China's AI computing power is adopting a new organizational approach

Over the past year, a familiar narrative has repeatedly appeared in global tech news: some company is about to build another multi-ten-thousand-card cluster, with investment figures hitting new records once again.

The numbers keep growing larger. Eventually, they become so big that people lose their sense of concrete scale.

One hundred thousand cards, hundreds of thousands of cards, hundreds of billions of dollars — these figures undoubtedly represent strength. But after wave upon wave of ever-larger numbers, the truly worthwhile question to ask may not be "who is building bigger," but a far more fundamental one: who will ultimately own this computing power, and who will get to use it?

Around this question, the world is moving toward two distinct organizational models.

01. Moat, or Public Network

The first model is to turn computing power into a competitive moat.

Tech giants pour massive capital into building their own clusters, operating them independently, and keeping full control. The larger the cluster, the more powerful the models, the higher the commercial revenue, and the more confidence they have for the next round of investment. Once this cycle starts spinning, latecomers will find it extremely difficult to catch up. In this model, computing power is no ordinary resource — it acts as an entry barrier. Part of its very purpose is to keep others out.

This logic is not hard to understand, and it does work. Large corporations have the capital, data, and engineering teams to centralize computing power for their own models and products, often achieving high efficiency. But the cost is also clear: computing power becomes further concentrated among a small number of players, leaving universities, research institutions, small and medium teams, and developers confined within the boundaries of the APIs, pricing, and rules set by the giants.

This is not a simple moral issue. Commercial companies naturally turn scarce resources into competitive advantages. The real problem is that as computing power increasingly becomes infrastructure for the AI era, if it is only organized in a private manner, the ticket to innovation will keep getting more expensive.

The second model is to organize computing power into a public network.

It does not focus solely on who owns each individual cluster, but rather on whether these clusters can be interconnected, scheduled, and used by more people. The National Supercomputing Internet follows this path. Over 3.5 million CPU cores, 250,000 GPU cards, and 1.4 million registered users — viewed separately, these figures represent scale; viewed together, they represent a new way of organizing resources. Computing power is no longer merely fixed assets on an institution's balance sheet; it begins to transform into a flowing, callable public capability.

History has seen similar shifts. In the 1950s, what truly transformed global trade was not a single larger cargo ship, but standardized shipping containers. By loading goods into boxes of uniform dimensions, they could flow seamlessly between ships, trains, and trucks. Ships remained ships, ports remained ports — but the organizational model changed, and with it, the cost of global trade.

What the National Supercomputing Internet is doing resembles "containerization" for computing power. Scientific computing, model training, and inference services previously often ran on separate systems, each with its own APIs, rules, and queues. The goal now is to enable these tasks to flow freely across a single network. What truly reshapes the landscape may not be making a single point larger, but enabling interconnection between all these points.

02. Viewing 100,000 Cards in the Right Context

Placed in this context, the significance of the Sugon 8000 becomes much clearer.

As China's first fully domestic 100,000-card AI supercluster, the Sugon 8000 is undeniably an engineering achievement. It is large enough and newsworthy enough. But engineering feats are not the most enduring narratives. Records are always meant to be broken, and parameters will eventually be surpassed by later arrivals.

What matters more is which developmental path it has been placed on.

A private cluster typically only needs to serve one owner, one set of business objectives, and one category of core tasks. The goal is clear: make its own model training faster, iteration quicker, and commercialization speedier.

But a 100,000-card AI supercluster connected to a core node of the National Supercomputing Internet faces a different set of problems. It does not only serve a single model — it must simultaneously support scientific computing, model training, inference services, and industry applications. Scientific computing demands high precision, model training requires massive throughput, and inference services need low latency. The requirements of different tasks are not the same, and sometimes they even pull against each other.

This is precisely the question that "Supercomputing-AI Integration" aims to answer.

It is not a catchy concept to stick on posters, but a very concrete systems challenge: can scientific computing and AI computing be organized on a single shared infrastructure? Can different types of tasks stop running in isolated silos? Can massive-scale computing power be turned into a stable, callable service?

What the Sugon 8000 needs to prove is not just whether a fully domestic 100,000-card cluster can be built — that question already has an answer. The harder part lies in the second half: can computing power organized through the public route match the efficiency of the private route? Under the real-world pressure of multiple users, diverse tasks, and varied scenarios, can it run stably, be dynamically scheduled, and be practically usable?

The answer will not appear at a launch ceremony. The real measure of success may be an indicator that rarely gets the spotlight: the social utilization rate of computing power. It is not about how high the peak performance of a single cluster is, but how much of a country's advanced computing power is actually used by scientific research and industry; not about how many resources exist in data centers, but whether those resources are ultimately turned into papers, models, simulation results, industrial applications, and enterprise services.

The private model can achieve extremely high efficiency at a single point, but it naturally creates barriers. The public model faces the opposite challenge: easy to connect, hard to make usable; easy to network, hard to schedule; easy to build a system, hard to operate it stably long-term.

Connecting 100,000 cards to the network is just the beginning. What truly matters afterward is how scheduling works under mixed workloads, whether long-term operation remains stable, how resources are allocated between different users, and whether service quality gets diluted as the user base grows.

As such, the Sugon 8000 is not just a massive machine — it is effectively a stress test for China's public computing power strategy.

03. The Harder Path to Take

To be fair, the path of public computing power is far more difficult to traverse.

The private model follows far simpler logic: capital investment, clear objectives, self-serving operation. As long as commercial returns are large enough, continued expansion is a rational choice.

The public model is far more complex. It requires unified standards, cross-party collaboration, hardware and software ecosystem adaptation, and the patience for long-term operations. Its outcomes may not arrive in a glamorous launch moment. More often than not, they manifest in understated changes: shorter task queue times, research teams able to run more experiment iterations, lower costs for enterprises to validate their models, and more developers gaining access to large-scale computing environments that were once out of reach.

Conversely, if this path fails, there may not be a single spectacular, high-profile collapse. The far more likely outcome is that the system gets built but remains underused; resources are connected but scheduling is dysfunctional; the platform exists, but users find it impractical. Eventually, this will quietly show up on a balance sheet: massive investment, insufficient returns.

This is the real test for the public computing power strategy. Connecting computing power into a network does not automatically make it flow. Building a 100,000-card cluster does not guarantee immediate results for research and industry. There are still layers to cross: scheduling systems, software ecosystems, task adaptation, pricing mechanisms, data security, and operational capabilities.

But if this path succeeds, the changes it brings will run deeper than any single model can achieve.

The barrier to innovation will be redefined. A laboratory, a startup, or an individual developer will no longer need to own an entire data center to qualify for competing in AI and scientific intelligence. They will be able to call computing power on a per-task basis, devoting their money and energy to solving the actual problems, rather than sinking resources into foundational infrastructure just to get a seat at the table.

When shipping containers first appeared, few people cheered for a simple iron box. It was only years later that people truly understood their value — watching global trade costs fall, seeing goods flow faster from port to port, and witnessing the global economy get reconnected.

Computing power has now reached a similar turning point.

100,000 cards are important, but far more critical is how those 100,000 cards are organized. They can become a towering wall for the few, or a network that grants more people access to innovation; they can remain a narrative of scale locked inside data centers, or turn into practical capability that researchers, industries, and developers can truly leverage.

This era of computing power has only just begun.

This article originates from the WeChat Official Account "Bohu Finance" and is published by 36Kr with authorization.