The overseas ban on intelligent computing power has been the final piece of the puzzle for China's intelligent computing platform capabilities. From now on, we can fully sow, cultivate, and harvest AI innovation within our own borders, ensuring higher controllability and security.

Industry debate persists as to whether China, currently one of the top two AI powers globally, can ascend to the sole throne. Judging by our current infrastructure developments, this milestone could realistically be achieved within a decade.

Only by being Chinese first can we eventually become global leaders.

--Introduction

In October 1934, a pivotal decision was made: the Red Army embarked on the Long March.

Throughout the Long March, the Red Army continually adapted its strategies and tactics based on evolving circumstances, transforming from a passive position of being "constrained everywhere by the enemy" to one of strategic maneuverability. Through brilliant military maneuvers such as crossing the Chishui River four times, ingeniously traversing the Jinshajiang River, forcibly crossing the Dadu River, and swiftly capturing the Luding Bridge, the Red Army successfully evaded enemy encirclement, pursuit, and blockade, securing a decisive victory in its strategic transfer.

The Long March was brutal, with the Red Army's ranks dwindling from nearly 206,000 soldiers at the outset to approximately 57,000 by its conclusion. However, its victory held profound significance, not only preserving the core of the Red Army but also laying the groundwork for subsequent victories and the development of the Chinese revolution.

This narrative underscores that the inception of great feats is often involuntary and forced. Nevertheless, as long as one retains a sense of direction and mission, strategic adjustments notwithstanding, the ultimate triumph will hold immense practical and spiritual value.

China's journey in developing advanced AI computing power, from its inception to current breakthroughs, mirrors that of the Long March.

On January 13, 2025, a Western power announced further restrictions on the export of AI chips and technologies, aiming to retain advanced computing power within itself and its allies while seeking additional means to thwart China's acquisition of such capabilities.

Previously, they would justify their actions by including Chinese companies on the "Entity List," citing threats to US information security or impacts on the US technology industry.

The sinister aspect of this latest move is its candor; they are no longer disguising their true intentions—to monopolize advanced computing power globally and obstruct China's access to it.

This unprecedented measure marks a shift in the China-US technology war, escalating it from a "cold war" to a "hot war."

China's response has been equally tough. Besides routine objections and protests, it has emphasized a critical fact—when the US government announced a new round of export restrictions against China on December 2, we urged domestic enterprises to proactively counter these measures, prudently consider purchasing US chips, seek expanded cooperation with chip enterprises in other countries and regions, and actively utilize chips produced and manufactured by domestic and foreign-invested enterprises in China.

In essence, we are no longer disguising our stance. From past criticisms and appeals, we are now directly charting a new path for the industry—reduce reliance on US chips and actively embrace domestic alternatives.

Both sides are at odds. Yet, why are we particularly confident this time?

Superficially, it's because we hold our own cards; deeper still, this presents an exceptional opportunity for us to overtake on a different trajectory.

Honestly, if we could freely purchase advanced computing power infrastructure software and hardware through open channels, we might not have embarked on the path of independent innovation so swiftly and decisively.

Weak nations would have succumbed; China chose to confront the challenge head-on.

Consider our achievements in recent years. Beyond hardware breakthroughs, the evolution of the entire intelligent computing platform and the supply of intelligent computing power have advanced significantly.

It could be argued that without the US blockade, China's AI would not have accelerated its self-improvement and rise so rapidly. Sometimes, we need to thank our adversaries.

However, merely accelerating chip development is insufficient. The 10,000-card and 100,000-card clusters required for large model training represent the most complex and powerful high-parallel computing cluster forms in human history. Computing power is merely the tip of the iceberg; it also necessitates the support of storage capacity, communication networks, framework layers, model layers, and application ecosystems to achieve true high availability.

In simpler terms, solving the self-research and production of chips is akin to laying the "bricks and mortar" for constructing this formidable system. The truly functional large model-as-a-service platform (MaaS) is where our adversaries most desire to choke us.

So, what kind of platform do we need to withstand the US encirclement and blockade?

I believe there are at least three essential elements:

First, full-stack self-research, self-production, and self-sufficiency. From hardware, infrastructure, computing power clusters, architectures, to software applications, all must embody independent controllability.

Second, providing robust and stable system-level computing power, as Huawei Executive Chairman Xu Zhijun stated—Artificial intelligence is becoming the predominant demand for computing power, prompting structural changes in computing systems. We need system computing power, not just single-processor computing power.

Third, creating a platform deeply integrated with autonomous hardware and cloud services. Cloud computing is currently the most suitable and flexible form for users to leverage AI capabilities on demand. Therefore, it has become a critical requirement to enable users from various industries to harness AI capabilities with minimal technical barriers, assist customers in developing, training, hosting, and applying models through cloud services, and cultivate a "black soil" of hundreds of models and thousands of applications.

The US does not want us, or any country besides its close 17 allied nations and regions, to possess these capabilities. Essentially, the Americans have recognized that in the paradigm shift from the post-internet era to the intelligent era, AI capability is the most crucial development factor and even the core driver for leapfrogging advancements in human science and technology.

Therefore, besides its allies, restricting the AI development of middle powers and stifling the AI development potential of formidable opponents like China is a blatant strategy of the US. The US not only seeks to forge its own path but also to obstruct others' paths.

Yet, China also has its experts who never place the initiative of development in others' hands overly optimistically. Consequently, our current achievements extend beyond advanced chips.

Our two prominent advantages are:

1. The hardware and software integrated platform and ecosystem that provides advanced intelligent computing capabilities, models, and services can already achieve independent controllability, with some local innovations even surpassing those of the US.

