After the Trump administration recently banned, as part of the escalating trade war between the superpowers, Nvidia from continuing to export any type of AI chips for data centers to China, a local player is trying to fill the void and keep China strong in the intensifying artificial intelligence race. According to various reports, Huawei is currently accelerating the production of a new AI chip, Ascend 910C, designed to serve as an alternative to Nvidia's powerful chips.
Huawei has already begun supplying samples of the new chip to local cloud providers and AI companies like Alibaba and Deepseek, and aims to produce hundreds of thousands of units of the new chip this year. Huawei's effort to provide an alternative to Nvidia is not just a matter of commercial competition in the AI infrastructure market, but an issue of national-strategic importance for China in its pursuit of technological independence and in its struggle against the United States for AI supremacy. Some are asking whether the American restrictions not only failed to constrain China's progress but actually transformed it into an equal superpower.
Huawei has not officially and openly announced the new chip, and reports about its performance are somewhat contradictory. According to some sources, apparently coming from within Huawei, the new chip provides processing and memory capabilities equivalent to Nvidia's previous generation chip, the H100. This is a dramatic claim, but should be taken with a grain of salt. According to a recent report in the technology blog Tom's Hardware, researchers from Chinese AI company Deepseek recently examined the chip and found it lacking in model training capabilities, running AI models "at a rate of 60% compared to Nvidia's H100."
The engineering structure of the chip also reveals Huawei's development challenges. The 910C is a kind of "hybrid creation" that connects two previous generation 910B chips in a special package. In other words, unlike Nvidia, which offers a new breakthrough with each new generation, the 910C is a kind of "improvisation."
Another question mark concerns Huawei's ability to produce the chip in quantities that will satisfy the enormous thirst among Chinese cloud and AI companies, now that Nvidia's supply pipeline to the Chinese market is blocked. Manufacturing such advanced chips requires an entire ecosystem of innovative technologies, and in recent years the tightening American embargo has made it increasingly difficult for China to gain access to this essential equipment.
China's vulnerability lies in two critical links in the chip manufacturing process: American restrictions block China's access to two suppliers, Taiwan's TSMC and the Netherlands' ASML, without which advanced AI chips like Nvidia's cannot be produced. ASML is the only company in the world that supplies EUV lithography machines. These machines, using advanced optics, enable "carving" tiny transistors as small as 7 nanometers and below on silicon wafers. ASML's CEO recently estimated that "denying these machines to China will perpetuate a 10-15 year lag behind the West."
Chinese manufacturers are also failing to close the gap with TSMC. For illustration, TSMC is expected to begin production of 3-nanometer transistors this year, and even unveiled 1.4-nanometer production about a week ago, to be launched in 2028. In contrast, China's SMIC, which manufactures Huawei's chips, is still "stuck" at 7 nanometers, which is the process Nvidia used to manufacture its AI chip in 2020.
The jump from 7 nanometers and below is complex, and China will struggle to accomplish it without access to equipment suppliers like ASML and testing equipment manufacturers like Israel's Nova. The smaller the transistors, the more efficient and powerful chips can be developed and manufactured. This means that by the end of the decade, Nvidia's chips could be thousands of times more powerful than Chinese alternatives, which would decide the AI battle in favor of the United States.
At a recent meeting between China's ruler, Xi Jinping, and Huawei founder Ren Zhengfei, the latter tried to convey a reassuring message, claiming that the Chinese chip industry's dependence on the West is decreasing, and that Huawei is promoting cooperation with thousands of Chinese companies to achieve by 2028 about 75% independence in the supply chain for chip development and production. Indeed, satellite images reveal that Huawei is establishing three new advanced manufacturing facilities in China.
China's scientific community is also trying to contribute its part, and recently two groundbreaking studies from Chinese universities were published in the journal Nature in the field, one dealing with the development of ultra-fast "flash" memory that can accelerate AI processing in data centers, and the other on an efficient and fast transistor not based on silicon. If these discoveries mature commercially, it could advance China a big step forward in its technological independence.
Last week, Nvidia founder Jensen Huang said that contrary to popular belief, China is not lagging behind the Americans in the field of AI at all. Indeed, about half of the world's AI researchers today are Chinese, and in 2024 Chinese companies and researchers registered more than 13,000 patents in the AI field, more than twice as many as their American counterparts. However, American patents are cited 7 times more often and American AI companies launched more than 40 new foundation models in 2024, compared to 15 models launched in the Chinese ecosystem.
In addition to its research investment, China also does not hesitate to use means to circumvent American sanctions, and according to Western claims, uses shell companies, smuggling, and exploitation of lax enforcement. For example, an advanced processor manufactured by TSMC was recently discovered in a Huawei chip. The processor reached Huawei through a third company that purchased it from TSMC and transferred it, in violation of export restrictions, to Huawei. Additionally, The Wall Street Journal recently reported that smuggling networks are successfully smuggling Nvidia's new "Blackwell" processor into China.
These loopholes in regulations and enforcement have only led the American administration to tighten restrictions. Just days before the end of his term, the Biden administration issued very strict regulations limiting the ability of American technology companies to sell advanced technologies to a very wide range of countries, including Israel, to prevent these technologies from leaking to China. These regulations drew much criticism from the chip industry, including from Nvidia, claiming they would make it difficult for American technology companies to maintain proper trade relations with many countries around the world. Today it was reported that the Trump administration is expected to cancel these regulations and offer an alternative framework, and time will tell if it succeeds in addressing the phenomenon without further burdening American companies.
The defining event that made it clear to everyone that China is breathing down America's neck regarding AI was, of course, the appearance of Chinese AI company Deepseek. In January, on Trump's inauguration day, Deepseek, which until then was completely unknown, launched a thinking model, R1, that was equivalent in performance to the thinking models of American AI companies like OpenAI and Google – and it managed to do so using vastly reduced processing resources.
Deepseek's surprising appearance on the AI stage prompted two types of responses. On one hand, some were quick to dismiss Deepseek as a product of intellectual property theft and communist propaganda, while on the other hand, some rushed to declare that American sanctions had failed and only spurred China to become an AI superpower. It seems that both claims are lacking. Although Deepseek used controversial methods, its algorithms offer significant improvement in the efficiency of training and running models, and indeed today AI companies around the world are using the same methods to optimize their models.
However, contrary to the prevailing assessment that Deepseek's new efficiency methods would reduce the demand for processors, they have only increased it. The reason is simple: the ability to develop and run AI models with fewer processing resources just makes AI technology more available and accessible to more players, not just technology giants with billion-dollar pockets. Indeed, Deepseek CEO Liang Wenfeng said in an interview with local media, "Money has never been a problem. The ban on technology exports is the problem."
