DSX Facilities Infrastructure Reference Design Overview

This table presents example site figures for a 250 MW-class IT load design with 96 Scalable Units. It is a sizing reference independent of the rendered campus image, not a description of the pictured site. Actual site layouts and capacities vary by project.

GPU rack count assumes a maximum of eight racks per compute row.

DSX design value: Creates a scalable power backbone that aligns campus expansion with available utility capacity, bringing grid requirements, BESS functions, and long-lead equipment into the design before they become schedule or operating risks.

Two buses in the same 34.5 kV air-insulated switchgear (AIS) or gas-insulated switchgear (GIS) distribute power across the site. The Service Substation lands utility lines above 100 kV and transforms them through HV-to-34.5 kV transformers into the campus medium-voltage backbone.

This reference pattern separates Network Core and GPU Compute paths for different availability objectives. On-premises generation, if used, and BESS share the same 34.5 kV bus, making the substation the integration point for flexible power and DSX Flex grid services.

• Utility in: 100 kV+ transmission.
• Switchgear: 34.5 kV, 2,500 A, 40 kA.
• Includes HV breaker scheme plus HV/34.5 kV transformers, two buses for Core and GPU Compute, integration for utility, on-premises generation, and BESS, plus utility grid operator coordination through DSX Flex.

Twin engineered basins on the east and west of the campus capture, slow, and treat rainwater runoff.

The two detention basins are positioned just outside the building perimeter on the east and west edges of the site. Their job is to buffer storm flow from an impervious-heavy campus, detain that water, allow solids to settle, and discharge slowly at a metered rate that approximates pre-development runoff.

See the 250 MW / 96 SU Example Site Figures for the example stormwater figure.

• Quantity: two basins.
• Permitting: EPA NPDES plus state stormwater.
• Cooling-tower blowdown: none in this design.

DSX design value: NVIDIA’s 45°C liquid-cooling design point expands the operating window for rejecting campus heat without full mechanical chilling, leaving more of the facility power budget for AI compute.

Dry coolers are positioned at the far east and west of the site, directly north of each Central Utility Building. Warm facility water from the chillers flows through finned-tube heat exchangers; fans pull ambient air across the fins, dumping heat to the air with no evaporative water use.

The CUB facility-water loop serves mechanical gallery CDUs and CRAHs, while the paired dry coolers reject its heat to the atmosphere.

Dry coolers run on 480 V alongside the CUB pumps and controls. On the GPU side, the cooling plant is not UPS-backed; during utility outages it relies on chilled-water loop thermal mass until backup generation comes online and the system auto-restarts.

• Function: air-side heat rejection.
• Evaporative water use: none.
• Bus: 480 V.
• UPS: Core CUB only.

DSX design value: Adds fast-response power flexibility around dynamic AI loads: BESS stabilizes load behavior and bridges grid events, while standby generation supports continuity for critical Core systems.

AI loads are dynamic by nature, while the grid and generators are built to serve steady loads. In this reference campus, BESS sits on the 34.5 kV switchgear bus as a controllable buffer between them.

Applicability and sizing are evaluated case-by-case with OEMs, grid operators, and regulators based on the site power architecture, operating objectives, grid requirements, and any product-specific requirements.

Central standby generators provide backup power through an N+2 configuration at the 34.5 kV switchgear level, supporting five-nines (99.999%) reliability for critical Core network and mechanical loads.

• BESS functions: facility-level buffer, abnormal-event response, bridge power, and grid dispatchability.
• BESS applicability and sizing: deployment-specific and evaluated case-by-case.
• Shared connection: 34.5 kV MSG.
• Backup generation: N+2.
• Reliability supported: five-nines (99.999%).
• Core DSX Technologies: DSX Flex and DSX MaxLPS.

DSX design value: Centralizes dual-temperature cooling for liquid- and air-cooled systems in a shared plant, which can reduce equipment and cooling overhead while improving site acoustics and maintenance conditions.

Two CUBs at the east and west ends of the building house chillers, pumps, and facility-water distribution serving CDUs and CRAHs in the mechanical galleries. Paired outdoor dry coolers reject heat at both ends.

The CUB is fed from dedicated MV loops. Chillers run on a decoupled 4.16 kV bus while the rest of the CUB, including pumps, coolers, and controls, runs on 480 V. UPS is provided only for the Core portion of the CUB.

