From NVIDIA’s GB200 NVL72 to Huawei’s CloudMatrix 384, and including products from Inspur, Sugon, Alibaba Cloud, Moxie, and many others, a growing number of companies are launching their own super-node solutions. Based on current trends, super nodes are likely to remain the dominant computing-power solution for the foreseeable future.

This represents the evolution of the computing power industry, but it also poses a significant challenge for IDC service providers.

Huawei’s CloudMatrix 384 super node on display

The Inevitability of Supernode Emergence

Data center cabinet power has evolved from a few kilowatts to tens, hundreds, and even megawatt levels—a transition that, under the original projections, might have taken 10 or 20 years. However, following the explosive growth of the AI industry, this timeline has been compressed to just a few years, leaving little to no room for the IDC sector to adequately accommodate the necessary cycles of technological advancement and industry-wide adoption.

At the 20th China IDC Industry Annual Gala held at the end of 2025, Chaojufen CTO Dan Tong stated that, AI chips have an update cycle of 1–2 years, whereas data centers typically have a lifecycle of 10 to 20 years.

In the past, this was never a concern: regardless of the type of servers or chips inside, or whether the services delivered were cloud, virtual machines, or bare-metal, the IDC provider’s offerings—such as server racks, power, cooling, and networking—remained unchanged.

However, in today’s era of intelligent computing, each new chip iteration drives a comprehensive upgrade—spanning server hardware, power requirements, and cooling needs. Moore’s Law continues to hold true in the AI domain: from the H100 to the forthcoming B200 series, the time required to double compute density has been compressed to less than 18 months.

Moore’s Law has also translated into a dramatic surge in chip power consumption: a single DGX H100 server equipped with eight GPUs draws as much as 10 kW, while the B200 server with eight GPUs sees a 50% increase, approaching 15 kW. A NVL72 rack built from GB200 GPUs pushes total power consumption to roughly 120 kW—and that’s just the beginning. NVIDIA says its newly unveiled “Kyber” system can deliver up to 1 MW per rack.

Caption: Nie Lei, General Manager of Zhuhai Hengqin Neogene Intelligent Technology Co., Ltd., recently shared insights on the power dynamics of NVIDIA’s SuperPOD at the “New Technologies” Private Networking Event—AIDC New Architecture, co-hosted by the China IDC Circle Corporate Club and SenseTime Beijing Co., Ltd.

(The latest “New Technology” private gathering—Space Computing Power—will be held on February 3. Please refer to today’s second post for details.)

However, Moore’s Law does not apply to data centers built with reinforced concrete. The power distribution, HVAC systems, and other infrastructure that are designed and installed today—along with the carefully planned gray and white spaces—may become obsolete within a year, unable to meet the demands of next-generation technologies.

The super node integrates computing, networking, power supply, and thermal management systems into a single standardized module. Externally, it provides a relatively uniform interface for water and electrical grids; internally, it autonomously absorbs the complexity introduced by chip iterations. Whether it’s the “prefabricated module” mentioned by Huawei experts at the conference or the “2.5-megawatt micro-module” proposed by Superfusion, the underlying concept is the same: by adopting a standardized modular design, long-life infrastructure can be made compatible with short-cycle compute equipment.

The emergence of supernodes essentially provides a buffering layer. By leveraging high levels of integration, they resolve the contradiction between the rapid iteration of computing hardware and the slower pace of data-center upgrades.

Role Transformation under TCO Logic

While hyperscale data centers have solved major challenges, the changes they have brought about are hardly positive for data center service providers.

Over the past several years, the industry has also been exploring development paths such as full-cabinet and modular architectures. In particular, open computing organizations like OCP have launched products including the Scorpio full-cabinet solution. However, these offerings are largely confined to highly customized applications, and, at their core, full-cabinet designs represent merely physical-packaging optimization—without significantly disrupting the IDC business model.

However, under the hyper-node architecture, a substantial amount of value is integrated directly into the cabinet. Switches, optical modules, power distribution buses, and even liquid-cooling distribution units (CDUs) all become part of the computing infrastructure.

Lin Hai of SenseTime shared TCO data that corroborates this point: over a five-year operational cycle, server depreciation accounts for 80% to 90% of total costs, while infrastructure expenses (including cooling, fire protection, power, and water) represent a relatively smaller share.

For users like SenseTime, the core priority is ensuring zero downtime and keeping their expensive compute resources continuously and fully utilized. While infrastructure services remain important, their relative importance has diminished in the face of the high depreciation costs associated with such costly compute assets. Moreover, many of the traditionally emphasized efforts—such as efficiency and energy conservation—are now largely handled internally by the super nodes themselves. Even stability, which customers value most, is primarily the responsibility of the super nodes.

Consequently, the valuation framework for IDC service providers has changed—or, to put it another way, has been simplified.

Moreover, transformation and reasoning also fall within the supernode’s coverage area.

Although the industry is abuzz with discussions about megawatt-scale single cabinets and all-liquid cooling, real-world demand is complex. For instance, when traditional data centers are retrofitted into intelligent computing centers or inference-oriented data centers, the theoretical requirement for compute density is not as high, allowing for a wider range of deployment architectures.

In the future, data centers are likely to be categorized into two types: one is high-computing-density AI computing centers geared toward training, and the other is inference-oriented AI computing centers with relatively lower computing-density requirements, which can leverage existing data-center infrastructure or even be deployed at edge data centers.

Yang Bifei, Vice President of Huawei’s Data Center Energy and Critical Power Product Line, pointed out that AI data centers will not be entirely dominated by liquid cooling. With the widespread adoption of models such as DeepSeek, inference workloads now account for more than 60% of total AI computing power. Many inference workloads do not require ultra-high-density environments; air cooling and low-density liquid cooling still have a wide range of applications.

Of course, for customers unable to undertake large-scale liquid-cooling upgrades, Huawei has not given up; instead, it has launched the air-cooled version of the Atlas 850 super node, targeting this market segment. In other words, inference and edge computing will also become target customers for supernodes.

The emergence of supernodes marks the industrialization of computing power infrastructure.

It is no longer merely a stack of discrete devices; instead, through deep integration, it resolves the temporal and spatial misalignment between computing power and infrastructure. As we push beyond physical limits, adapting to this transformation has become a shared challenge for both upstream and downstream players in the industry chain.

For the industry chain, OEMs have gained greater influence through integration. But for IDC service providers, after supplying power strips, piping, and network cables, how can they demonstrate their value in ensuring the stable operation of hyperscale nodes?