Primary sources: NVIDIA's technical blog "Delivering Flexible Performance for Future-Ready Data Centers with NVIDIA MGX" (2025-12-15) and the official RTX PRO 6000 Blackwell Server Edition specification page. This article is a secondary analysis written for procurement and operations decisions; every calculated figure is labeled as such.
Do the Power Math Before the Compute Math

NVIDIA opens its blog with a blunt diagnosis: the AI boom keeps accelerating into 2026, and enterprise data centers are approaching the limits of conventional server and rack architecture. The pressure lands on three things — power budgets, cooling headroom, and facility layout. That is not rhetoric. Line up the official numbers and the arithmetic makes the case for you.
The MGX 6U reference design carries eight liquid-cooled RTX PRO 6000 Blackwell Server Edition GPUs per node. Each GPU is rated at a maximum of 600 W (configurable), so the GPUs alone draw 8 × 600 W = 4.8 kW per node (calculated). Fill a standard 42U rack with 6U nodes — seven of them, a theoretical maximum that ignores space for switches and cable management — and you are looking at 56 GPUs and 33.6 kW of GPU power alone (calculated). That figure says nothing yet about CPUs, DPUs, NICs, or power-conversion losses.
Most server rooms designed for traditional enterprise workloads cannot feed or cool a node like that, let alone a rack full of them. Which is why the real audience for MGX 6U is not just server OEMs. It is every facility owner planning to "buy GPUs and do AI": ask whether the room can take the load before you ask when the cards arrive.
6U, 800 mm, and the HPM: Turning a Generation Change into a Module Swap

MGX is a modular reference architecture — standardized building blocks that shorten time-to-market (TTM). This iteration adds a 6U chassis configuration at 800 mm depth. Three details deserve a procurement reader's attention.
First, one chassis spans multiple compute generations. NVIDIA states explicitly that the same chassis design supports multiple generations of compute platforms and multiple workload configurations, reducing dependence on disruptive redesigns. In procurement terms: the chassis, power delivery, liquid-cooling plumbing, and service aisles are heavy assets you can reuse across generations. When the next platform arrives, you swap compute modules — not the rack's infrastructure.
Second, the host processor module (HPM) supports both x86 platforms and NVIDIA's next-generation Vera CPU, so a single server design can be standardized across CPU architectures. For Pearl River Delta customers running x86 virtualization clusters who also want to test the Arm ecosystem, this is a rare "one chassis, two roadmaps" option. Stay clear-eyed, though: an HPM option does not mean an OEM will actually ship both configurations. Confirm the real BOM before you order.
Third, the 800 mm depth and larger chassis volume buy serviceability: field-replaceable units (FRUs) such as NICs and power supplies are easy to reach. The nightmare scenario for a liquid-cooled node is "replace one hose, tear down half the server." How the maintenance path is designed shows up directly in your operating costs.
For context: HGX is closer to a turnkey GPU baseboard; MGX is closer to standard blocks plus a reference design. OEMs get more freedom, but more of the system-integration and validation burden lands on them too. When you buy an MGX-based machine, you are really buying the quality of that OEM's implementation of the reference architecture.
The Liquid-Cooled RTX PRO 6000: Single-Slot, 96 GB, and a Configurable 600 W

