NVIDIA DSX POD and Domestic Modular Intelligent Computing: AI Competition Shifts Toward Compute Delivery

I saw a few tech news items these past two days, and I think looking at them together is more interesting than looking at any one of them alone.

First, Nvidia. On October 21, it released a new solution called DSX POD, preparing to integrate in advance the computing equipment, rack-level power supply, network cabling, and liquid cooling systems needed for AI data centers, then deliver them to the site for deployment in a modular fashion. According to the targets Nvidia announced, when the site's basic conditions are already in place, on-site deployment work that might originally have taken six months can be compressed to as fast as 14 days.

Nvidia has already made a fortune from GPUs over the years, and now it has even started seriously studying how data centers are built, installed, and delivered.

But what really interests me is not that container, nor the number 14 days, but that Nvidia is extending its competitiveness from chips to the delivery of entire computing power systems.

At almost the same time, there was also a series of moves in China. ZTE launched a new generation of modular intelligent computing infrastructure products, Huawei previously released a 3D data center architecture, and there was also new progress on a domestic 1024-card liquid-cooled super node. On the other side, Samsung and SK Hynix are advancing the technology upgrade of their NAND flash memory factories in China.

These news items appear to have no direct relationship, but I think they happen to illustrate one thing: the next stage of competition in the AI industry may not just depend on who has more and stronger chips, but on who can turn computing power into real productivity faster and more efficiently.

And as competition starts to develop in this direction, the manufacturing, communications, power, and engineering capabilities that Chinese companies have accumulated over the long term may usher in a new release of value.

1. Why has Nvidia started worrying about how data centers are built?

But there is actually an easily overlooked question here: does buying GPUs mean you can immediately use that computing power?

Obviously not.

From equipment procurement to actually putting an AI data center into operation, there is still a great deal of work in between. Servers need a stable power supply, liquid cooling systems need to handle increasingly high heat loads, high-speed networks need to connect large numbers of computing devices, and the entire system still has to go through installation, debugging, and validation. If any link cannot keep up, it may leave expensive computing equipment unable to be put into use for a long time.

The DSX POD released by Nvidia this time is an attempt to solve part of the problem. It hopes to move a large amount of integration work that originally had to be done on-site at the data center into the factory in advance. Equipment is integrated and tested at the factory, then shipped to the site for deployment, reducing the uncertainty of construction and debugging as much as possible.

Of course, modular data centers were not invented by Nvidia. DSX POD is currently mainly a reference architecture designed around the new-generation Vera Rubin computing platform, and relevant partners are expected to begin offering products from early 2027. The so-called fastest 14-day deployment cycle is likewise based on the premise that basic conditions such as external power supply, cooling, and networking are already prepared, and does not mean that building a data center from scratch takes only two weeks.

But these do not affect my judgment on this matter.

In the past, Nvidia's most important advantage came from GPU performance and the CUDA ecosystem. After customers bought equipment, they still had to solve a large number of system integration and deployment problems. Now, Nvidia is further intervening in these links, trying to help customers turn purchased hardware into computing power that can actually run faster.

The business logic here is very direct. For AI data centers with huge investment, every day that computing equipment sits idle means a loss of cost and potential revenue. As AI infrastructure becomes increasingly expensive, the importance of deployment efficiency, operational efficiency, and long-term cost will naturally continue to rise.

Nvidia is no longer satisfied with letting customers buy stronger chips; it has started to care about when these chips actually start running.

In my view, this may be more worthy of attention than the promotional figure of 14-day deployment.

2. The opportunities for Chinese companies are hidden precisely in these "inconspicuous" links

If we continue following Nvidia's line of thinking, we will find an interesting phenomenon: as AI infrastructure increasingly emphasizes modularization, standardization, and rapid delivery, manufacturing and engineering capabilities that were originally less dazzling than GPUs have instead become increasingly important.

Just one day before Nvidia released DSX POD, ZTE launched the ZXEDC M8000 intelligent computing micro-module solution, modularly integrating systems such as liquid-cooled cabinets, networking, power supply and distribution, and monitoring, aimed at high-density AI computing scenarios.

Such products are not as easy to generate buzz as new-generation GPUs, but the problems they solve are very practical. In the past, when building data centers, many devices had to be purchased and installed separately, and then different teams had to complete connection and debugging. Now, manufacturers are beginning to try to integrate this work in advance, gradually shifting data center construction from a large amount of on-site construction to more standardized system delivery.

What is more noteworthy is that this time ZTE did not only have a product launch.

