GB300
GB300 NVL72
16 compute racks plus 16 network and service racks, hosting 1,152 GPUs.
MDC2000 · AI INFERENCE AND TRAINING
Give 1,152 GPUs the infrastructure they need.
MDC2000 targets 2.5 MW of IT load for larger AI clusters. Power, direct liquid cooling and supporting air cooling are coordinated around the selected NVIDIA server configuration.
- 2.5 MW
- Target IT load
- 142 kW
- Maximum IT load per compute rack
- 1,152
- GPUs in the GB300 NVL72 layout
- 1.10–1.15
- Design PUE target
Design targets are checked against the selected hardware and complete load schedule. Energy and water performance depend on the site climate, operating load and cooling configuration.
Power and cooling sized for the same cluster.
High-density GPU racks need electrical capacity, liquid interfaces, network space and service access that fit together. Supporting equipment still needs air cooling.
MDC2000 resolves those systems in one modular facility, designed around the selected NVIDIA reference design and the site.
Four systems designed together.
- 01
Power subassembly
Low-voltage receiving interface, modular UPS, batteries and independent A/B/C/D rack distribution. Upstream MV switchgear, power transformers and standby generators are outside the base scope unless specifically included.
- 02
IT environment
32 racks in the GB300 layout: 16 compute racks and 16 network and service racks, with access to power, liquid connections and cabling.
- 03
Thermal & facility subassembly
Cooling-loop pumping, filtration, controls, leak monitoring and air-side cooling, with fire-protection and monitoring interfaces defined for the project.
- 04
Heat rejection
External heat rejection selected for the site climate and availability requirements. Dry operation can be supported by stored-water adiabatic assistance where needed.
Module geometry, transport splits and site connections are agreed during project engineering.
16 compute racks. 16 for networking and service.
The current GB300 NVL72 configuration provides 32 racks and hosts 1,152 GPUs in its 16 compute racks. Network and service equipment are selected around the cluster fabric.
| Rack allocation | Quantity | Design role |
|---|---|---|
| GB300 NVL72 compute racks | 16 | 1,152 GPUs in the GB300 NVL72 configuration |
| Network and service racks | 16 | Networking, storage and service equipment for the selected cluster |
| Total racks | 32 | Compute and support space in one coordinated facility |
The current target is 2.5 MW IT load, with 142 kW per compute rack. The rack maximum is a design limit, not an instruction to multiply every rack by 142 kW. The final equipment schedule sets the total.
Select the complete NVIDIA configuration.
B300 / B200
Other NVIDIA server configurations
Discuss the selected B300 or B200 server arrangement. Rack power, cooling, networking and service geometry are checked for that equipment.
Platform positioning covers GB300, B300 and B200 configurations. The 1,152-GPU count belongs to the GB300 NVL72 layout; other server arrangements need their own rack and load plan.
Direct liquid cooling for the main load. Air for the rest.
The current product direction uses an approximately 85% direct liquid / 15% air cooling split. Coolant carries heat from the main compute equipment; an air-side system serves the remaining components and supporting infrastructure.
85%Direct liquid cooling in the current design direction
15%Air cooling for residual and supporting loads
- GPU/CPU cold plates
- OEM-specified coolant loop
- rack CDU
- Facility cooling loop
- Site-selected heat rejection
- Residual rack + network/support heat
- air-side system
- facility cooling loop and heat rejection
The coolant, CDU interfaces, flow rates and operating temperatures follow the selected OEM equipment and project design.
The split describes the design direction. The final thermal schedule establishes the actual liquid and air duties.
Low water use starts with the heat-rejection design.
Dry heat rejection is the starting strategy. Stored, treated water can support adiabatic assistance when the climate requires it.
MDC2000 targets PUE 1.10–1.15. The PUE range means facility energy is 10–15% above IT energy, making the infrastructure overhead clear. It depends on climate and load; it is a design target, not a measured result for every installation.
Annual modelling establishes the energy and water use for the site. Storage volume and refill arrangements are designed separately from annual consumption.
Designed around NVIDIA reference requirements.
The electrical design coordinates rack supply paths, UPS protection and supporting loads with the selected NVIDIA reference design.
Upstream sources, maintenance, fault domains and control sequences are reviewed as one availability design.
Uptime Institute requirements and OCP recommendations are assessed where relevant to the project. A reference-design statement does not claim a completed facility certification.
Design the site around the cluster.
- 01
Confirm the equipment and location.
Define the server configuration, network, storage, IT load, site climate and deployment plan.
- 02
Approve the basis of design.
Freeze electrical, thermal, civil, controls, fire and OEM boundaries.
- 03
Complete detailed engineering.
Review fibre routes, rack access, power paths, liquid interfaces, transport and outdoor heat rejection.
- 04
Factory-integrate the AI Block.
Assemble power, thermal, controls and monitoring systems in a controlled environment.
- 05
Complete factory acceptance testing (FAT).
Test the agreed electrical, thermal, control and alarm functions before shipment.
- 06
Install and connect.
Complete site works in parallel, then position and integrate the system.