The key numbers from the official specification page: 96 GB of GDDR7 with ECC, a 512-bit bus, 1,597 GB/s of memory bandwidth, 24,064 CUDA cores, 4 PFLOPS of FP4 compute and 2 PFLOPS of FP8. Maximum power is 600 W and configurable. The air-cooled card is a dual-slot FHFL design; the liquid-cooled version is single-slot FHXL.
The liquid-cooled variant matters for more than running cooler. The single-slot form factor is what makes eight GPUs in 6U possible at all. Deleting the GPU fans saves parasitic power and moves heat rejection from room air into the coolant — the precondition for high-density racks. And the configurable 600 W ceiling is an efficiency dial: inference-heavy workloads can trade power for better output per watt instead of passively running at the limit.
The blog also stresses that the GPU is certified by more than 50 enterprise ISVs and ships with the AI Enterprise, Omniverse, vGPU, and Run:ai software stack. Its positioning is general-purpose enterprise acceleration — agentic AI, physical AI, simulation, graphics, and video — not a training-only card.
Networking is the increment most readers will miss. The ConnectX-8 SuperNIC integrates a PCIe Gen 6 switch, and at a 2:1 GPU-to-NIC ratio each GPU gets up to 400 Gb/s of network bandwidth. Against designs that hang a discrete PCIe switch off the side, NVIDIA claims double the network bandwidth per GPU and up to 2× the NCCL all-to-all performance. One detail is easy to misread: the GPU's own specification page lists PCIe Gen 5. The Gen 6 switching lives on the SuperNIC side. The two statements do not conflict — but do not write "the GPU supports Gen 6" into your proposal as a selling point. The BlueField-3 DPU, meanwhile, offloads line-rate encryption, microsegmentation, and threat detection from the host CPUs and GPUs, and the blog notes future ConnectX-9 support.
Serviceability, TTM, and Cross-Generation Compatibility: What to Put in the Contract
The blog's timeline: global system builders are expected to launch liquid-cooled RTX PRO 6000 systems and MGX 6U liquid-cooled RTX PRO servers in the first half of 2026. Note the tense — that was the plan at announcement time. If you are procuring in November 2026, verify actual availability and configurations with the OEM rather than quoting the blog.
A reference architecture is not a finished system. MGX lowers the OEM's engineering barrier and shortens TTM, but power-delivery implementation, liquid-cooling interfaces, firmware quality, and spare-parts logistics still vary by vendor. At minimum, lock the following into the contract: an MGX compliance statement with a list of deviations; the HPM roadmap and cross-generation compatibility commitments; CDU and quick-disconnect interface specifications; a FRU list with spare-parts SLAs; and the firmware-update support window. Cross-generation compatibility only pays off when the next generation actually arrives — a verbal promise is worth less than a written clause.
Guangzhou and Pearl River Delta Facilities: A Retrofit Assessment Checklist
Given the local climate and the existing building stock, here is the assessment order we recommend (engineering advice, not figures from the sources).
Power first. Per the calculations above, the GPU side of a single node starts at 4.8 kW, and the full node adds CPUs, networking, and conversion losses on top. Three-phase PDUs, busway capacity, and end-of-row distribution redundancy must be sized against measured loads. Do not apply old rules of thumb from low-density cabinets to AI nodes.
For liquid cooling, focus on three things: CDU redundancy (start at N+1), secondary-loop water quality and leak detection, and condensation risk. Pearl River Delta summers are hot and humid; if supply-water temperature drops below the room's dew point, condensation on the pipework becomes an incident waiting to happen. The supply-temperature window for liquid cooling can be wider than for air-cooled chilled-water systems, but it has to be designed together with the room's dehumidification capacity. On the plant side, weigh cooling towers against air-cooled chillers on year-round efficiency, not just equipment price.
On retrofit sequencing, existing facilities should start with a "liquid-cooling-ready zone": carve out an area where both power and cooling meet the bar, deploy MGX 6U there, and keep the rest of the room on air-cooled workloads. Greenfield projects should design power and pipework for liquid-cooling density from day one. The biggest risk is not buying the wrong card — it is the cards arriving before the room can take them. GPU lead times and facility-retrofit schedules belong on the same Gantt chart.
Sources
- NVIDIA Developer Blog (Chinese site): "Delivering Flexible Performance for Future-Ready Data Centers with NVIDIA MGX" (利用 NVIDIA MGX 为未来数据中心提供灵活高效性能), 2025-12-15. https://developer.nvidia.cn/blog/delivering-flexible-performance-for-future-ready-data-centers-with-nvidia-mgx/ (fetched and verified directly)
- NVIDIA official product page: RTX PRO 6000 Blackwell Server Edition specifications. Page undated; accessed 2026-11-15. https://www.nvidia.com/en-us/data-center/rtx-pro-6000-blackwell-server-edition/ (fetched and verified directly)
Note: the 4.8 kW, 33.6 kW, and 56-GPU figures are calculations from official numbers (labeled above). The HGX comparison, contract clauses, and facility-retrofit advice are this article's original analysis, not source content.
