WorldIndustrialWeekly Energy Desk: What the NVIDIA DSX Cooling-Water and Power Numbers Mean for Industrial Sites

Ask a plant energy manager what limits a compute expansion and the answer is rarely floor space. It is the incoming power envelope, the make-up water a heat-rejection circuit consumes, and the utility's tolerance for a load that behaves unpredictably. The build-out now underway around the NVIDIA DSX AI Factory Platform in Europe lands on precisely those three constraints, which is why it deserves an engineering read rather than a press-release skim. This Energy Desk breakdown maps the published architecture, trial results and equipment qualifications onto the decisions a factory, campus or dedicated compute site actually has to make.

Three energy constraints addressed in one program

Europe’s leading ecosystem partners are taking up the platform to build AI factories that reduce cooling water use and consume less power for the same AI output. The same program also supports faster bring-up of new capacity and lets operators adapt workload power consumption to support grid reliability. Translated into utility terms, that is one water line, one demand line and one interconnection line — the three meters that a facilities or energy team is measured on.

Codesigned as one system, not handed off after the spec

As a full-stack AI factory reference architecture, NVIDIA DSX was co-innovated with leading infrastructure companies worldwide. Compute, networking, storage, software, power and cooling are codesigned as one system, optimized from design to operation. AI clouds, together with leading design, construction, energy and cooling providers, are using it to build and operate infrastructure that serves Europe’s enterprises, developers and governments. The practical consequence for industrial readers: facility engineering is inside the reference architecture rather than downstream of it.

Module map: where NVIDIA DSX platform touches the energy chain

Module	Energy-side function	What the operator gains
DSX MaxLPS	Dynamic power allocation inside a provisioned budget	The same number of tokens with up to 30% less power
DSX Flex	Coordinates workload power consumption with grid conditions	Response to utility needs, demand-response events and energy prices
DSX Reference Design	Validated infrastructure architectures	A defensible starting point for power and cooling engineering
DSX Sim	Simulation of power, cooling and infrastructure configurations	Virtual validation before equipment is installed
DSX OS	Infrastructure management	Lifecycle automation and consistent runtime environments
DSX Exchange	Connects computing and facility systems	No air gap between the IT domain and the utility/domain side

Within the NVIDIA DSX platform, these technologies are described as helping operators design, deploy and run AI factories that save valuable resources. The value for an energy team is that each layer has a meterable outcome attached to it.

Cooling retrofit: replace evaporation with a closed loop and dry heat rejection

Conventional cooling towers remove heat by evaporating water and require a continual supply of replacement water. DSX reference designs instead connect liquid-cooled computing with dry-cooler-based heat rejection that avoids evaporation: a closed loop lets cooling liquid absorb heat from computing equipment, while outdoor dry coolers — effectively large radiators — release that heat into the air without evaporating water.

NVIDIA Vera Rubin infrastructure supports 100% liquid cooling with coolant temperatures of up to 45 degrees Celsius. Dry-cooler-based designs can reduce facility cooling water consumption to near zero. At 2-gigawatt scale, NVIDIA DSX can save 54 million liters of water a day compared with conventional cooling-tower-based systems, equivalent to the daily water use of 180,000 households.

Read in sequence, the engineering logic is: a higher allowed coolant temperature makes dry heat rejection viable, eliminating evaporation turns make-up water into a marginal line item, and the water-scarcity risk that often gates a siting approval is removed. Whether supplemental mechanical cooling is needed during the hottest periods depends on local climate — a caveat that belongs in any feasibility model.

Power allocation: capacity is defined by what you are allowed to draw

How much capacity operators can bring online is defined by power availability. DSX MaxLPS and DSX Flex address this constraint. MaxLPS dynamically allocates power within an AI factory’s provisioned budget, helping operators generate the same number of tokens with up to 30% less power. That efficiency can also support more GPU capacity and higher throughput within existing power budgets. The software is being explored by CoreWeave, Firmus, Google Cloud, IREN, Lambda, Nebius, Nscale and Sharon AI.