2. The enthusiasm of China's various industries to embrace AI is unparalleled. Based on the practice of the aforementioned advanced computing power cloud-native platform, it has transitioned from "experimental implementation" to "practical application."

A prominent example is Huawei.

Currently, Huawei Cloud Ascend AI Cloud Service integrates large-scale computing power clusters, the computing engine CANN, the AI development framework MindSpore, the ModelArts AI development pipeline, and the ModelArts Studio large model-as-a-service platform. It can already provide stable and reliable full-stack computing power support for the training, inference, and development and operation of AI applications based on full-stack self-production and self-research.

Moreover, in a sense, the capabilities provided by Ascend AI Cloud Service are not merely "equivalents" but "superior substitutes," offering more targeted services that align with the needs and practices of the Chinese AI industry.

Those who truly understand trillion-parameter AI training know that one of the core indicators of a 10,000-card cluster is not just the number of GPUs but the duration of continuous, uninterrupted training.

To put it another way, when purchasing a car, you are not just concerned about its paper specifications; what truly matters is how many kilometers it can run without failure, which is the core experience indicator for buying and using a car.

When China first began building large clusters, interruptions of a few minutes to over ten minutes were common. Even now, the average uninterrupted duration for trillion-parameter model training is only about 2.8 days.

However, Ascend AI Cloud Service can achieve 40 days of uninterrupted operation; simultaneously, the average cluster fault recovery time in the industry is about 60 minutes, which Ascend AI Cloud Service can reduce to 10 minutes.

With the support of these advanced technologies, Ascend AI Cloud Service has been able to shorten the resource provisioning time for large models from months to days, accelerating the development of these models.

When China successfully tested its atomic bomb, a great leader remarked that a medal weighing one ton should be sent to Khrushchev because, had it not been for their withdrawal of experts, we would not have developed the atomic bomb independently; even if we had succeeded with their support, it would have been a debt hard to repay.

In other words, without the intense competition from the US in advanced computing power, we might not have possessed an advanced platform like Ascend AI Cloud Service so swiftly.

Another crucial factor for us is the enthusiasm of Chinese enterprises to embrace AI.

Transitioning from daring to use domestic products to truly using them and finding them "delightful" is the result of countless enterprises' trust and embrace of China's AI industry.

For instance, iFLYTEK, another influential enterprise in the field of large models, has proven through its practice over the past year or so, in collaboration with Huawei Cloud, that building an independently controllable and general large model base based on domestic computing power is entirely feasible. This viewpoint was validated on October 24 last year when iFLYTEK, in partnership with Huawei, launched China's first 10,000-card large model computing power platform, "Feixing No. 1," and based on this platform, iFLYTEK Spark 4.0 Turbo fully benchmarks GPT-4.

There are also more and more industry users. Huang Jing, Chief Cloud Service and Big Data Engineer at GAC Research Institute, shared a set of data—GAC has created GAC AI Smart Cockpit, GAC AI Smart Driving, GAC AI R&D Assistant, and GAC AI Large Model Platform based on large model technology. Among them, GAC AI R&D Assistant, relying on Huawei Cloud CodeArts, has built a code assistant that achieves automatic code writing and detection, improving R&D efficiency by over 10%.

Currently, Huawei Cloud AI has been implemented in over 30 industries and more than 400 scenarios.

Of course, having just one Huawei is insufficient. Internet industry giants such as Baidu, Alibaba, ByteDance, and Tencent are also providing robust computing resources to the market based on their respective strengths.

But this is just the beginning. China's intelligent computing power market is continuously growing. According to an IDC report, China's overall intelligent computing service market increased by 79.6% year-on-year in the first half of 2024, significantly outpacing the growth of traditional data centers; among them, the intelligent computing integration service market increased by 168.4% year-on-year, and the GenAI IaaS market increased by 203.6% year-on-year.

In AI development, China is the only Asian country in the quadrant of global AI development leaders. Simultaneously, it boasts advantages such as the largest single market scale and the largest total scale of single users, with tremendous potential for growth.

The AI competition may be the most crucial battle in the past 20 years, determining the development paradigm of the digital world for at least another 20 years. In a way, we have to thank the US.

Since ancient times, knowing war does not mean liking war, and the Chinese people dislike war even more. However, the Americans forced us to fight the War to Resist US Aggression and Aid Korea, a war that established our nation. The Soviets forced us to fight a crucial war that reshaped the strategic triangle between China, the US, and the Soviet Union.

Therefore, at this juncture, when the US emerges to push us, although their actions are aggressive, they inadvertently help us accelerate and complete the crucial piece of the puzzle for independently controllable advanced computing power. While our shortcomings in building AI innovation sources cannot be said to be completely filled, they can be considered fully adequate.

Now, Huawei Cloud's global storage and computing network, KooVerse, covers 30 geographical regions and 88 availability zones, while establishing core training and inference clusters around the three major data centers in Ulanqab, Wuhu, and Gui'an. Through the cloud, inference services can access corresponding AI computing power centers nearby, achieving an ultra-low latency premium service experience of 10ms, meeting the latency requirements of large model applications.

The emergence of this platform has completed the final piece of the puzzle for China's intelligent computing platform capabilities. Henceforth, we can fully sow, cultivate, and harvest indigenous AI innovations on our own soil, with enhanced controllability and security.

There is ongoing debate within the industry about whether China's AI capabilities, currently ranked among the top two globally, can ascend to the sole pinnacle. Judging by our current infrastructure development, this possibility is increasingly plausible within the next decade.

Only by firmly rooting ourselves as Chinese can we ultimately achieve global prominence.