tech

The United States strengthens control, intensifying the fragmentation of the AI chip market.

Sources reveal that U.S. officials have slowed down the issuance of licenses for chip manufacturers like Nvidia and AMD to ship large quantities of artificial intelligence (AI) accelerators to the Middle East, while also reviewing the AI development in the region.

According to media reports citing sources, it is currently unclear how long the review will take, and there is no specific definition of what constitutes "large-scale shipments." Sources point out that officials are particularly concerned about large sales due to countries like the United Arab Emirates (UAE) and Saudi Arabia wanting to import a large number of AI chips for data centers.

AI accelerators can help data centers process the massive amounts of data required for developing AI chatbots and other tools, making them essential equipment for businesses and governments seeking to build AI infrastructure. Currently, Nvidia holds a dominant position in this field.

Sources say that the purpose of slowing down exports is to give the United States time to formulate a comprehensive strategy on how to deploy advanced chips overseas, including coordinating who will manage and protect facilities used to train AI models.

Advertisement

The United States restricts Nvidia and AMD from selling AI chips to the Middle East, causing a significant drop in the stock prices of both companies. After the news was reported, Nvidia fell to an intraday low, closing down 3.77% at $1,105 per share, with Nvidia's market value evaporating $106.4 billion overnight. At the same time, AMD gave up earlier gains, ultimately closing up less than 1% at $166.75 per share.

Nvidia has previously stated on multiple occasions that the United States will upgrade its regulatory measures on Chinese AI, and once the new ban is implemented, it will cause the company to lose the opportunity to compete and lead in one of the world's largest markets.The U.S. Department of Commerce stated through a declaration that its top priority is to protect national security. As part of the ongoing discussions, Thea Kendler, the head of the Bureau of Industry and Security at the Commerce Department, visited the United Arab Emirates, Saudi Arabia, Qatar, and Kuwait earlier this month. A source familiar with the matter said that after her visit to the UAE, she indicated that progress had been made in cooperation on semiconductor export controls.

The U.S. concern is that, although Chinese companies generally cannot directly obtain advanced U.S. technology, they can use these chips through data centers in the Middle East.

Some insiders said that U.S. officials have delayed or not responded to shipment license applications submitted in the past few weeks. A source said that this includes U.S. companies planning to sell products to customers in the UAE, Saudi Arabia, and Qatar.

In addition to Nvidia and AMD, Intel and AI chip startup Cerebras Systems also produce accelerator chips. All four companies declined to comment.

The UAE and Saudi Arabia have been vying for regional leadership in the field of AI, hoping to reduce their economies' dependence on oil. Both countries regard the United States as a key partner in achieving this goal. Government officials and corporate executives have expressed that they will meet the requirements of the U.S. side.

In October last year, the U.S. Department of Commerce also imposed chip export controls on most of the Middle East. These control measures were originally aimed at China and other foreign competitors, meaning that companies must obtain special permission from the U.S. government to ship advanced semiconductor products and chip production tools to Middle Eastern countries such as the UAE or Saudi Arabia.

What impact will the U.S. upgrade of AI technology export controls have?

On May 8, a bipartisan group of members of the U.S. House of Representatives unveiled a bill. If passed, the bill would make it easier for the Biden administration to implement export controls on AI technology.

The bill, named the "Enhancing National Frameworks for Overseas Restriction of Critical Exports Act" (ENFORCE Act for short), states that it aims to restrict the export of U.S. AI systems to prevent foreign competitors from using American AI technology.The bill mentions that "AI systems" refer to all software and hardware related to AI, including AI models and all numerical parameters related to the implementation of AI technology.

On May 9, the Chinese Embassy in the United States responded to this, saying that this is a "typical act of economic coercion and unilateral bullying, which China firmly opposes," and added that it will take "necessary measures" to protect its own interests.

At present, China's large model industry still relies on American technology companies. In terms of computing power, large models mainly use computing power for training and inference. The training process is to continuously adjust parameters through a large amount of data to enhance the capabilities of large models, so higher performance computing power is required. The inference process is more about verification, inputting new data or tasks, and obtaining results to verify the capabilities of large models. The inference process pays more attention to the low latency characteristics of computing power, and the requirements for performance are relatively lower.

Many Chinese technology companies will handle the training and inference processes separately to optimize costs and efficiency. Usually, the training process will use the cloud services of American companies, and the inference process will use domestic computing power.

It is understood that despite previous cloud service control requirements, there are still many Chinese companies using American company cloud services or choosing to build data centers overseas to train AI large models. However, this regulatory signal has had a direct impact on the industry. A person from an American cloud service provider said that the industry feels uncertain, so even if the policy is not completely tightened, it will still make some Chinese companies hesitate when choosing American cloud manufacturers to train large models, and thus give up.

A founder of an AI startup company said that currently, China's GPU clusters still have some gaps with Azure and AWS, and it is difficult to catch up in a short time. In addition, data control related to large models may also have an impact. At present, it seems that high-quality knowledge-based data is mostly English data. "China's large model industry needs to rethink the technical path, otherwise it is difficult to avoid being sanctioned."

In terms of open source models, many Chinese technology companies' models are based on LlaMA released by Meta in February 2023 and LLaMA2 released in July 2023.

On April 18 this year, Meta released two open source models - Llama 3 8B and Llama 3 70B, which are free for external developers to use. Many people in the industry have commented that this is the most powerful open source model so far, and developers can make large model products comparable to the level of GPT-4 based on this.

In addition, some large models in China are trained using data generated by GPT-4. A technical director of a major AI company believes that the principle of this method is to train low-order models with high-order models. Large model training requires high-quality data, and this method can reduce data processing costs, "most of the data generated by GPT-4 is higher than the quality of online databases."

Therefore, if the new bill is implemented, adding restrictions to all large models in the United States will directly increase the cost and reduce the efficiency of some domestic technology companies in training large models.A representative from Microsoft Research Asia stated that an export ban on AI would not completely seal off the export of large models. Just as "resellers" made a fortune by reselling Nvidia chips after the chip ban, this ban may also give rise to a group of people who become wealthy by transporting large models. He believes that while policies have a real impact, one should not ignore the fact that some people are deliberately exaggerating the impact of the AI competition between China and the United States. "There are too many people making money by creating opposition and anxiety."

The industry generally believes that if the United States intensifies restrictions on AI exports to China, the gap in AI technology development between China and the United States will become increasingly large. A divided market is not only not conducive to the overall development of AI technology, but also makes it difficult for companies to conduct business and further exacerbates the gap between the supply chains of China and the United States.