Trial and deployment data laid out for comparison

Evidence point	Hardware or scale	Reported result	What was held fixed
Lambda proof of concept	NVIDIA HGX B200 systems with DSX MaxLPS	24% more cluster token throughput and 19% more GPU capacity	A fixed power budget
Nscale evaluation run (Iceland)	four NVIDIA GB300 NVL72 racks with DSX MaxLPS	49% higher inference throughput and 37% more managed GPU capacity	The same 264-kilowatt provisioned power budget
Emerald AI, EPRI, Nebius and U.K. National Grid early trial	An NVIDIA Blackwell Ultra GPU cluster	Reductions of up to 40% in the cluster’s power draw in response to simulated grid events	Critical workloads continued running
Water use at scale	2-gigawatt scale versus conventional cooling-tower-based systems	54 million liters of water a day saved, equivalent to 180,000 households’ daily use	Same AI output

Case file: Nscale turns a fixed envelope into headroom

Through DSX MaxLPS, DSX Flex and DSX Sim, Nscale is demonstrating how the platform connects power optimization, grid-responsive operations and design. An evaluation run in an Nscale AI factory in Iceland used DSX MaxLPS across four NVIDIA GB300 NVL72 racks and showed 49% higher inference throughput and 37% more managed GPU capacity within the same 264-kilowatt provisioned power budget — evidence that dynamic power allocation can unlock additional capacity from infrastructure already installed.

Nscale is also exploring DSX Flex to align workload power consumption with grid conditions. For facilities in Portugal and Iceland, it is using DSX Sim to create digital twins that help evaluate power, cooling and infrastructure configurations before deployment. Virtual validation lets the operator identify potential issues earlier, reduce deployment risk and accelerate the path from planning to productive capacity.

Grid interaction: the load becomes a controllable asset

By helping AI factories become flexible assets for the power grid, DSX Flex enables operators to adapt workload power consumption to utility needs, demand-response events and energy prices. It can also coordinate electricity from the grid, onsite renewables and energy storage, so sites can operate as flexible power loads that support grid reliability and enable faster grid connection. For industrial owners dealing with interconnection queues and volatile tariffs, that framing — load as a dispatchable asset — is the part worth testing first.

Validate before you break ground

SimReady Central Library has been launched, including validated power and cooling equipment assets from partners like Vertiv, and accelerates the creation of DSX AI factory digital twins. When used with DSX Sim, ecosystem partners including PTC are helping create simulation-ready assets with computer-aided design. These assets enable SimReady workflows that let builders model infrastructure configurations and evaluate power and cooling decisions before equipment is installed, so operators can validate designs earlier, reduce deployment risk and move faster from planning to construction.

Lifecycle control: the NVIDIA DSX OS production layer

The modular, open source NVIDIA DSX OS portfolio is expanding with three components in developer preview to help operators turn installed infrastructure into productive capacity:

• NVIDIA Config Manager automates network configuration for AI factories.
• NVIDIA Rack Management Service provides rack-level management of power, inventory, firmware and switches.
• NVIDIA Cluster Readiness Engine runs distributed workloads before production workloads arrive to validate GPU clusters and isolate failures to specific nodes.

These components, together with NVIDIA Infrastructure Controller, NVIDIA AI Cluster Runtime and NVSentinel, help operators automate infrastructure lifecycle management, maintain consistent runtime environments and strengthen health validation, fault attribution and recovery. Aleria, Cassava, CoreWeave, Crusoe, Firmus, Firebird, G42, GMI Cloud, IREN, Lambda, Nebius, Nscale, Radiant, Sharon AI and Volta are cloud partners exploring the latest components to automate infrastructure lifecycle management and improve resiliency and reliability.

The vendor stack that makes the engineering executable

Industrial software, power and cooling partners are advancing Europe’s buildout through DSX reference designs, lifecycle digital twins and qualified equipment that connect facility infrastructure with accelerated computing. For NVIDIA Vera Rubin infrastructure, Schneider Electric has developed power and cooling reference designs and is working with AVEVA on digital twins spanning infrastructure design, construction and operations. With its construction management platform integrated with DSX Sim, Procore is helping manage builds.

