“This Rack May Reach 80 kW”: Before Quoting Liquid Cooling, Find the Heat Boundary
A cooling-vendor BD lead should verify what the 80 kW figure means, what the servers support, where remaining heat goes and whether the site wants a study or a retrofit quote.
Signals to watch
- The rack-power figure is identified as measured, peak, nameplate or planned
- The exact IT equipment and the portion of heat still handled by air are known
- Facility water, heat rejection and an outage or survey window are connected to the request
An “80 kW rack” message should start a boundary check, not an equipment quote. The cooling-vendor BD lead first needs to learn whether 80 kW is a measured sustained load, a short peak, a nameplate sum or a future design target. Then the team must locate the heat: what the IT equipment can transfer to liquid, what remains in air, and where the facility can finally reject it.
This reader watches colocation, AI-infrastructure and facility-engineering Telegram groups the company deliberately connected and may access. The desired Signal is an existing hall approaching a real survey, feasibility or retrofit decision. Seeing it a day late can mean missing the only site walk before a change window.
Start with the number’s measurement basis
Use this as an illustrative composite fragment, not a customer statement or market statistic:
“One rack may hit 80 kW on the next GPU batch. Existing hall, no idea if rear-door is enough. Need options before maintenance window.”
The number is intentionally hypothetical. It does not represent average rack density, an ASHRAE recommendation or a threshold at which liquid cooling becomes mandatory.
Four meanings could sit behind it:
- the sum of device nameplate ratings;
- a future design maximum;
- a short observed peak; or
- sustained measured IT load.
Those values lead to different engineering work. Ask for the time basis, measurement point and whether the figure applies to one rack or an average across a row. Until then, the salesperson has a reason to investigate—not a cooling duty to quote.
Check 1: what can the installed equipment actually support?
The same rack power can lead to different options. Some servers may support direct-to-chip cold plates; others may remain air cooled. A rear-door heat exchanger changes heat removal at the rack boundary. Immersion places equipment in a dielectric fluid and changes service, materials and operating procedures. These are not interchangeable product labels.
The exact server and accelerator models matter because the equipment vendor controls supported configurations, fluid interfaces, firmware dependencies and warranty conditions. “GPU rack” is not an adequate bill of materials.
ASHRAE TC 9.9’s 2021 liquid-cooling white paper explains why rising compute, memory and storage power makes liquid cooling necessary in some cases. It also says careful analysis is needed to architect a cold-plate water-cooling system. That supports an engineering review, not a design approval from one chat message; the current project still needs the relevant equipment and site records.
Check 2: how much heat remains in the room?
Direct-to-chip cooling moves heat from selected components into a liquid loop, but other components may still cool through air. The ASHRAE paper gives one explicit hybrid example: a CPU is cooled by liquid while the memory subsystem remains air cooled. The fraction depends on the exact server design, so a generic architecture label cannot supply the project-specific split.
That residual fraction changes the question. A rack can gain cold plates and still overload room airflow if the remaining air load, neighboring racks and failure condition are ignored.
The first calculation therefore needs two loads, not one:
- heat expected to enter the technology coolant; and
- heat that room air must continue to carry.
The site may not know the split yet. That is a reason to request OEM thermal data or a measured assessment, not to choose an optimistic percentage.
Check 3: where does the liquid release its heat?
A CDU, or coolant distribution unit, circulates and controls the technology-side coolant. Depending on the architecture, it may transfer heat to facility water or reject it back to room air.
That distinction exposes two very different retrofit limits.
With a liquid-to-air arrangement, the existing room cooling plant still receives the heat. ASHRAE notes that a one-rack system can use an in-rack liquid-to-air heat exchanger where facility water is unavailable, but the populated rack can still add significant heat to the data center and consume remaining air-handler capacity. With facility water available, the paper describes a rack-mounted CDU. The site must then check connection, temperatures, flow, pumping, controls and final heat rejection.
The group fragment says none of this. It also leaves water chemistry, filtration, materials, pressure, leak detection, drainage, floor loading, pipe routes, electrical distribution and outage tolerance unknown.
Check 4: what is the site buying this week?
“Need options” can mean three commercial outputs:
A feasibility answer. The site wants to know whether its next hardware batch can fit the existing hall. The first deliverable is a data request and engineering review.
A site assessment. The team wants load measurements, equipment verification and a survey of power, water, space and heat rejection. A dated access window makes this materially more actionable.
A retrofit proposal. The site has selected the equipment family, confirmed responsible engineers and wants a defined scope, commissioning plan and commercial offer.
Do not move directly from the first output to the third. A request for “options” is not proof that a project is funded or that the poster may approve the work.
What the BD handoff should contain
After the first review, the record might say:
Illustrative high-density rack planned for an existing hall. The poster cites 80 kW but has not confirmed whether it is measured or designed. Server models, liquid-supported configuration and remaining air load are unknown. A maintenance window is mentioned; site, facility-water capacity, heat rejection, power, outage allowance, decision owner and budget remain unconfirmed. Route first to a technical site-assessment conversation.
That statement separates the reason for urgency from the missing engineering facts. It also keeps the commercial question visible: the site may be buying an assessment before it buys hardware.
TOP Prospect can find and connect relevant fragments in Telegram groups the user intentionally connected, preserve the original messages, source and time, remove clear duplicates and place the candidate in a human review queue. It cannot read private chats, measure rack load, check warranties, enter a facility or select a cooling design.
For a broader capacity request, compare when a colocation request is ready. The Southeast Asia capacity scenario separates temporary power need from a permanent build, while the data-center demand-discovery article explains why technical fragments often appear before a formal request.
Frequently asked questions
Does an 80 kW rack automatically require liquid cooling?
No. In this article 80 kW is an illustrative value, not an industry average or engineering threshold. The load basis, IT equipment, airflow, heat capture, facility water, power and operating constraints determine the options.
What is a CDU in a liquid-cooling project?
A coolant distribution unit manages and circulates the technology-side coolant and, depending on its design, transfers heat to facility water or back to room air. Its suitability depends on temperatures, flow, pressure, chemistry, capacity and resilience requirements.
Why does residual air cooling still matter with cold plates?
Direct-to-chip cold plates do not necessarily capture every watt from a server. Memory, power supplies, networking and other components may continue to reject heat to air, so the room cooling boundary must be checked.
Who should select the cooling architecture for a retrofit?
The accountable facility and project engineers should decide with the relevant server, cooling and controls vendors. They need verified equipment support, heat loads, water loops, safety, warranty and operating constraints.