- 07
Complete SAT / IST and commissioning.
Verify the transported and site-connected facility before operational handover.
The delivery programme separates factory work, transport, site installation and commissioning, with the dependencies for each stage agreed in the proposal.
Current product targets and project scope
- Target IT load
- 2.5 MW
- Maximum IT load per compute rack
- 142 kW
- GB300 rack layout
- 16 compute + 16 network and service racks
- GB300 GPU population
- 1,152 GPUs
- Purpose
- AI inference and training
- Target equipment
- NVIDIA GB300, B300 and B200 configurations
- Cooling direction
- Approximately 85% direct liquid / 15% air
- Design PUE target
- 1.10–1.15; site climate and load dependent
- OEM interfaces
- Coolant, CDU, power and network interfaces agreed for the selected equipment
- Heat rejection
- Selected for the climate and availability requirements
- Power architecture
- Coordinated with the selected NVIDIA reference design
- Project scope
- Infrastructure, OEM equipment and site works defined separately
- Acceptance
- FAT before shipment; site acceptance, integrated testing where agreed, commissioning and handover
The final load and cooling schedule reconciles rack limits, total IT load, OEM interfaces and site climate. Numerical PUE values are design targets, with annual performance modelled for the project.
Agree the infrastructure and equipment scope.
INSIDE THE COORDINATED MDC2000 DESIGN
- IT environment and rack service access
- LV receiving interface, UPS, batteries and rack distribution
- Facility cooling-loop distribution and air-side support cooling
- Controls, monitoring and leak detection
- Coordinated fire-protection, security and BMS/DCIM interfaces
- Factory integration and the agreed FAT scope
CONFIRMED BY PROJECT
- Compute, network and storage hardware
- OEM rack and CDU secondary-loop implementation
- Upstream MV switchgear and transformers
- Standby generation
- External heat-rejection equipment and redundancy
- Site civil works, foundations, utilities and access
- Fire detection/suppression equipment and local approvals
- SAT, integrated testing, certification and operational support
FOR SITE OWNERS AND ENERGY COMPANIES
Build capacity for a tenant’s computing equipment.
MDC2000 can form the infrastructure for a colocation business. Alongside the facility, we can discuss introductions to prospective tenants and the requirements they bring. Tenant placement and commercial terms are agreed by project.
Tenant introductions do not guarantee occupancy or investment returns. MDC4000 is a coming-soon option for future capacity planning.
Frequently asked questions
What rack density can MDC2000 support?
The current product direction is 142 kW per compute rack, within a 2.5 MW IT-load design. The GB300 NVL72 layout has 16 compute racks and 16 network/service racks, hosting 1,152 GPUs in the compute racks. Per-rack limits and total capacity are checked together against the final hardware and load schedule.
Is the equipment cooled by air or liquid?
Conventional CPU and storage systems and MDC60 use air cooling. MDC500 has air-cooled and proposed immersion configurations, selected by project. MDC2000 uses an approximately 85% direct liquid / 15% air cooling split. MDC4000 is planned for 100% direct liquid cooling. Compatibility follows the exact server configuration.
What PUE and water-use performance should I expect?
MDC2000's current design target is PUE 1.10–1.15. It is a design target rather than a measured result for every site. Water use depends on the country and climate of the installation. The project model must state its climate, IT-load profile, cooling configuration and annual energy and water boundaries.
Does liquid cooling need a permanent water connection?
Liquid circulates in closed cooling loops. Water consumption depends on the outdoor heat-rejection system. MDC2000's dry-first approach can use stored, treated water for adiabatic assistance in hot conditions, without a permanent utility-water connection. Storage, annual use and refill arrangements follow the site climate.
Is IT load the same as total facility power?
No. IT load is the electrical power used by servers, storage and networking. The site supply also serves cooling, power-system losses and other facility equipment. Its required capacity comes from a project load study, including peak conditions and operating modes; annual PUE alone cannot size the connection.
Do you supply the infrastructure or the servers as well?
Our core offer is the data-center infrastructure. We can plan integration around customer-supplied equipment or discuss an equipment package in the project scope. The proposal states who procures and supplies the servers, GPUs, networking and storage; they are not automatically included.
Can you help find a tenant for a colocation project?
For suitable projects, we can help connect a site owner with a potential infrastructure tenant. We supply the data-center infrastructure; the tenant operates its computing equipment. Introductions and commercial terms are project-specific, with no guarantee of occupancy or investment return.
What standards and certifications apply?
The design is reviewed against the requirements relevant to your project, including the selected NVIDIA reference design and any specified availability criteria. Uptime Institute requirements and OCP recommendations are assessed where applicable. A reference-design statement is separate from certification; any certificate must identify the product, design or facility it covers.
What footprint and site access will I need?
We issue the layout, foundation loads, service clearances and delivery and lifting requirements for your configuration. Share the site plan and access route early so the module and external cooling equipment can be placed and maintained.
Plan the facility for your AI cluster.
Share the hardware plan, available power or location. We will discuss the infrastructure and, for colocation projects, the potential tenant requirements.