Biden’s Grand AI Gift to China

cs_opinion_img
Before the United States realizes that sanctioning China further will harm its own interests, it will continue unwittingly aiding China.
January 22, 2025
author_image
Chinese finance blogger with a million followers
Click Register
Register
Try Premium Member
for Free with a 7-Day Trial
Click Register
Register
Try Premium Member for Free with a 7-Day Trial

On January 13th, local time in the United States, the Biden administration issued the “Interim Final Rule on Artificial Intelligence Diffusion.”

To be honest, after reading the contents of this report, I had two distinct impressions.

Firstly, I was ecstatic to see that the United States has expanded its scope of influence globally, a move that disperses their impact extensively and is strategically misguided.
They have categorized nearly 200 countries and regions into three tiers:

Tier one includes the United States and its 18 allies, such as Australia, Belgium, Canada, Denmark, Finland, Germany, France, French Guiana, Ireland, Italy, Japan, the Netherlands, New Zealand, Norway, South Korea, Sweden, Taiwan (China), and the United Kingdom. There are no restrictions on purchasing computational hardware from the United States. Interestingly, India is notably absent from this list, indicating restrictions on India’s acquisition of computational hardware from the United States.

Companies headquartered in the United States can deploy up to 25% of computational power in countries outside the first tier, with a limit of 7% in individual non-tier one countries.

Tier two consists of over 140 countries/regions, including Singapore, Mexico, India, Malaysia, Israel, the United Arab Emirates, Saudi Arabia, Portugal, Turkey, and others. Each country/region can acquire a maximum of about 50,000 AI GPUs (based on NVIDIA’s H100 GPU) from 2025 to 2027. It is noteworthy that Singapore, India, Israel, and Mexico are included in this tier, with Israel’s inclusion being particularly surprising.

However, the allocated quantity is significantly limited. The three-year allowance of acquiring 50,000 equivalent H100 GPUs falls short when compared to the purchases made by individual companies in the United States. For example, in 2023, Meta (formerly Facebook) alone acquired over 150,000 NVIDIA GPU H100 units. Companies like Microsoft made even larger purchases, while others such as Amazon, Oracle, Google, and Tencent acquired around 50,000 units.

This means that the cumulative purchases of these tier two countries over three years are far less than what a single American internet company bought in one year in 2023, essentially capping the computational power for these countries.

Furthermore, the report even specifies the quarterly purchase limits from the first quarter of 2025 to the fourth quarter of 2027.

The United States has provided three exceptions:

These countries and regions can enter into government agreements with the United States. If they agree to U.S. political conditions such as compliance with U.S. export controls and human rights requirements, the quota can be increased to 100,000 units.

Individual entities in these countries can purchase a maximum of 1,700 GPUs, in addition to the 50,000 limit. This suggests that the U.S. believes that even with 1,700 GPUs, a single entity cannot achieve much. The U.S. is likely to monitor and control these purchases, such as reviewing customer lists annually to prevent any attempts to circumvent the restrictions.

Entities headquartered in these countries can apply for “National Validated End User” (NVEU) status. If granted, they can purchase up to 320,000 GPUs over the next two years. This allowance seems to target specific large-scale enterprises, especially those based in tier two countries like Singapore, Israel, and Portugal.

Moreover, beyond GPU chips, the United States has also imposed restrictions on the export of advanced proprietary large language models (in contrast to open-source language large models) to tier two countries.

Large models contain numerous parameters, and the weights represent the significance of these parameters. The report mentions that “Model weights for advanced AI models can, moreover, be produced only by training the model on vast quantities of data using thousands of advanced computing ICs over a period of several months or more.” These weights are crucial for training, as poorly trained models can produce meaningless outputs.

Therefore, weights are crucial in preventing the export of advanced large models that could pose various threats to U.S. security, such as the development of nuclear weapons, biological weapons, and so on.

The third tier includes 22 countries and regions such as Mainland China (including Hong Kong and Macau), Iran, Belarus, Russia, Venezuela, Nicaragua, Syria, among others. These countries are almost entirely prohibited from importing advanced AI chips from the United States, and the export of proprietary large model weights is strictly forbidden as well.

Initially, I was pleased to see these U.S. sanctions as precise measures. For instance, in 2018, the U.S. first targeted ZTE and then turned its focus to Huawei. At that time, I believed this strategy was somewhat correct. I even thought that if the U.S. managed to bring down Huawei and halt its progress, China’s ICT industry development would significantly slow down.

Huawei was unparalleled in China’s ICT industry, especially in hardware. This was evident when Huawei gradually exited the smartphone market in the fourth quarter of 2020 and made a comeback by the fourth quarter of 2023. During these three years, no domestic smartphone brand had the capability to challenge Apple’s global position.

However, as time passed, it seems that Americans have forgotten the effectiveness of concentrating their efforts. Instead of solely targeting Huawei, they have added more and more Chinese companies to the Entity List, including DJI, iFlytek, Hikvision, Dahua, OFilm, SMIC, Naura, Cambricon, and hundreds of other Chinese enterprises. This dispersal of U.S. punitive measures has narrowed the market for American companies. For example, iFlytek, also on the Entity List, turned to purchasing Huawei’s Ascend chip hardware since they were unable to buy American hardware.

