NVIDIA’s GB300 NVL72 rack will lead AI rack shipments in 2026 before Vera Rubin VR200 racks become the largest shipment share during 2027, according to TrendForce, which projects combined GB300, VR200 and VR300 NVL72 output value above US$710 billion in 2027 and a 214% year-on-year surge in NVL72 rack output value that year. That shift turns data centre planning into a per-rack problem: operators must provision power, liquid cooling and NVLink topology at the rack level, not the server level.
What is the full-rack power envelope for GB300 NVL72, and how should power delivery be provisioned?
GB300 NVL72 has a full-rack power requirement of 142 kW. The reference architecture spells out the components behind that figure: power shelves, NVSwitch trays, management switches and liquid-leak detection, alongside 72 Blackwell Ultra GPUs, 36 Grace CPUs and 130 TB/s of NVLink bandwidth. Because NVL72 shipments across GB and VR platforms will grow more than 50% year-on-year in 2027, a single deployment can aggregate multiple 142 kW racks on one power distribution run. Facility teams should plan upstream switchgear and busway capacity around that per-rack envelope, and coordinate with rack vendors on power shelf redundancy and monitoring, as described in power and cooling considerations for AI deployments.
What liquid-cooling and CDU capacity should operators plan for the GB300-to-Vera Rubin transition?
The available data does not state an exact CDU kilowatt rating for either GB300 or VR200, but the 142 kW full-rack power requirement and the inclusion of liquid-leak detection in the GB300 NVL72 reference architecture make direct liquid cooling a mandatory planning assumption. Cambridge Nexus, which operates full GB300 NVL72 racks as a bare-metal AI Factory service, structures deployments around seven layers, including power and cooling, and its racks are pre-manufactured in a Taiwan factory before installation at a selected data centre site. Because Vera Rubin system ASP is roughly double GB300, data centre teams should expect higher per-rack thermal loads even though the specific VR200 rack power is not disclosed. For now, the 142 kW GB300 NVL72 figure is the only verified rack-level power envelope; operators should use it as the planning floor for liquid-cooling loops, CDU isolation valves and leak-detection zoning. The Open Compute Project (OCP) provides open specifications for power shelves and liquid-cooling interfaces that can guide these decisions.
Practical considerations for 2027 volume deployments:
- Plan power delivery for a 142 kW full-rack GB300 NVL72 envelope, not per-server power budgets.
- Require liquid-leak detection in rack specifications and ensure CDU/facility water loops can isolate a leaking rack.
- Align installation and acceptance timelines with CNEX-style 60-day contract-to-installation commitments, but verify rack availability before site readiness.
- Map every GPU to its NVLink domain before scheduling; all GPUs in one NVL72 rack report the same cluster UUID.
- Treat pre-manufactured racks as factory-integrated units to reduce on-site cabling and commissioning, but confirm management switch and power shelf integration.
How does rack-level NVLink topology affect scheduling and memory sharing within a 72-GPU domain?
All GPUs in one NVL72 rack report the same cluster UUID, meaning the entire rack is a single NVLink domain for topology-aware schedulers. NVIDIA Mission Control maps cluster UUID to NVLink domain and clique ID to NVLink partition, and it integrates with Slurm and NVIDIA Run:ai workload management platforms. An IMEX daemon runs in the host OS of every compute tray, enabling shared-memory multi-node CUDA, while Slurm's topology/block plugin exposes NVLink partitions as blocks for placement decisions. MNNVL then enables shared GPU memory across nodes within the rack via IMEX-capable trays. For operators, this means the rack is not just a power and cooling unit; it is a scheduling and memory-coherency boundary. Mission Control capabilities will extend to the Vera Rubin platform, including Rubin NVL8, so planning for rack-level NVLink should not be delayed until VR200 ramps. More detail on fabric choices is covered in choosing the right interconnect for AI workloads.
What rack-level integration and deployment timelines should operators expect from GB300 to VR200?
The deployment timeline shows GB300 racks leading NVIDIA AI rack shipments in 2026, with demand strong into first-half 2027, and VR200 racks ramping later to become the largest shipment share during 2027. NVL72 shipments across GB and VR platforms grow more than 50% year-on-year in 2027, and TrendForce projects combined output value above US$710 billion. On the integration side, CNEX commits to a 60-day contract-to-installation and acceptance window, but that speed depends on rack availability. At the same time, buyer mix is shifting: five major North American CSPs remain primary buyers, but their procurement share falls from approximately 70% in 2025 to 60% in 2026, while Tesla (with xAI and SpaceX) and CoreWeave drive rising non-CSP rack demand. For data centre operators, this means floor plans must accommodate both hyperscale and AI-native tenants with different contract lengths and power requirements. NVIDIA Data Center revenue now approaches 93% of total company revenue, which reinforces how much of the rack-scale transition is concentrated in AI infrastructure. Follow the AI data centres coverage hub for ongoing updates on these rack generations.