How China Can Break the Monopoly of AI Chips

[baoshi.rao]

After the ban on NVIDIA's A100 and H100 chips to China, the U.S. government has issued another round of chip export restrictions.

On October 25, the U.S. Securities and Exchange Commission (SEC) disclosed on its official website that the GPU chip export ban mentioned in NVIDIA's regulatory filing would take immediate effect.

Notably, this move occurred against the backdrop of improving U.S.-China relations. The ban was initially announced by the U.S. Department of Commerce on October 17 with a 30-day notice period, but this brief window was abruptly closed just eight days later.

As a result, NVIDIA's A800/H800 chips, which had seen skyrocketing prices due to frantic purchases by domestic AI companies, are now completely cut off. Other major manufacturers, such as AMD's MI250 chips and Intel's Gaudi2 data center products, are also prohibited from being sold to Chinese customers.

The new export controls impose specific requirements on chip performance density, the scope of export destinations, and equipment licensing for chip manufacturing. Additionally, 13 Chinese GPU companies have been added to the 'Entity List.'

The intention is clear: to use every possible means to stifle China's AI development, much like the 'Sophon' blockade in The Three-Body Problem that halted human technological progress.

The foundation of AI is computing power, and the foundation of computing power is chips. Faced with increasingly severe chip sanctions, China's AI industry has only one path forward: domestic substitution.

01

The New Ban

A comprehensive and all-encompassing blockade.

"How can one tolerate others snoring beside their bed?"

This was said by Zhao Kuangyin, the founding emperor of the Northern Song Dynasty, to Xu Xuan, a debater sent by Li Yu, the last ruler of the Southern Tang Dynasty.

Today, the U.S. government holds a similar mindset. Witnessing China's rise in high-tech industries such as power batteries, new energy vehicles, and photovoltaics, the U.S. cannot allow China to challenge its dominance in AI, the pinnacle of future technology.

In 2018, ZTE was banned by the US; in 2020, Huawei faced sanctions, rendering its Kirin high-end chips obsolete; in October 2022, the US targeted the AI chip market, prohibiting the supply of Nvidia's top-tier A100 and A800 chips to China.

A year later, the US Department of Commerce has once again wielded its regulatory power, banning even the 'alternative versions' of Nvidia chips supplied to China. Compared to last year, the updated regulations have further intensified, aiming to close all potential loopholes:

1. Performance Density Replaces Bandwidth Parameters: Chips with a speed of 300 TFLOPS (trillion operations per second) or higher are banned. Chips with speeds between 150-300 TFLOPS and a performance density of 370 GFLOPS (billion operations per second) per square millimeter or higher are also prohibited.

2. Expanded Export License Requirements: The scope now covers over 40 countries and regions to prevent these chips from entering China through third-party routes.

3. Restrictions on Chip Manufacturing Equipment: Licensing requirements are imposed on equipment for 21 countries, expanding the list of banned devices and limiting China's ability to produce advanced chips below 14nm.

4. Blacklisting Chinese GPU Companies: 13 Chinese GPU firms, including Biren Technology and Moore Threads, are added to the 'Entity List,' severing their ties with TSMC.

According to estimates, mainstream AI chips on the market are effectively blocked under these performance requirements. Nvidia's A100, A800, H100, H800, L40S, AMD's MI250 series, and Intel's Gaudi2 are all affected.

Consumer-grade RTX 4090 GPUs and lower-performance L40 chips are not included in the 'sales ban,' but companies must report exports of chips exceeding 300 TFLOPS to prevent their use in AI model training.

Additionally, the US has abused its long-arm jurisdiction, warning chip foundries that exports of chips with 50 billion or more transistors and high-bandwidth memory require special attention and licenses for shipments to China.

For the 13 blacklisted companies, any foundry work must obtain approval from the US Bureau of Industry and Security (BIS). Without access to high-end foundries, these companies' chips remain theoretical.

