Not every AI workload will get a purpose built data centre. AIRSYS is betting that rack-level liquid cooling will help operators squeeze more compute from the facilities they already have
The AI data centre boom has created an assumption that every new GPU cluster needs a purpose built facility, whilst this may be true for the largest deployments, it ignores a much bigger and messier problem: what happens to the thousands of data centres that already exist? – those buildings cannot simply be knocked down and rebuilt every time rack power rises.
Operators need ways to introduce higher-density compute into facilities designed for an earlier generation of IT which makes cooling less of a new-build problem, and more of a retrofit problem and this is where AIRSYS’s strategy becomes interesting; the company is increasingly positioning its cooling technology around that gap – combining liquid and air cooling with a particular focus on bringing higher-density capability into existing or more constrained environments and their latest LiquidRack platform is explicitly aimed at retrofit, edge and mid-density AI infrastructure.
The rack may be the easiest place to start
The obvious response to higher heat loads is to build more cooling capacity at facility level, however, that can be expensive, disruptive and slow. AIRSYS is taking a different route by localising fluid distribution, pumping and thermal management at the rack, the company says this approach is designed to reduce reliance on centralised cooling infrastructure and make the move to liquid cooling more incremental, which is an important proposition – if an operator has a data hall with spare power but limited cooling headroom, then the problem is not necessarily a lack of building space, it may be that the existing thermal architecture was never designed for today’s GPU loads, so putting more of the cooling system at rack level potentially gives operators another way to deal with that constraint.
Liquid cooling does not have to mean a total rebuild
There is a tendency to talk about AI cooling as though operators have only two choices: stay with air or make a complete move to liquid – the reality is likely to be more complicated than this as some workloads need direct liquid cooling whilst others can continue to use advanced air cooling and many facilities often end up with a mixture of technologies. This middle ground seems to be where Airsys appear to be building for and flexibility matters, because AI deployments do not happen in identical ways – a hyperscale campus, a colocation facility, an enterprise data centre and an edge site may all require AI capacity – but they will not have the same power, cooling or physical constraints.
Betting on the physical layer
The company’s strategy is also becoming more substantial on the manufacturing side – following a $40 million investment earlier this year, they have opened their new global headquarters campus in Woodruff, South Carolina with the site set to support manufacturing for high-efficiency cooling technologies and AI infrastructure – they have also positioned the facility as a manufacturing base for customised liquid-cooling components, including 3D-printed components for high-density computing.
This is worth noting – AI data centre growth is increasingly exposing physical infrastructure constraints across the industry and cooling equipment is part of that supply chain – so scaling manufacturing becomes just as important as developing another cooling technology.
Could the retrofit market be the bigger opportunity?
For data centre operators, the interesting part of AIRSYS’s positioning may ultimately have little to do with a single product – it is the underlying assumption that AI infrastructure has to work within buildings that operators already have. The next wave of AI capacity will most certainly include enormous new facilities, but there will also be existing data centre footprints trying to accommodate GPUs, inference workloads and higher rack densities without replacing their entire mechanical infrastructure.
This changes the cooling question from “How do we build a data centre designed for AI?”, to, “How do we make our existing site capable of running AI?”. AIRSYS is effectively building it’s argument around that second question and if AI deployment increasingly becomes a retrofit exercise, then the rack could prove to be one of the most important places to begin solving the cooling problem.
