Summary

• Nvidia's DSX reference architecture combines liquid cooling with dry coolers, targeting near-zero facility cooling water consumption under suitable conditions and modelled savings at very large sites.
• DSX MaxLPS is designed to reduce power consumption by up to 30% for equivalent AI output. Early Lambda and Nscale evaluations report improved throughput and GPU capacity within fixed electrical budgets.
• DSX Flex can adjust selected workloads in response to grid conditions. Its reported 40% reduction in cluster power draw was demonstrated during simulated events, not as a permanent facility-wide saving.

Nvidia is expanding its AI data centre platform with technologies designed to reduce cooling water consumption, improve the amount of computing delivered within available electricity capacity and allow selected workloads to respond to power grid conditions.

The company’s DSX platform combines liquid cooling and dry cooling reference designs with software for managing computing power, modelling infrastructure and coordinating the electricity used by large AI installations.

At GTC Berlin on 21 October, Nvidia reported early evaluations involving European infrastructure developers, including Nscale, alongside results from computing provider Lambda and a grid-response demonstration involving British electricity system participants.

Nvidia says its architecture could reduce facility cooling water consumption to near zero in suitable installations and allow the same number of AI output tokens to be generated using up to 30% less power. The figures are conditional technology claims, not independently verified savings achieved across an installed European data centre fleet.

Dry cooling changes the water requirements

High-density computing equipment produces substantial heat, which must be removed continuously to keep processors, memory and supporting systems within acceptable operating temperatures.

Some conventional data centre cooling systems use towers that reject heat through evaporation. This can be effective, but it requires replacement water as part of the cooling process.

Nvidia’s DSX reference architecture combines liquid-cooled computing equipment with outdoor dry coolers. Liquid circulates through a closed system, carrying heat from the computing equipment to infrastructure that releases it into the surrounding air without relying primarily on evaporation.

The architecture uses Vera Rubin computing systems designed for liquid cooling with coolant temperatures reaching 45°C, allowing heat rejection at temperatures compatible with the proposed dry cooling configuration.

Under appropriate conditions, the reference design is intended to bring direct facility cooling water consumption close to zero. Nvidia estimates that a hypothetical two-gigawatt installation could save 54 million litres each day compared with conventional cooling tower infrastructure.

That is a modelled comparison for a very large specified configuration, not a measurement from a completed two-gigawatt facility using the design.

Nor does reducing the use of water within a cooling system eliminate the wider water footprint associated with computing. Electricity generation, equipment manufacturing and other supporting activities can consume water outside the data centre boundary.

Local climate is another important variable. Nvidia acknowledges that supplemental mechanical cooling may be necessary during the hottest periods, affecting the cost and performance of an installation.

Operators must consequently evaluate dry cooling against the temperatures experienced at their site, the availability and price of water, and the electricity required by alternative cooling equipment.

Computing output becomes a power management target

Electricity connections place a physical limit on the capacity that data centre operators can deploy. An installation may have sufficient floor space for additional equipment but lack the power supply needed to operate it.

DSX MaxLPS is designed to manage the power allocated to computing equipment dynamically, increasing useful output within a provisioned electrical budget.

Nvidia says the software can generate an equivalent number of AI tokens while consuming up to 30% less power. Operators could alternatively use efficiency improvements to support more computing capacity without increasing their allocated electricity supply.

Those outcomes are related but different. Reducing energy use while maintaining output could lower operating expenditure, whereas using the same electricity to produce more output may increase revenue opportunities without reducing total power consumption.

The appropriate choice depends on customer demand, the price of computing services, electricity contracts and the financial characteristics of the hardware installed.

Lambda has completed a proof of concept using Nvidia HGX B200 systems and DSX MaxLPS. Nvidia reports 24% higher cluster token throughput and 19% more supported GPU capacity within a fixed power budget.

The results concern the evaluated configuration rather than a guaranteed improvement for every AI workload or computing cluster.

Nscale has separately tested MaxLPS at its Icelandic AI facility, using four Nvidia GB300 NVL72 racks.

Within an unchanged provisioned power budget of 264 kilowatts, the company reported 49% higher inference throughput and 37% more managed GPU capacity.

