Nvidia's 45°C Warm-Water Cooling Drops AI Datacenter Water Use to Near Zero

Chips are no longer the only thing gating AI datacenters. Water and power have become the real constraint. Nvidia’s answer, shown at Climate Week on June 21-22, is a design that barely sends any coolant down the drain.

What Happened

Nvidia unveiled the DSX AI factory reference design for its next-generation Rubin systems, calling it the world’s first 100% liquid-cooled datacenter design. The trick is temperature and a closed loop. The coolant is a mix of 75% water and 25% propylene glycol, filled once and recirculated for the life of the facility. It runs at 45°C — hotter than a hot tub — and that heat is warm enough to reject without a mechanical chiller in the right climate. The payoff shows up in water. A cooling-tower setup evaporates roughly 2.6 million gallons of water per megawatt per year; this design drops that to near zero on-site. Warm-cooling reference kits ship to partners in Q4 2026, with the first hyperscale deployments slated for mid-2027. A 50 MW pilot is already running in Finland, tied into Fortum’s district-heating network, where the datacenter’s waste heat warms more than 20,000 homes in Espoo. One honest caveat: this cuts on-site datacenter water, but it does not erase AI’s broader water and energy footprint, including the water and power consumed by electricity generation upstream.

What This Means for Founders

The larger signal here is that the bottleneck in AI infrastructure has moved from chips to siting. The right to use water and the ability to pull enough power now decide where a datacenter can even be built. For anyone scaling a compute-dependent product, that lands two ways. The first is cost. When cloud providers spend more on cooling in water-stressed regions, that expense flows down into GPU rental rates and inference pricing. Conversely, where chiller-free warm cooling spreads, operating costs in some locations have room to fall. Either way, AI unit economics no longer end at a model price list — they are tied to where the datacenter physically stands. The second is waste heat as a new resource. The moment a facility sells its heat into district heating, as in Finland, a datacenter shifts from cost center to energy supplier. In the US, siting in water-stressed regions and grid-interconnection queues are already slowing buildout. Business models that route waste heat into nearby residential or industrial heating, technologies for coolant recirculation and water treatment, and siting advisory aimed at cold or water-rich regions are all openings this shift creates.

What You Can Do Now

If you run an AI product, fold cloud-region water and power constraints into how you price inference. The same GPU carries different cost and reliability depending on which region’s datacenter it sits in. If you work in infrastructure or hardware, map now where the technologies that ease siting constraints — waste-heat reuse, coolant recirculation, water treatment — actually become a business. And don’t copy the “zero water” headline straight into your marketing. On-site water falls, but the upstream and generation footprint remains. If you lean on sustainability, drawing an honest line around which stage you’re cutting holds up far longer under regulatory and investor scrutiny.