According to ZTE's official disclosure, in 2025 it obtained an order for 666 IT modules for the Yangtze River Delta computing center project of a domestic internet company, of which 345 M8000 IT PODs have already been delivered. From order placement to equipment arriving at the site, it took about one month.

It should be noted here that one month refers to the period from order placement to equipment arriving at the site, not the entire data center construction or commissioning cycle, and it cannot be directly compared with Nvidia's 14-day deployment target.

But this case at least shows that modular intelligent computing infrastructure in China is no longer just a product concept at launch events, but has entered actual project delivery.

I think this is more valuable than simply arguing over who proposed modularization first. Because the real competition is not about who tells a prettier story about the solution, but about whether complex equipment systems can be made into products that can be repeatedly produced and stably delivered.

The 3D data center architecture released by Huawei in September responds to the same question from another direction. It attempts to organize cooling, IT equipment, power supply, and backup power systems in a three-dimensional layered manner, improve the deployment conditions for high-density computing equipment, and combine factory prefabrication and on-site assembly to improve construction efficiency.

ZTE, Huawei, and Nvidia's solutions are not exactly the same; their technical routes and applicable scenarios also differ, so one cannot simply compare which is more advanced. But they are all answering the same question: as AI computing equipment becomes increasingly dense and increasingly expensive, how can this infrastructure be built and operated more efficiently?

I think there is a change here that is very much worth watching.

In the past, when discussing domestic AI, it was easy for us to focus our attention on models and chips. Which model topped the rankings, and how much the performance of a domestic chip improved, tended to become the hot topics more easily. But as AI enters the stage of large-scale application, the importance of power supply, cooling, networking, and engineering delivery will be re-recognized by more and more people.

And these links are precisely closely related to the capabilities China's manufacturing industry has accumulated over the long term.

China has a vast electronics manufacturing and communications equipment industry, a mature power supply and distribution, cooling, and engineering construction system, and also a great deal of experience in data center construction and operations. As AI infrastructure gradually moves from customized engineering to standardized products, these capabilities have the opportunity to be recombined to form new competitive advantages.

This does not mean that Chinese companies are already leading in all technologies, but at least it shows one thing:

The industrial change Nvidia is driving will not only expand Nvidia's own advantages, but may also allow Chinese companies to find new competitive space.

Moreover, there has recently been another case in China that gave me a more concrete understanding of this possibility.

III. 1024-card super node: what is truly worth watching is not the number 1024

On October 15, according to public reports from the same period, the first 1024-card Ascend Atlas 950 SuperPoD liquid-cooled super node was lit up, and adaptation for the Step 5 Preview model from StepFun was advanced in parallel. At the same time, domestic computing platforms such as MetaX, Biren, and Moore Threads also announced support for this model.

This news is easily understood as the scale of domestic computing power having grown again, but I think what is truly worth watching is not the number 1024, but that domestic computing platforms are trying to solve a more difficult problem: how to make a large number of chips not just placed together, but actually able to work together efficiently.

Large-model computing, especially models using the MoE architecture, places very high demands on data communication between chips. If interconnection efficiency is not high enough, even if there are a large number of computing devices, they may spend a great deal of time waiting for data and ultimately fail to fully utilize their theoretical computing power.

Huawei's super node adopts a high-speed interconnect architecture, attempting to organize 1024 Ascend NPUs into a more tightly coupled computing system, while also addressing the cooling and operation problems brought by high-density deployment through liquid cooling and system-level design.

This matter and the data center productization Nvidia is advancing are actually located at different links in the same industrial chain. Nvidia hopes to enable computing systems to complete deployment faster, while domestic super nodes are exploring how to enable a large number of computing chips to work together more effectively.

Of course, lighting up equipment does not mean long-term production environment validation has been completed. The actual computing efficiency, stability, and cost performance of the super node still need to be proven by more real business workloads and independent testing.

On the same day, multiple domestic computing platforms announced support for StepFun's new model, which also reflects another layer of change. For developers, no matter how powerful a hardware system is, if new models cannot be adapted in a timely manner, its actual value will also be limited. The time gap between model release and computing platform adaptation is becoming a practical factor in measuring ecosystem capability.

Of course, announcing support does not mean performance optimization for all scenarios has been completed, but the ability to establish adaptation relationships between models and hardware more quickly is itself a capability that the domestic computing ecosystem needs to continuously improve.