Those measurements provide more specific evidence than the broader headline efficiency claim, although the results remain company-reported and tied to the hardware and workloads used in the evaluation.

Other infrastructure operators, including CoreWeave, Google Cloud, Nebius and IREN, are exploring the software. Their involvement should not be described as proof that each has completed an equivalent production deployment or achieved the same performance gains.

AI workloads begin responding to grid conditions

DSX Flex addresses the relationship between computing operations and the wider electricity system.

Some workloads require continuous processing, while others can be scheduled, delayed or operated with reduced performance for limited periods without undermining essential customer services.

The software is designed to use that distinction when electricity prices change or grid operators request temporary reductions in power consumption.

It can also coordinate consumption with on-site renewable generation and energy storage, potentially allowing operators to adjust how much electricity they draw from the network.

An early demonstration involving Emerald AI, EPRI, Nebius, Nvidia and Britain’s National Grid reduced cluster power draw by up to 40% during simulated grid events while maintaining critical workloads.

That result does not establish that a data centre can permanently operate using 40% less electricity. It demonstrates temporary flexibility under defined conditions, with selected workloads continuing to run.

Commercial deployment would require agreements defining when reductions can be requested, the power response available and how operators are compensated for providing flexibility.

Power grid operators would also need confidence that the expected response can be delivered reliably without creating unexpected shifts in demand elsewhere.

Software flexibility may help electricity networks accommodate some new computing capacity, but it cannot automatically provide a connection where suitable grid infrastructure is unavailable.

Techopia previously examined electricity delays affecting Microsoft’s planned Nscale development in Essex. The DSX evaluation in Iceland is a separate deployment and does not resolve the connection problems associated with that British project.

Simulation extends the platform into facility design

Nvidia is also expanding DSX Sim, which allows engineers to evaluate computing equipment alongside cooling, electrical and other supporting infrastructure before installation.

Large AI facilities require closely coordinated designs because the placement and operating characteristics of computing equipment affect both electricity distribution and heat removal.

Changes to rack density or cooling equipment can alter the requirements placed on other systems, making design coordination important before expensive components arrive on site.

Digital representations can help engineers identify conflicts and compare alternatives during planning, although the usefulness of a simulation depends on the accuracy of its equipment models and operating assumptions.

Nscale is applying DSX Sim to facilities in Portugal and Iceland, while Nvidia is introducing SimReady Central Library with equipment assets contributed by industrial technology partners.

These models are intended to represent relevant characteristics of power and cooling equipment, allowing developers to compare infrastructure configurations before physical construction.

Actual performance must still be verified after installation, particularly where temperatures, equipment behaviour or other conditions differ from the assumptions used in the model.

Infrastructure management remains under development

DSX OS adds tools for configuring network equipment, managing racks and evaluating the readiness of computing clusters before they begin production workloads.

Nvidia has introduced three additional components in developer preview: Config Manager, Rack Management Service and Cluster Readiness Engine.

The tools are intended to automate selected infrastructure tasks and identify faults before customers begin using the equipment. However, developer preview status means the new components should not be treated as generally available production services across every DSX installation.

The broader platform also depends on industrial suppliers providing electrical, thermal and mechanical infrastructure suitable for high-density computing.

Companies including Vertiv, Trane Technologies, Hitachi Energy, Siemens Energy and GE Vernova are participating in aspects of the DSX ecosystem, from cooling systems to electricity management and digital modelling.

Their involvement does not establish that every participating organisation has implemented the entire platform or that its equipment has achieved the headline efficiency figures in commercial service.

Nvidia’s latest results show different kinds of progress: modelled cooling water reductions, computing performance measured in specific evaluations and temporary flexibility demonstrated through simulated grid events.

European infrastructure developers must now establish whether those capabilities can be combined economically within individual sites, accounting for electricity supply, cooling requirements, customer demand and the cost of increasingly powerful computing equipment.

The early results suggest opportunities to improve output from constrained infrastructure, but wider adoption will depend on independent validation and operational evidence rather than applying the strongest individual results to every future AI factory.