The US is particularly vigilant about high-end chip equipment like lithography machines to prevent their transfer to China. While EUV lithography machines were already banned, the new restrictions also target ASML's DUV lithography systems, preventing the shipment of older DUV models and parts to Chinese chip factories.

Reviewing Nvidia's product lineup, only the V100 seems to have escaped the ban. However, with a maximum speed of 125 TFLOPS and a communication rate of 300 GB/s, its performance is inadequate for current AI applications.

The U.S. Commerce Department's new regulations aim to comprehensively block China's AI industry through 'performance + channels + entity lists + long-arm jurisdiction,' leaving no room for escape. Like iron bars, they stand between China's high-tech vision and the real market demand.

02

Computing Power Hegemony

Endless suppression and blockade. As Meng Wanzhou said, the scale of computing power determines the speed of AI iteration and innovation, as well as the pace of economic development.

Globally, the U.S. dominates the top of the high-end chip pyramid and uses state power to control the supply chain, courting companies like TSMC to isolate China.

The U.S. can act arbitrarily due to its strong technological prowess. Data shows that the training computing power consumption of ChatGPT-3 is enormous, reaching 3640 PF-days (equivalent to computing at 10^15 operations per second for 3640 days), which equals 6,000 Nvidia A100 chips. Factoring in interconnection losses, 10,000 A100 chips are needed as a computing power foundation.

In other words, for Chinese AI companies to keep up with OpenAI, they need at least 10,000 A100 chips as a baseline. GPT-4 requires training on approximately 25,000 A100 chips for 90 to 100 days. As for the next-generation model, GPT-5, Elon Musk estimates it may require 30,000 to 50,000 H100 chips.

And this is just the training phase before market launch. For operations, taking GPT-3 as an example, maintaining 50 million daily active users requires 16,255 A100 chips.

Thus, in the booming large-model movement, Chinese companies' demand for Nvidia chips has surged. The Financial Times previously reported that Chinese tech giants like Alibaba, Tencent, ByteDance, and Baidu placed $5 billion in orders for A800 chips from Nvidia, expected to be delivered this year and next. But now, these hopes have been dashed.

Today, the 'entry ticket' to the AI world is firmly in the hands of U.S. companies like Nvidia, which currently holds 82% of the global data center AI accelerator market and a staggering 95% share in AI training. This dominance may be even more formidable than the 'Wintel alliance' of the PC era or the Qualcomm + Google (Android) duo of the mobile era.

The highly oligopolistic market has propelled Nvidia to become the biggest winner in the AI battlefield, with its market capitalization exceeding $1 trillion, making it the world's most valuable chip company. Its performance has also surged dramatically, reaching historic highs.

On one hand, Nvidia, AMD, and Intel dominate the AI chip market. On the other hand, given the globalized division of labor in the semiconductor supply chain—especially the reliance on TSMC for manufacturing—the U.S. has pressured TSMC to build factories in America.

Meanwhile, the U.S. has arbitrarily revised trade access rules. Last year's regulations prohibited products with computing power exceeding 600T from being manufactured by TSMC. This effectively restricted China's AI chip technology to the 7nm threshold, though Chiplet packaging technology could still be used to interconnect multiple chips for higher performance. However, the newly enacted ban, which uses performance density as a restriction criterion, aims to close this loophole.

It's clear that the U.S. is employing every possible method to maintain its technological dominance. The moment it perceives a challenge to its position, it resorts to relentless suppression and blockades.

With the new restrictions, Nvidia has also suffered. A $5 billion order is likely to fall through, and its stock price dropped nearly 12% from October 17 to October 31 after the new rules were announced.

It's worth noting that about 50% of Nvidia's revenue comes from data center chips, with China contributing 21.4% of its income. The Chinese mainland has long been its largest market outside the U.S.

Breaking the Iron Bars

Domestic AI chips are not unusable—they're just not as good.

With the new export controls in place, China's AI chip industry has no choice but to become self-reliant.

Fortunately, Huawei's breakthrough in 5G chips has given Chinese companies hope. Despite heavy blockades, Huawei managed to carve out a path, proving that China's chip industry has made significant progress and can find alternatives without U.S. technology.

