Loading page…
The rollout pairs new operating-software previews with cooling designs and power-management results. Its largest water-saving figure remains an estimate, while grid-response evidence comes from an early trial.
Listen to this story
At GTC Berlin on October 21, NVIDIA expanded DSX, its integrated data-center reference architecture, through European cloud operators and infrastructure partners, and previewed software for facility design and operations. The evidence is at different stages: NVIDIA estimates that dry cooling could save 54 million liters of water daily at a 2-GW facility, while Nscale reported gains in a four-rack evaluation and a separate grid-response trial used simulated events. For operators, DSX offers a coordinated approach to facility planning and power management, but results depend on hardware, climate and workload.
NVIDIA says Vera Rubin supports liquid cooling at coolant temperatures up to 45°C; dry coolers avoid evaporative water loss, though hot periods may require supplemental mechanical cooling.
DSX MaxLPS targets the same token output with up to 30% less power, or lets operators use their power budget for greater output.
In Iceland, Nscale reported 49% higher inference throughput and 37% more managed GPU capacity across four GB300 NVL72 racks within a 264-kW budget.
NVIDIA is bringing European partners onto a platform it says can run AI data centers with less electricity and almost no facility cooling water. Its October 21 announcement pairs DSX adoption with new software previews, but the evidence spans different stages: estimated water savings, partner power-management evaluations and an early trial using simulated grid events.
Announced at GTC Berlin, the rollout connects European AI cloud operators with infrastructure suppliers serving businesses, developers and governments. DSX is a reference architecture—a coordinated design for computing, networking, storage, software, power and cooling. NVIDIA’s pitch is to manage these parts together, from planning a facility to operating it, rather than optimize each in isolation.
The water claim rests on replacing evaporative cooling towers with closed-loop liquid cooling and outdoor dry coolers. Liquid absorbs heat from the computing equipment, then circulates to equipment resembling large radiators, which releases that heat into the air without evaporating water. Conventional cooling towers, by contrast, need replacement water as evaporation removes heat.
NVIDIA says its Vera Rubin infrastructure supports fully liquid-cooled operation with coolant temperatures up to 45 degrees Celsius. With dry coolers, it estimates a two-gigawatt deployment could save 54 million liters daily against cooling-tower-based systems. That is a facility cooling-water estimate, not a claim that the entire operation uses no water. Local climate also determines whether supplemental mechanical cooling is needed during the hottest periods.
DSX MaxLPS dynamically allocates power within an operator’s provisioned budget. NVIDIA says it can generate the same number of tokens—the units of text a model produces—with up to 30% less power. Operators can alternatively use the efficiency to support more GPUs and higher output. CoreWeave, Google Cloud, Lambda, Nebius and Nscale are among the companies exploring it.
The announcement cites Nscale’s evaluation across four GB300 NVL72 racks in Iceland: 49% higher inference throughput, or AI response output, and 37% more managed GPU capacity within a 264-kilowatt budget. Lambda’s proof of concept on HGX B200 systems reported a smaller gain. These are results from different hardware setups, not a head-to-head comparison.
DSX Flex addresses a different problem: adjusting workload power demand to utility needs, electricity prices and grid events. It can coordinate grid electricity, onsite renewables and storage. In an early trial with Emerald AI, EPRI, Nebius, the U.K. National Grid and NVIDIA, a Blackwell Ultra cluster reduced power draw during simulated grid events while critical workloads continued.
For design work, NVIDIA launched SimReady Central Library on Hugging Face. Its more than 2,000 free, ungated simulation-ready 3D assets include racks and cooling infrastructure. Unlike geometry alone, assets can carry properties needed for simulation, including electrical or thermal metadata. Developers must select assets whose supported capabilities match their particular workflow.
Nscale is using DSX Sim to build digital twins—virtual representations of facilities—for Portugal and Iceland, evaluating power, cooling and infrastructure configurations before deployment. For installed systems, NVIDIA is adding three components to its open-source DSX OS portfolio in developer preview:
Loading discussion...
Join the conversation
Explain which goal should take priority and why.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.