I see a lot of posts about power these days and how much we’re going to need, but much less about the optimisation everybody is working on. After our conversations with APR Technologies, ZutaCore, and Accelsius with different cooling solutions, I think the cooling efficiency of next-generation AI data centres will improve significantly. The ZutaCore/Munters comparison gives a sense of the potential: modelled annual cooling energy roughly 74% below its specified single-phase chilled-water baseline, with partial PUE falling from 1.115 to 1.03. Just to be clear, this is cooling-energy savings, not a 74% reduction in the facility’s total electricity consumption. But it would still be a huge step. The figures are modelled for a particular design and climate, with a reference-design proof of concept planned for Q1 2027. From our liquid-cooling sector research, we see how much time, energy and engineering effort is being committed to solving thermal issues. Accelsius reported B200 server benchmarks in which its two-phase system matched the single-phase baseline’s thermal performance with facility water at 54°C rather than 45°C. Warmer water gives operators more opportunity to reject heat without running energy-intensive chillers. Accelsius benchmark Part of APR’s approach is local coolant control. Management described directing coolant towards hotspots to reduce pumping energy and improve heat transfer at the chip interface. These designs target energy savings, thermal control and, depending on the system, opportunities to reduce cooling infrastructure and CAPEX. If you look at the route to market, there are interesting options for how these technologies could scale. In our call with APR, management said that even when large companies send their best engineers, they still need to teach them how to use and integrate the technology at a really deep level. Management stressed the need to get involved early in the hardware design, helping with simulations, coolant selection and materials. To me, that gives these niche players significant strategic value as partners to server manufacturers and chip designers. There’s a real specialist edge here that’s what we’ve gathered from multiple calls. So I see specialists teaming up with larger partners as a likely route, alongside potential M&A. Time to market is still something we’re researching. We have follow-up calls planned with these companies, but also with a few major players to look at it from their side. The combination of commercial scale and getting more compute out of the electricity a site can access is really interesting in these times. Lowering cooling overhead can leave more power available for running the servers. In my view, this sector is still very undercovered by analysts and people on X. $NVDA $NBIS $IREN $VRT