# The Thermodynamics of NVIDIA's 45°C Liquid Cooling > Source: > Published: 2026-06-24 22:04:37+00:00 [Cloud & Infra](https://www.devclubhouse.com/c/cloud)Article # The Thermodynamics of NVIDIA's 45°C Liquid Cooling High-temperature closed loops eliminate data center water consumption but expose a deeper dependency on upstream power grids. [Ji-ho Choi](https://www.devclubhouse.com/u/jiho_choi) The AI buildout is colliding with physical reality. As rack power densities climb past 100 kW, traditional air-cooling methods are no longer viable. The industry has long relied on evaporative cooling towers, which consume millions of gallons of water annually. [NVIDIA](https://www.nvidia.com) is attempting to break this cycle with its Rubin architecture and the associated DSX AI factory reference design. By introducing a 100% liquid-cooled system that operates with coolant temperatures up to 45°C (113°F), the design aims to cut facility water consumption to near zero. This is not just an incremental efficiency gain. It is a fundamental shift in thermodynamic design that moves the thermal bottleneck from the facility to the silicon. However, while it solves the immediate localized water crisis for data center operators, it also exposes a deeper, upstream sustainability paradox that platform engineers must reckon with. ## The Thermodynamics of 45°C Coolant Traditional data center design operated on a simple, expensive assumption: colder is better. For decades, server rooms were kept cold enough to require jackets, using energy-intensive chillers to push massive volumes of cooled air through complex hot and cold aisle configurations. The 45°C architecture flips this model. Silicon can operate safely at much higher temperatures than human comfort levels dictate. In the DSX reference design, a coolant mixture of 75% water and 25% propylene glycol enters the server rack at up to 45°C. It flows through microchannel cold plates mounted directly on the processors, networking chips, and power components, absorbing heat at the source. By the time the coolant exits the rack, its temperature has risen to approximately 55°C (131°F). This 10°C temperature delta is the key to the system's efficiency. Because the coolant is so warm, the temperature difference between the loop and the outdoor ambient air is highly favorable. This large delta allows the system to reject heat passively using outdoor dry coolers (essentially large radiator coils) without relying on mechanical refrigeration or evaporative cooling towers for the vast majority of the year. ``` php flowchart TD A[Outdoor Dry Cooler] -->|45°C Coolant| B[Coolant Distribution Unit] B -->|45°C Coolant| C[Microchannel Cold Plates on Chips] C -->|55°C Coolant (Heat Absorbed)| B B -->|55°C Coolant| A ``` The cold plates themselves require advanced manufacturing. To handle the extreme heat flux of chips operating at these densities, the cold plates use microchannels machined to micron-level precision. These channels are fabricated using techniques like precision milling, skiving, laser processing, etching, or 3D printing. Copper is the mandatory base material here, chosen over aluminum for its superior thermal conductivity. ## Eliminating Fans, Chillers, and Noise Transitioning to a 100% liquid-cooled architecture alters the physical footprint of the data center. Traditional facilities are notoriously loud, with cooling fans generating noise levels at or above 85 decibels, a threshold that requires ear protection for technicians. The Rubin platform eliminates server fans entirely. By removing the fans and the need for constant mechanical chilling, the energy overhead of the facility drops precipitously. Industry data shows that raising chiller plant temperatures by just one degree Celsius cuts cooling energy costs by roughly 4%. By running the loop at 45°C and bypassing chillers entirely in favorable climates, operators can reduce a facility's Power Usage Effectiveness (PUE) from a typical 1.35 down to 1.15. [Shadow GPS — know where it is, always Real-time GPS tracking for vehicles, gear and loved ones. No monthly contracts.](https://www.devclubhouse.com/go/ad/12) For a 50-megawatt hyperscale facility, this thermal efficiency translates to more than $4 million in annual savings on energy and water costs. More importantly, it reduces facility water consumption from approximately 2.6 million gallons per megawatt per year (for conventional cooling towers) to virtually zero. The loop is filled once at commissioning and runs closed for the life of the hardware. ## The Developer and Platform Angle: The Upstream Paradox For platform engineers and infrastructure architects, this shift introduces two major considerations: geographic flexibility and the upstream resource paradox. First, the elimination of evaporative cooling towers removes a major geographic constraint. Historically, building a high-density AI cluster required securing massive water rights, making deployments in arid regions like the US Southwest politically and environmentally fraught. With a 45°C dry-cooler design, you can deploy high-density Rubin NVL72 racks (which exceed 100 kW per cabinet) in warm climates without triggering local water crises. The system can operate chiller-free for most of the year, relying on minimal chiller intervention only during extreme peak summer days. However, solving the water problem at the facility level does not solve the water problem for the workload. This is the upstream paradox. While Scope 1 (direct) water consumption at the data center drops to near zero, the Scope 2 (indirect) water footprint remains heavily tied to the local power grid. Electricity generation is highly water-intensive. Fossil fuel power plants require substantial water for cooling, and even clean energy sources like hydropower lose massive volumes to reservoir evaporation. **Coal generation:** Consumes approximately 2.2 liters of water per kilowatt-hour.**Natural gas generation:** Consumes approximately 1.17 liters of water per kilowatt-hour.**Wind and solar:** Consume only a tiny fraction of these amounts over their lifecycles. If a "zero-water" data center is powered by a grid dominated by coal or natural gas, a heavy LLM training run still drives massive water consumption upstream at the power plant. To build truly sustainable infrastructure, platform engineers cannot rely on hardware cooling innovations alone. Workload orchestration must become grid-aware. Scheduling massive training jobs or batch inference pipelines must factor in not just the local PUE, but the real-time water and carbon intensity of the grid supplying the power. ## A Necessary Step, Not a Total Cure The transition to 100% liquid cooling is an engineering necessity. At power densities exceeding 100 kW per rack, air is simply no longer a viable medium for heat transfer. NVIDIA's partnership with cooling specialists like Motivair, a division of [Schneider Electric](https://www.se.com), to develop high-impedance coolant formulations shows that the supply chain is aligning rapidly behind this standard. Operating at 45°C is a elegant thermodynamic solution to the localized water and energy demands of AI factories. It proves that we can run high-performance silicon warm, eliminate mechanical chillers, and stop boiling away millions of gallons of local municipal water. But developers should not mistake a zero-water facility for a zero-water workload. The ultimate environmental cost of AI is still determined by the grid. ## Sources & further reading - [45°C cooling design cuts data center water use to near zero](https://blogs.nvidia.com/blog/liquid-cooling-ai-factories/)— blogs.nvidia.com - [Nvidia’s New Cooling System Cuts Data Center Water Use to Near Zero—But Not AI’s - MIT Sloan Management Review Middle East](https://www.mitsloanme.com/article/nvidias-new-cooling-system-cuts-data-center-water-use-to-near-zero-but-not-ais/)— mitsloanme.com - [Edgen](https://www.edgen.tech/news/post/nvidia-rubin-runs-45c-liquid-cooling-cutting-water-use-to-near-zero)— edgen.tech - [Can AI data centres run without water? Nvidia introduces near-zero-water cooling technology](https://www.wionews.com/photos/can-ai-data-centres-run-without-water-nvidia-introduces-near-zero-water-cooling-technology-1782220113124)— wionews.com [Ji-ho Choi](https://www.devclubhouse.com/u/jiho_choi)· Security & Cloud Editor Ji-ho covers the increasingly tangled overlap between cloud architecture and security, drawing on a background as a penetration tester to keep his reporting grounded in real-world attack paths. He never lets a vendor claim go unquestioned and insists that every buzzword come with a proof of concept. ## Discussion 0 No comments yet Be the first to weigh in.