- Home
- Industries
- AI Data Centers
GPU racks that service in minutes, not shifts.
Average rack density hit 27 kW in 2026 — up 69% in a single year — and B200-class cabinets run near 130 kW with roughly 100 kW crossing the cold plates of an NVL72 rack. At $9,000 per minute of unplanned downtime, the couplings on every tray drop have to open dry, close dry and prove it with test data.
Tray swap: drain-and-fill versus dry-break.
The liquid loop is only as serviceable as its worst connection. Here is what changes when every tray drop breaks dry against the OCP UQD envelope.
| Maintenance step | Threaded drain-and-fill loop | DryLink UQDB dry-break |
|---|---|---|
| Coolant at disconnect | branch drained or spilled coolant wiped by hand | both valves seat shut at separation — inside the OCP 0.025 ml dash-04 spill budget |
| Tray swap procedure | unbolt, drain, refill, then bleed air from the branch | unlatch and lift — no drain cart, no refill, no bleed step |
| Air ingestion | every refill pumps air that has to be chased out | loop stays closed, so no air enters the circuit |
| Exposure per swap | hours with a loop open at $9,000/min average downtime cost | minutes per tray — the rest of the rack keeps circulating |
| Cross-brand interoperability | proprietary threads lock the loop to one vendor | OCP UQD Rev 1.0 envelope — compliant plugs and sockets interchange |
| Leak evidence | visual inspection and hope | 100% of sets leak-tested; the report ships in the box |
| Open coolant near live electronics | wicking and drips over the tray below | separation happens dry at both ends, by design |
The coupling is where outage risk concentrates.
Water and cooling-system faults account for roughly 15% of data center outage events, and a single leak incident has been assessed at $190,000 to over $6.32 million once hardware replacement, downtime and remediation are tallied. With 57% of operators reporting their last major outage cost more than $1 million, the maintenance method — not just the hardware — decides the bill.
- $9,000/min — average cost of unplanned downtime, about $540k per hour
- $190k–$6.3M+ — assessed total impact of one leak incident
- 76% (2026) — share of AI servers shipping liquid-cooled, from 15% in 2024
DryLink builds the rack side accordingly: UQDB blind-mate sets that self-align inside up to ±1.3 mm of build tolerance in the UQDB class, rack manifolds ported to your drawing, and a 100% leak test record for every batch that enters your loop.
What a GPU loop asks of its connections.
Floors quoted from OCP UQD Rev 1.0 and the LBNL open specification — the same reference points a hyperscale hardware team audits against.
| Parameter | Value | Why it matters in a GPU rack |
|---|---|---|
| Interface standard | OCP UQD Rev 1.0 (Sept 2020) | the envelope your server and CDU vendors qualify against |
| Sizes | UQD02 / 04 / 06 / 08 — 1/8" to 1/2" | one family from cold-plate drop to manifold port |
| Rated flow class | 0.55 – 4.7 GPM across the four sizes | covers node drops sized on the 1.2 L/min/kW rack basis |
| Spill per disconnect | 0.020 – 0.070 ml @0 psi, by size | tray swaps stay inside a drip-free budget |
| Working / burst pressure | 100 psi working · 300 psi burst (OCP floors) | pump head plus static head, with margin to spare |
| Mate durability | 5000 cycles minimum | years of maintenance windows at every position |
| Node envelope | 1U · 44.45 mm pitch | fits switch and tray layouts without re-layout |
| Blind-mate compensation | 0.8 – 1.3 mm class across UQDB | absorbs rack build tolerance on tray insertion |
| Manifold ΔP budget | < 1/40 of the IT loop (LBNL open spec) | the manifold never becomes the flow bottleneck |
| Rack flow basis | ≈ 1.2 L/min per kW — 80 kW ≈ 96 L/min | port and bore sizing starts from heat load |
| Wetted seals | EPDM standard · FKM / FVMQ options | qualified for DOWFROST LC 25 / JEFFCOOL ISF-25 chemistry |
| Batch documentation | 100% leak test report per batch | evidence for the commissioning and M&O file |
What a rack program order looks like.
Described as a representative configuration, not a customer claim — the numbers below are our standard commercial terms.
“Batch of UQDB04 blind-mate sets for every tray drop, one 8-port supply/return manifold pair per 42U GPU rack — first articles approved and shipping in 15 days, leak reports enclosed carton by carton.”
Rack-side questions, answered with numbers.
Can blind-mate UQDB drops still be serviced by hand during a maintenance window?
What do you need to quote a full rack of tray drops?
How fast can sample sets ship for a pilot rack?
Send the rack loop diagram. Hold the maintenance window to minutes.
UQDB sets and rack manifolds quoted within 48 hours — MOQ from 50 pieces, standard sets ship in 15 days, FOB · CIF · DDP.