See the 250 MW / 96 SU Example Site Figures for reference-scale Compute and campus load figures.

• Chiller bus: 4.16 kV.
• CUB / coolers bus: 480 V.
• UPS: Core CUB only.

DSX design value: Places redundant CDU groups in mechanical galleries to reduce equipment count and support concurrent maintenance, while rack-level control and isolation limit cooling-fault impact and support tighter data-hall access control.

This representative pin marks cooling equipment inside the mechanical gallery, not exterior site equipment. Larger shared piping arrangements reduce the number of liquid-to-liquid CDUs while N+1 group operation supports concurrent maintenance.

CDUs receive high-temperature facility water from the CUB and use it to cool a separate, technical-grade secondary loop serving GPU cold plates.

CRAHs receive medium-temperature facility water from the CUB and provide air cooling for the data halls and other occupied spaces.

The validated CDU partner list is dynamic. See the live NVIDIA Enterprise Marketplace catalog for current MP-Ready and Sample-Ready products.

• DSX form factor: liquid-to-liquid.
• TCS design flow: at least 1.5 LPM/kW.
• Validation: CDU self-qualification suite.
• CDU redundancy: N+1.

DSX design value: Standardizes power, cooling, networking, and rack arrangements as scalable AI-capacity building blocks that can evolve across campus phases and GPU generations, simplifying design, procurement, and construction.

A Scalable Unit is the smallest repeatable AI factory building block: 1 Compute Hot Aisle Containment (HAC) plus 1 Support HAC, with a consistent power budget per SU.

In the reference pattern, each data hall contains 18 Scalable Units, or 24 in the NVIDIA DSX MaxLPS design.

Cabinet TDP scales from MGX Gen 1.1 at 198 kW to Vera Rubin NVL72 at 330 kW.

See the 250 MW / 96 SU Example Site Figures for the example compute-hall, GPU-rack, and GPU counts.

• Cabinet TDP: 198 kW to 330 kW.
• Air cooling for remaining equipment is provided via CRAHs in the Mechanical Gallery.

DSX design value: Centralizes shared network services to keep fiber paths within design limits, simplify backup-power distribution, and support phased construction without disturbing installed cabling.

The Network Core and Shared Services areas sit between the Compute Data Halls on both sides of the CIN Spine. They connect into the CIN Spine and house the Tenant Access Network, Secure Management Network (SMN), High-Speed Storage, Control Nodes, Meet-Me Rooms, and Fiber Entrance Rooms.

See the 250 MW / 96 SU Example Site Figures for reference-scale Core and campus load figures.

• Network relationship: CIN Spine, TAN, SMN, HSS.

DSX design value: Centralizes the CIN spine so separate data halls can operate as one compute system within optical limits, with direct fiber paths preserving reach and supporting phased cluster growth.

The Cluster Interconnect Network spine is the campus’s east-west GPU fabric, connecting every Compute Hall into one cluster. Its central location and direct fiber routes keep connections within the 500 m CIN distance limit.

The base single-story form factor is optimized for speed and cost-efficient development. Where CIN fiber runs would exceed 500 m, the modular building components can be arranged in an H-shaped form to keep links within the distance limit.

Diverse-path carrier fiber lands at the MPOE in the Network Core. Inter-DSX cross-connects can link this campus to other DSX deployments over the same fiber spine.

• Function: east-west CIN spine.
• Scale steps: 10 / 16 / 21 / 32 / 42 / 64 / 128 SU.
• Centralized, modular CIN design supports phased cluster growth while keeping fabric connections within design distance limits.
• East-West CIN scales linearly; smaller configurations are available. For requirements between listed steps, use the next larger CIN configuration.
• TAN and SMN sizing varies by deployment.
• External fiber: diverse-path carrier.
• Cross-connects: Inter-DSX site links.

DSX design value: Organizes stable core services, adaptable perimeter space, and a full-width back-of-house corridor so operations areas can scale or change while preserving access and layout flexibility.

The south side of the campus brings together the Primary Site Entrance, Security Guardhouse, parking lot, and Entry and Offices block. This is where badging happens, vendor tickets are triaged, and the NOC, SOC, and FOC monitor the campus 24/7.

• Entries: primary south plus secondary north.
• Operations: 24/7 NOC/SOC/FOC.
• Full-width back-of-house corridor and regular structural grid support layout flexibility.