Rising component costs may adjust prices on selected items without prior notice. Supply constraints may add shipping time — we’ll notify you if affected. OEM costs and supply tightening — contact us to secure stock and pricing. For projects or bulk orders, choose “Request for Quotation” on any item.
Call our specialists(65) 6978 3502 Email our sales teamsales@sourceit.com.sg

Servers & Power

The First 30 Days: What Really Happens When a Liquid-Cooled AI Rack Arrives

Project plans for AI racks tend to end with a line that says "delivery". Then the crate arrives and everyone discovers that delivery is where the careful work begins.

An air-cooled server can be racked and running in an afternoon. A liquid-cooled rack cannot. Between the loading bay and the first useful job there is a month of steps, and skipping any of them is how expensive hardware gets damaged.

The month at a glance

When Stage What happens
Before arrival Prepare the pipes Site pipework is pressure tested, cleaned and flushed before any rack is connected
Week 1 Position and connect Rack moved in, inspected for transit damage, power and coolant lines connected
Week 2 Test, fill and bleed Leak tests, coolant fill, air purged from the loop, CDU settings applied, coolant chemistry checked
Week 3 Burn-in The rack runs at sustained load so weak components fail now, under warranty and before production
Week 4 Accept and hand over Performance, network and storage checks, documents handed over, operations team trained

The four-week layout is our indicative framing. Real schedules vary with the site, the vendor and how much preparation was done in advance. The steps themselves follow the sources cited below.

Before arrival: clean pipes or nothing

The channels inside a GPU cold plate are extremely fine. Debris left in site pipework from welding or cutting will find them. That is why the Open Compute Project publishes guidance specifically on preparing pipework before racks are connected: hydrostatic testing, cleaning, flushing and preparation for commissioning. Industry guidance on filtration often recommends a final target in the low single-digit micron range.

None of this involves the rack. All of it has to be finished before the rack is plumbed in.

Sources: Open Compute Project: pre-commission preparation of TCS row manifolds, Data Centre Magazine: from commissioning to operations

Week 2: the sequence that matters

One integrator's commissioning guide sets out the order plainly: flush, pressure test, fill, purge air, set the CDU, verify chemistry, then soak under load. The pressure test comes in two parts. First a dry test with air or nitrogen, held for typically 30 to 60 minutes with no measurable drop. Only then a wet inspection at operating pressure.

The dry test exists for an obvious reason. If a joint is going to fail, you want it to leak gas, not coolant.

Source: RDP: Commissioning a liquid-cooled AI rack

The sign-off sheet

Every step should leave a number behind. If you are the customer, these are the records to ask for.

Record What it proves
🧹 Flush and filtration results The pipework was clean before the rack was connected
📉 Pressure-test readings and duration The loop held pressure, dry and wet
🧪 Coolant chemistry report The fluid meets the server maker's specification on day one
🌡️ CDU setpoints, flow and temperatures The cooling is set for your room, including the margin above dew point
🔥 Burn-in log Every GPU ran at load for the agreed period without errors or throttling
💾 Storage and network test results The GPUs will not be starved of data or bandwidth
📁 Handover pack As-built rack drawings, serial and firmware inventory, configuration exports, runbooks, credentials and licences

Source for the handover pack contents: RDP knowledge base: deployment and commissioning

Week 3: let it fail now

Electronics tend to fail early or late, and rarely in between. Burn-in runs the system hard for days so that the early failures happen while the installer is still on site and the parts are plainly under warranty. A rack that passes a quick power-on check and goes straight into production has skipped the cheapest test it will ever get.

The step that gets skipped

Here is the uncomfortable part. Vertiv, citing an Uptime Institute update from January 2026, warns that the pressure to bring AI capacity online quickly is leading some operators to skip verification steps such as factory witness testing of coolant distribution units, and to rely on the manufacturer's assurances instead.

The incentive is easy to understand. A rack of this class is costly every day it sits idle. But the trade is poor: days saved against the risk of contaminating or overheating the most expensive components in the building.

The counter-view: not every site needs every test. A single rack with an in-rack CDU, installed by the vendor's own engineers, is a different proposition from a hall of field-built pipework. Proportion is reasonable. Skipping the records is not.

Source: Vertiv: why "flush and fill" is no longer enough

Five questions for your installer

  1. Who performs each stage, and are they certified by the server maker?
  2. What are the pass criteria for the pressure test and the coolant test?
  3. How long is burn-in, and what workload is used?
  4. What happens to the schedule if a component fails during burn-in?
  5. What is in the handover pack, and who trains our operations team?

Agree the answers before the purchase order. They are much harder to negotiate with a crate in the loading bay.

Related on SourceIT

SourceIT Approved

SourceIT Approved

Researched and verified by the SourceIT team for Singapore buyers.

Share This Article
Facebook

More from our blog


← Older Post Newer Post →