IV. Samsung and SK Hynix's Chinese factories remind us not to overlook the manufacturing system

On October 21, Reuters reported that Samsung and SK Hynix are advancing technology upgrades at their NAND flash memory factories in China. Among them, Samsung's Xi'an factory involves a higher-layer NAND process, and SK Hynix's Dalian factory is also carrying out related production line upgrades. The two companies confirmed the upgrade work, but some specific technical targets in the report still came from people familiar with the matter.

This matter is taking place against the background that export restrictions on advanced semiconductor equipment still exist, so it is particularly worth watching. The relevant upgrades are still constrained by requirements such as U.S. equipment export licenses, and cannot simply be understood as all technology upgrade plans having already been approved. Samsung also clearly stated that this upgrade is not an expansion of capacity or the construction of new factories.

But I think that, rather than simply discussing why companies still need to continue upgrading Chinese factories, the more valuable question is: why do these manufacturing bases that have been operating for many years still have such important value?

The value of a mature semiconductor factory has never been only equipment and buildings. Behind it there are also process experience, engineering talent, production management systems, and supply chain coordination accumulated over many years. Even with funds, it is very difficult to completely replicate these conditions in another place in a short period of time.

For global memory companies such as Samsung and SK Hynix, Chinese factories are an important part of their long-term global manufacturing layout. As memory technology continues to iterate, maintaining the competitiveness of existing production lines is itself a normal industrial operating need for companies.

At the same time, China's domestic memory industry is also developing. According to Omdia data cited by Reuters, in the second quarter of 2026, YMTC's share of China's NAND market reached 38.5%. This is only data for the Chinese market, and does not mean that Chinese companies have already taken a comprehensive lead in the global memory industry, but it shows that domestic companies are forming increasingly strong market competitiveness.

I do not think it is necessary to write this matter as a simple win-or-lose relationship between international companies and Chinese companies. What is truly worth watching is that China is forming a memory industry system that has both the long-term manufacturing layout of international companies and the continuous innovation of domestic companies.

Although NAND flash memory is not a GPU, nor is it the HBM high-bandwidth memory required for large-model training, it is an important foundation for storage devices such as enterprise SSDs. As the scale of AI data centers expands, storage, networking, and other infrastructure beyond computing equipment also need continuous upgrading.

The production line upgrades at Samsung and SK Hynix have no known direct supporting relationship with NVIDIA's DSX POD. But it reminds us from another angle:

Advanced computing systems can be continuously updated, yet the industrial manufacturing system that supports them requires long-term accumulation, and that accumulation itself is competitiveness.

For China, this may be an industrial value more worthy of long-term attention than simply discussing whether a particular technology leads or lags.

V. In the next stage, China has a chance to compete in its own way

This question is of course important. The gaps in advanced computing chips, manufacturing processes, and software ecosystems will not automatically disappear just because China has a huge manufacturing base. For domestic AI to truly build long-term competitiveness, breakthroughs in core technologies remain an unavoidable task.

But after reading these recent news stories, I have instead become increasingly convinced that in the next stage, there may be another thing worth watching.

This does not mean domestic AI has already won, nor does it mean system integration can replace performance progress in core chips. But if future AI competition really shifts from "comparing who has more chips" to "comparing who can turn computing power into productivity faster," China should not be just a pursuer.

Because this competition requires not only advanced chips, but also manufacturing capability, supply chain organization, engineering experience, software adaptation, and the ability to operate stably over the long term. Chinese companies have already accumulated considerable experience in some of these areas and are continuously exploring through new products and projects.

The real thing that still needs to be proven is whether these scattered capabilities can form complete, reliable, and reasonably costed solutions and withstand long-term verification in practical applications.

NVIDIA's new move shows us that the boundaries of global AI industry competition are expanding. And a series of recent advances by domestic companies have made me more convinced that the future competitiveness of Chinese AI does not necessarily have to be found only along the paths already taken by overseas giants.

In this game, we have our own way of playing.

Sources

NVIDIA Newsroom, October 21, 2026, official release of the DSX POD AI Factory.

ZTE official website, October 20, 2026, ZXEDC M8000 intelligent computing micro-module product and project delivery status.

Huawei official website, September 16, 2026, release of the 3D data center architecture.

IT Home and contemporaneous reports, October 15, 2026, the 1024-card Ascend liquid-cooled super node and Step 5 Preview model adaptation progress.

Reuters, October 21, 2026, technology upgrades at Samsung and SK Hynix's China NAND factories, as well as Omdia's China NAND market share data.

This article comes from the WeChat public account "Tech Vortex," author: Wang Qinzhou, published by 36Kr with authorization.