Huawei is preparing to significantly increase the production of its most advanced artificial intelligence chip next year, aiming to win customers in the world's largest semiconductor market, while its competitor NVIDIA is struggling with geopolitical headwinds.

According to a report by Bloomberg citing sources, Huawei plans to produce about 600,000 of its flagship chip, Ascend 910C, next year, doubling this year's output. Due to U.S. sanctions, Huawei will have difficulty bringing these products to the market for most of 2025. According to the sources, in total, the Shenzhen-based company will increase the output of its Ascend product line to 1.6 million bare chips in 2026. A bare chip refers to the basic silicon element that contains the circuitry of a chip.

The report states that if Huawei can achieve these goals, it would represent a breakthrough in technology, as Huawei is seen as the biggest hope for China to reduce its reliance on foreign chips, which are the engines driving artificial intelligence development in the world's second-largest economy.

This indicates that Huawei and its main partner SMIC have found a way to alleviate some bottlenecks that not only hindered its AI business but also impeded China's goal of achieving self-reliance. The sources said that the predictions for 2025 and 2026 included the number of bare chips in Huawei's inventory, as well as internal estimates of yield or defect rates during the production process.

From Alibaba to DeepSeek, