOM |
Rack Loop Kit — Manifolds, Hoses, Detection
速答： Every liquid-cooled rack needs manifolds, hose lengths keyed to elevation drawings, quick disconnects, drip trays or leak sensors, and blanking for unused ports. Order these with the CDU — custom hose kits are a classic late-path item that holds GPU go-live after power is already live.
- Manifolds: Port count matches node cold-plate attach points; spare ports for growth if the design allows.
- Hoses / QDCs: Length matrix by rack U and aisle; NVIDIA- or OEM-qualified connectors where the platform requires them.
- Leak detection: Sensor rope or point sensors wired to CDU interlock and ops alert — not a “phase two” aftercutover.
- Filtration / chemistry: Coolant type, inhibitors, and filter change cadence as named BOM lines or service SKUs.
Lead-Time and Compatibility Risks in 2026
速答： CDU and custom manifold ATP often trail GPU crates. Parallel-order cooling with compute; verify DLC vs RDHx language on every quote; lock facility water assumptions; and keep a spare kit. Treat cooling quote windows like any other enterprise hardware — do not assume the loop holds while node quotes move.
- Allocation skew: GPU and CDU lines clear on different calendars — place both in one approval package.
- Architecture bait-and-switch: “Liquid cooling” without DLC/RDHx and kW is not a comparable quote.
- Facility mismatch: Wrong water temperature or flow assumptions force late redesign — attach site data to the RFQ.
- Commissioning lag: Loop fill, leak test, and thermal validation need calendar days after hardware land — budget them.
- Fabric still required: Cooling readiness does not replace InfiniBand or Ethernet fabric BOMs — see QM9700 / ConnectX-7 guide.
Buyer Checklist Before You RFQ
速答： Attach node fill, peak kW per rack, DLC or active RDHx choice, CDU capacity with N+1, manifold/hose matrix, leak detection, and facility water data to the RFQ. Incomplete cooling packages get incomplete quotes — and incomplete quotes miss go-live even when GPU power feeds are already live.
- GPU model and nodes per rack frozen (H100 / H200 / B200 / B300 / other)
- Average and peak kW per rack written (with derating assumptions)
- Cooling architecture locked: DLC / active RDHx / hybrid kit SKUs
- CDU kW, count, and N+1 policy listed
- Manifold, hose length matrix, and QDC class included
- Leak detection and interlock path specified
- Facility water temperature, flow, and flange standards attached
- Commissioning window and spare kit percentages noted
真实应用场景
速答： Map the three scenarios below to your workload, facility, and commercial constraints before freezing CapEx. Equal-BOM RFQs should cite which scenario drives the shortlist.
Scenario 1: Brownfield 20–30 kW Air Hall (Facility Bound)
Workload Profile: Existing CRAC rooms hosting Hopper H200 air nodes with headroom under 30 kW/rack.
Architectural Fit: Air remains viable for H200-class density. B200/B300 liquid-cooled HGX nodes typically force CDU + facility water loops; treating DLC as optional here is a CapEx trap.
Scenario 2: Greenfield 40–100 kW AI Hall (Density Bound)
Workload Profile: New Blackwell Ultra densification targeting 45–100 kW/rack with RDHx or cold-plate DLC.
Architectural Fit: Budget CDU redundancy (N+1), manifolds, and leak detection in the same BOM as GPU servers—not as a facility afterthought.
Scenario 3: Phased Air-to-Liquid Migration (Program Bound)
Workload Profile: Finance approved GPUs first; cooling CapEx lagged one fiscal quarter.
Architectural Fit: Sequence RFQs so GPU PO and CDU PO share the same delivery window; parking liquid nodes on temporary air derates burns warranty and MFU.
常见问题
速答： Dense B300-class racks need active liquid cooling, not passive doors alone. A cooling BOM is CDUs plus rack loop kits and detection. Size CDUs to peak heat with redundancy. Order cooling in the same cycle as GPU nodes so ATP calendars do not strand crates on the dock.
Size CDUs to peak rack heat with N+1, not nameplate average. Spec fluid chemistry, quick disconnects, leak detection, and manifolds in the same PO cycle as the GPU nodes. GB200 NVL72 is rack-scale liquid; B300 dense rows need DLC or active RDHx. This page is the cooling BOM — CDU lines first — so ATP does not strand servers on the dock.
Yes for dense B300-class deployments. Plan CDUs with direct-to-chip loops or active RDHx sized to peak rack heat. Passive rear doors alone are not a safe default for Blackwell-class density.
CDU units with redundancy, loop interfaces, rack manifolds, hoses and quick disconnects, leak detection, filtration, monitoring, and spare kits — alongside the GPU nodes they serve.
Sum peak rack heat loads, apply diversity only with engineering sign-off, then size thermal capacity and N+1 against facility water temperature and flow. Undersizing the CDU throttles GPUs under synchronized training.
Yes whenever possible. CDU and custom hose ATP often trail GPU crates. Parallel approval avoids idle capital while the loop is still on order.
Submit node fill, peak kW per rack, and site water constraints via 提交 BOM / Contact TAOHAN or email info@021taohan.com. You receive a validated CDU / loop package with availability signals from a global free-port hub.
Key Takeaway
核心结论： Decide the architecture first: dense B300-class fills go direct-to-chip, while hybrid halls may still work on active rear doors rated to real rack heat. Then buy CDUs at peak load plus a spare unit, attach site water data, and release the loop package on the same approval as the nodes.