Previously, there was excessive reliance on Nvidia. Now, with the U.S. tightening its grip, nurturing a domestic supply chain has become the only way forward.

But this isn't as simple as replacing imported chips with domestic ones. As the saying goes, 'a slight move in one part can affect the whole situation.' It requires top-level policy support, market-driven operations, and, most importantly, leaps in foundational technologies across the entire industry chain.

Currently, domestic contenders in China primarily consist of three types of enterprises: first, large tech companies like Huawei and BAT (Baidu, Alibaba, Tencent); second, state-backed enterprises such as Sugon and Hygon; and third, private and startup companies including JingJia Micro, Biren Technology, Cambricon, Moore Threads, Tianshu Zhixin, and Enflame.

In terms of performance, Huawei's Ascend 910 and Hygon's Deep Computing No. 2 can directly compete with NVIDIA's A100, reportedly achieving about 80% of its performance. Liu Qingfeng, Chairman of iFlytek, even stated that Huawei's GPU technology is now on par with NVIDIA's A100, claiming that the Spark model developed based on it can rival or even surpass ChatGPT.

Among startups, Moore Threads is led by Zhang Jianzhong, former General Manager of NVIDIA China; Muxi's C100 is expected to compete with NVIDIA's H100, while Biren's BR100 is said to surpass the A100.

However, these products face several systemic issues. First, with TSMC's manufacturing blocked, they can only rely on SMIC's 7nm process. Second, there's the ecosystem problem. Although Hygon is the only one among these companies to achieve significant profitability and market success, its reliance on AMD's x86 architecture license is insufficient to compete with NVIDIA.

Third, paper specifications ≠ practical application. Domestic AI chips are not unusable but are not user-friendly. In June this year, some local governments proposed purchasing domestic AI chips for computing center construction but were rejected by contractors, including state-backed enterprises, due to inferior performance and usability compared to NVIDIA's chips.

Some local governments or companies, responding to the call for domestic substitution or to support local products, purchased small quantities of domestic chips, but these often end up unused, gathering dust in warehouses.

Moreover, due to late starts, slow iteration, and low efficiency, even if cheaper, domestic AI chips incur significant trial-and-error costs when entering the market. While Moore's Law is becoming less effective in today's semiconductor industry, products still undergo upgrades. The lack of large-scale practical applications for domestic products means they miss out on positive feedback, slowing their development pace.

Once pioneers establish an ecosystem monopoly, latecomers have little chance of success, as seen in the PC and mobile eras.

This creates a persistent negative cycle. Breaking this deadlock and revitalizing the market ecosystem is urgent.

Huawei has begun laying the groundwork, with its Ascend ecosystem covering hardware, software, and more. It collaborates with local governments, enterprises, universities, hospitals, and research institutions in key projects like the East-West Computing Initiative to utilize computing power.

This is a crucial move. Without application and promotion, even the best products are merely for display.

Final Thoughts

After the new regulatory measures were announced, Jensen Huang's response was thought-provoking:

NVIDIA will comply with the relevant regulations... We will also strive to support market demands and serve customers in mainland China. There are many excellent technology companies in China, such as Huawei, and NVIDIA must work hard to compete with the local industry while observing the outcomes.

His words reflect a sense of helplessness toward the U.S. government's hawkish stance, lingering attachment to the Chinese market, and respect for China's outstanding tech companies—a complex and contradictory mindset.

However, Huang's sentiments alone won't suffice. For China's AI development, self-reliance is the only viable path.

Learning from successful models is nothing to be ashamed of. The strength of the U.S. tech industry stems from comprehensive prosperity across all levels—OpenAI's R&D and products, NVIDIA's chips as a backbone, strong support from ally TSMC, Microsoft's investments and applications, a vast talent pool, and substantial capital—all key to its success.

Currently, China is not short on capital or market potential, and its talent pipeline is steadily growing. With Huawei's breakthrough in 5G as a precedent, achieving similar success in AI chips is not impossible.