Estimating cooling capacity is the first step in an engineering review, not final equipment selection. An oversized system can add unnecessary cost, while an undersized one can leave IT equipment without enough cooling at design conditions. The five steps below show how to build a transparent load schedule for AI racks. Every number in the worked example is an illustrative assumption, not an HY product specification or performance guarantee.
What heat load must the liquid loop actually remove?
Nearly all electricity consumed by IT equipment eventually becomes heat. Build a rack-by-rack schedule using the approved server configuration, typical operating load, peak design load, and expected growth. Nameplate ratings are useful limits but do not necessarily describe a site’s normal operating load. Check whether network equipment, storage, power conversion, and other loads are counted in the same boundary.
How do you split the load between liquid and residual air cooling?
The liquid heat-capture fraction is the portion transferred through cold plates and the liquid loop. The remainder still needs an air-cooling path. Memory, storage, network switches, PDUs, fans, and heat leaking from liquid-cooled components can contribute to room heat. Obtain the capture fraction from the exact server and cold-plate design; do not assume the same percentage for every rack. See how cold plates, CDUs and dry coolers work for the loop interfaces.
How do ambient conditions change the estimate?
Outdoor temperature and humidity affect the heat rejection equipment. Its rated capacity must be checked at the site’s design weather, desired supply temperature, and approach temperature, which is the difference between the cooling fluid temperature and the outdoor condition used in the equipment rating. High altitude and extreme temperature can require derating. Water availability and local limits may change whether dry cooling, evaporative assist, or a closed tower is appropriate.
How much redundancy should you add?
State the failure scenario first. In an N+1 arrangement, the remaining units must carry the required design load after one unit is unavailable. Redundancy is a capacity and control arrangement, not a universal percentage added to the load. Check pumps, CDUs, heat rejection units, power feeds, controls, and maintenance isolation separately. The container planning guide covers these project boundaries.
Worked example: what does a transparent estimate look like?
Illustrative assumptions only: ten racks, each with a 120 kW design IT load, produce a 1,200 kW combined IT load. Assume the specific server design captures 75% of this heat in liquid. The liquid loop therefore carries 1,200 × 0.75 = 900 kW, while the remaining 300 kW requires an air-cooling path. These figures are not HY equipment ratings, GB300 specifications, or measured operating results.
For an illustrative N+1 liquid-side arrangement, three equal 450 kW units provide 900 kW when any one unit is unavailable. This assumes each unit can deliver 450 kW at the site’s design conditions and that pumps, piping, controls, and power support that failure mode. Heat rejection capacity, air cooling, design margin, and auxiliary loads still need separate checks. Different rack loads or capture fractions produce different results.
When should an estimate become an engineering review?
Move to a project-specific review before equipment selection whenever the rack configuration is unsettled, design weather is severe, altitude is significant, redundancy is required, or a PUE or availability target must be committed. Bring the server bill of materials, rack loads, cooling fluid requirements, site climate, electrical service, and scope boundary to the review. The GB300 project design checklist provides a structured input list.
Frequently asked questions
Can I cool AI racks with air alone?
It depends on the rack heat density, server requirements, air distribution, room conditions, and available infrastructure. Confirm the server vendor’s cooling requirements and calculate the full load before choosing an approach.
What data do I need before sizing liquid cooling?
Provide the exact server and rack configuration, typical and design electrical loads, approved liquid capture fraction, coolant requirements, flow and temperature limits, site climate, redundancy target, and the boundary between supplier and site systems.
Does a bigger CDU always mean a safer design?
No. The CDU must match the required flow, pressure, temperature, heat exchange duty, controls, and redundancy. A higher nameplate capacity does not resolve a mismatch in server interfaces or heat rejection conditions.
Next step
Share your load schedule and site conditions through HY’s liquid cooling solutions page for a configuration review.