NVIDIA DSX: The Blueprint Turning Every Megawatt Into Tokens With 45°C Direct-to-Chip Liquid Cooling

What separates an AI factory from a traditional data center is that the facility itself becomes the product. Both the report and the five videos repeat the same line: if design, grid, power and cooling are not co-engineered, every megawatt is wasted. Revenue is now measured in tokens per watt, so any idle watt is lost revenue. NVIDIA DSX steps in to unify design, simulation, deployment and operations in one intelligent system.

DSX is not a single product but a platform and a blueprint. Built around the Vera Rubin generation, its reference design treats every layer from chip to grid as one stack. With an Omniverse digital twin, layout, power and cooling are simulated before a single rack lands and bugs are caught on screen. When the site powers on, DSX OS takes over to provision, monitor and self-heal the infrastructure, turning components from many vendors into one trusted multi-tenant capacity.

At its heart is the idea that power is the production constraint, embodied by DSX MaxLPS. Many sites over-provision for peak and leave up to 40% of capacity stranded. MaxLPS claims to dynamically manage power across GPUs, racks and workloads to deliver up to 40% more compute inside the same budget. Stranded watts are reclaimed rack to rack and in-rack smoothing flattens current spikes. Planning options MaxP, MaxQ and MaxLPS produce different GPU counts per fixed 1 MW, with MaxLPS using effective average utilized power.

The nervous system for that orchestration is DSX Exchange. It bridges IT and OT as an open event bus built on NATS, speaking MQTT 3.1.1, persisting with JetStream and federating via leaf nodes. A three-node HA cluster, topic-level ACLs with OAuth2/mTLS/NKey and Helm/ArgoCD declarative deployment let facility signals speak the same language. Then agents like Phaedrus can balance cooling and electrical systems in real time.

Grid flexibility lives in DSX Flex. Grid signals for shedding, demand response and price events are read live and interpreted by Emerald AI agents that throttle or defer workloads. The same layer orchestrates hybrid energy across grid, on-site renewables and storage. Instead of a one-way sink, the AI factory becomes a flexible asset that relieves the grid. The videos note 100 gigawatts of new AI factory capacity arriving before decade's end; without flexibility that scale would strain grids.

The design-validation chain is too complex for a single tool. SimReady assets are governed in PTC Windchill PLM, model-based systems engineering is done in Dassault Systèmes, Jacobs finalizes the design in its custom Omniverse app. External thermals are simulated in Siemens Star-CCM+, internal thermals in Cadence Reality, electrical in ETAP and networking in NVIDIA's DSX Air. Virtual commissioning through Procore then compresses construction schedules and reduces site surprises.

Another lever in power delivery is 800 VDC. High-voltage DC carries more power with less copper loss and eases scaling. Combined with MaxLPS, the site's fixed power envelope is used more efficiently and the need to re-cable for each GPU generation shrinks. Briefly mentioned in the report, this detail matters for Rubin-class dense racks.

The cooling breakthrough is direct-to-chip liquid cooling at 45°C. A copper cold plate sits on the die; coolant entering at 45°C absorbs heat at source and exits near 55°C. The coolant is a closed loop of 75% water and 25% propylene glycol, pumped from a coolant distribution unit through servers and back. Earlier hybrids liquid-cooled only GPUs and CPUs while the rest stayed air-cooled; the Rubin generation is described as the first fully liquid-cooled design with no fans anywhere.

The magic of 45°C is removing chillers. Conventional air cooling relies on chillers and cooling towers that burn energy and water on hot days. When inlet water tolerates 45°C, many climates can run chiller-less with dry coolers. NVIDIA's blog notes that in favorable climates this can cut facility cooling water from about 2.6 million gallons per megawatt per year with tower-based systems to near zero. Cooling with water hotter than a hot tub sounds counterintuitive, yet silicon stays within validated limits inside the cold plate.

The report's comparison table makes the gap concrete. Air cooling sits at PUE 1.5 to 1.8 while 45°C liquid drops toward 1.05. Density is 15 to 20 kW per rack for air versus 120 kW plus for liquid. Water and environmental impact diverges as well: open evaporation towers versus a closed loop with near-zero loss. Read together, liquid cooling cuts not only the energy bill but also land and water bills.

Scalability and test culture are DSX's other hardware differentiators. The Enterprise Reference Architecture is built on a four-node scalable unit, built and end-to-end validated in NVIDIA labs, predictably growing from four to eight, twelve and sixteen nodes. OEM partners Cisco, Dell, HPE, Lenovo and Supermicro pass their offerings through a design review board. Customers therefore buy not a bill of materials but a known-good architecture.

These architectures come in three flavors. RTX Pro AI Factory for universal acceleration at the entry, HGX for high-performance AI in the middle, NVL72 for gigascale training and frontier models at the top. All rest on NVIDIA Certified Systems and Spectrum-X Ethernet with NVLink and scalable network topologies. An enterprise can start in one tier and expand to another while network and operational assumptions stay consistent, lowering risk.

On top of hardware, the Enterprise Validated Design completes the stack. In Jensen's five-layer cake, the bottom three layers are energy-chip-infrastructure, the top layers are models and applications; reference architectures cover the bottom, validated designs the top. The lifecycle is framed as a data flywheel: data selection, curation, training and fine-tuning, agent skills, deployment and observability, then feeding production traces back into the next development cycle. Each bucket hosts jointly tested software from NVIDIA and independent vendors.

NVIDIA's own factory is presented as live proof. Internally, about forty thousand employees drive four trillion tokens per month and two hundred million inference requests per day at near 99.9% availability, with internal demand growing 40% month over month. ChipNeMo, used daily by almost all five thousand hardware engineers and evolving for three years, stands out. In IT, ticket deflection and self-service scaled on 24/7 virtual desktops where agents run autonomously; a system first sized for a thousand users burst to twenty-five thousand requests as agent usage exploded.

Ecosystem scale shows why DSX is more than paperwork. Foxconn, Schneider Electric, Vertiv, Bechtel and Caterpillar build physical infrastructure to the DSX spec; PTC, Dassault, Siemens, Cadence and ETAP cover simulation, Procore covers field execution. The IREN example, a multi-layer partnership across a 5 gigawatt global portfolio starting with a 60 megawatt cloud deal at Childress, grounds the strategy. At this scale each month of delay is framed as billions in lost revenue.

Looking ahead, two themes dominate. One is hybrid cloud plus on-prem capacity planning; with 6 to 9 month procurement cycles a cloud buffer is critical to meet demand. The other is confidential computing to bring frontier models on-prem securely, with encrypted and provably protected weights while enterprise data stays inside its perimeter. Together they will decide tokenomics for regulated industries worried about compliance and data mobility.