Since April 2023, iFlytek and Huawei have collaborated to establish a joint team and conducted a six-month joint research project in Hefei. By October 2023, they completed the world’s first Ascend ten-thousand-card computational power platform, named “Feixing No 1,” in Hefei. This platform can support the training of trillion-parameter large models domestically, making it the world’s first non-U.S. technology-based ten-thousand-card cluster platform.

This collaboration not only brought Huawei an estimated revenue exceeding 1 billion RMB but also provided iFlytek, as an AI large model manufacturer, with various technological feedback. This collaboration significantly aided Huawei in improving its Ascend chip and ecosystem software, thereby enhancing its capabilities.

In reality, this project was iFlytek’s first venture into a ten-thousand-card cluster platform with Huawei. However, if iFlytek had not been added to the Entity List, this cooperation might never have occurred.

Now, U.S. sanctions have expanded worldwide to approximately 160 countries and regions, further broadening the scope of restrictions and penalties. If the U.S. succeeds in diffusing its punitive measures globally, its impact will be dispersed worldwide.

Of course, the U.S. rationale is to prevent GPU cards from spreading to third-tier countries like China from second-tier nations, but it seems that in attempting to address one thorny issue, the U.S. has inadvertently introduced a larger one.

The second sentiment is that China’s opportunity has truly arrived as Americans have unexpectedly handed over the markets of around 170 countries and regions to us.

NVIDIA strongly opposed these sanctions following their announcement. Ned Finkle, NVIDIA’s Vice President of Government Affairs, stated that the extreme “national cap” policy would compel countries to seek alternative technologies, hastening the erosion of America’s technological leadership.

It appears to me that American politicians are increasingly disregarding democratic processes in their decision-making, focusing solely on the democratic form of voting during elections. In my opinion, this approach has two effects:

Firstly, it carries distinct traits of racialism and ideology. The first tier mainly consists of white-majority countries, with Japan and South Korea being semi-colonial territories where the U.S. has a strong military presence and significant influence. Taiwan, on the other hand, due to its semiconductor production capabilities and GPU manufacturing capacity, stands out. The second tier comprises predominantly Black, Arab, Latin American, Southeast Asian, South Asian, and other non-Caucasian nations.

In the third tier, countries like China, Russia, and Venezuela do not heed the United States’ directives. This tiered approach will further expose the fundamental white supremacist ideology of American elites.

Secondly, this is the first time in history that the United States has imposed such blatant restrictions on global production and technology, greatly stimulating a shift towards purchasing from China. While the U.S. has previously blocked technology access to most countries, such as with the military F35 aircraft limited to close allies, this strategic and expensive fighter jet’s inaccessibility might be understandable, as most countries cannot afford it and military aircraft are not productivity tools.

However, in sectors like GPU technology, which directly impacts national development capabilities, restricting sales to over 100 countries globally and curbing infrastructure development is unprecedented in my knowledge. Thus, if China can provide products to solve development issues for more than 100 countries and regions, why wouldn’t they turn to China? For China’s supply chain, it’s as if someone is handing us opportunities on a platter. For companies like Huawei and Cambricon in the Chinese computational hardware sector, the focus is currently on addressing the domestic shortage of domestically produced computational power, primarily due to manufacturing capacity constraints. If semiconductor manufacturing capacity issues are resolved, the opportunity for Chinese hardware and ecosystem to expand globally arises. Once established, the combined market of tier two and tier three countries will surpass that of the U.S. and its 18 allies, posing a significant threat to the U.S. semiconductor industry.

The current embargo has a 120-day buffer period, but it feels like the prelude to a major showdown. At present, domestically produced AI computational chips, led by Ascend, are in high demand, with even Cambricon, which has been in the red for years, showing significant improvement in its financial performance in 2024, possibly turning profitable in the fourth quarter. If there is overseas demand, domestic manufacturers will likely consider expanding into international markets, making 2025 potentially the inaugural year for Chinese AI chip hardware to go global.

I previously mentioned that observing the enhancement of domestic semiconductor manufacturing capacity could be seen through Huawei’s smartphone market share. In the third quarter of 2024, Huawei’s smartphone market share was 15.3%, a 42.0% increase from the same period in 2023, indicating a rapid improvement in domestic semiconductor production capacity. This growth trajectory is expected to continue in 2025.

According to IDC data, by the fourth quarter of 2024, Chinese smartphone manufacturers accounted for 56% of global shipments. If all Chinese smartphone manufacturers adopt domestic chips in the future, it would mean that domestic smartphone chips would account for over half of the global market. Furthermore, if over 100 countries and regions worldwide start purchasing Chinese AI chips in large quantities, it would signify that China’s AI chip market share could also exceed half of the global market, leaving the outcome of the China-U.S. competition without suspense.

For professionals in China’s semiconductor industry, time is of the essence!

Editor: Leo Cai

References
VIEWS BY

author_image
Chinese finance blogger with a million followers
Share This Post

Leave a Reply

Your email address will not be published.

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>

Comment
Cancel