Siemens is using DSX Sim to simulate power, cooling and acoustics. Dassault Systèmes and Vertiv are developing digital twin models of Vertiv’s OneCore infrastructure, enabling simulation and optimization through model-based systems engineering. ABB is developing SimReady assets with NVIDIA DSX Sim, letting engineers evaluate source-to-rack architectures before construction while advancing 800 VDC development through its newly announced Infinitus Direct Current portfolio.

Trane Technologies and Eaton recently collaborated with NVIDIA to help boost data center efficiency and cut installation costs with an integrated design built for the NVIDIA Vera Rubin DSX AI Factory Reference Design. The design brings together Eaton’s Beam platform and power infrastructure with Trane Continuum cooling and coordinated controls in a grid-to-chip architecture. To support thermal infrastructure planning at scale, Trane Technologies has published seven thermal management system reference designs, including three 250-megawatt DSX-based cooling reference designs.

Qualified equipment at a glance

Vendor and product	Category	Qualification status
Trane Technologies LiquidStack GigaModular	Coolant distribution unit	Qualified under its applicable NVIDIA DSX Ready category
Vertiv 2.3-megawatt CoolChip	Coolant distribution unit	Qualified under its applicable NVIDIA DSX Ready category
Hitachi Energy battery energy storage system	Storage	NVIDIA DSX Ready BESS qualification

On the supply and grid side, GE Vernova, Hitachi Energy and Siemens Energy are collaborating with NVIDIA on power infrastructure and digital twin initiatives for energy supply and grid integration, connecting energy-system engineering and grid planning based on the facility’s requirements. Hitachi Energy is developing an integration between its energy management system and DSX Flex to optimize grid, generation and storage resources with workload demand.

Five misreads industrial teams should avoid

1. Treating cooling water as an operating expense rather than a siting constraint. Evaporative towers need a continual supply of replacement water; the closed-loop, dry-cooler approach removes that dependency to near zero.
2. Assuming dry heat rejection is universally sufficient. Whether supplemental mechanical cooling is needed during the hottest periods depends on local climate, so the assumption must be tested per site.
3. Ordering more hardware before auditing the power envelope. Capacity is defined by power availability, yet the same AI output is reported with up to 30% less power under dynamic allocation.
4. Skipping virtual validation. Digital twins built with DSX Sim let power, cooling and infrastructure configurations be evaluated before deployment, which is where cost and schedule risk are normally created.
5. Running the facility and the workload as separate systems. DSX Exchange connects computing and facility systems, and DSX Flex ties consumption to utility needs — siloing them forfeits both effects.

A staged sequence for an energy-led implementation

1. Fix the baseline: document the provisioned power budget, make-up water consumption and any demand-response obligations before any equipment decision.
2. Adopt a validated architecture: start from a DSX Reference Design so compute, networking, storage, software, power and cooling are engineered as one system.
3. Rework heat rejection: move to 100% liquid cooling with dry-cooler-based rejection where coolant temperatures of up to 45 degrees Celsius allow it, and model local-climate exceptions.
4. Enable dynamic allocation: run the existing fleet under DSX MaxLPS to measure throughput and capacity recovered within the same budget.
5. Test grid responsiveness: pilot DSX Flex against utility signals, onsite renewables and storage coordination before committing new load.
6. Simulate the next build: use DSX Sim with SimReady assets to model power, cooling, acoustics and infrastructure configurations prior to construction.
7. Standardize lifecycle control: deploy the NVIDIA DSX OS components for configuration, rack management and cluster readiness, and select equipment qualified under applicable DSX categories.

Bottom line for plant and facility owners

The industrial relevance of this platform is not that it makes accelerators faster; it is that it changes what a site can promise to its utility, its water supplier and its own maintenance schedule. More throughput and more GPU capacity inside a fixed budget, near-zero cooling water at the design level, and a documented ability to shed load during grid events are all claims a plant energy team can put into a business case. The sensible first move is a measured one: baseline the site, simulate it, and prove the allocation and response behavior on installed equipment before any new capacity is ordered.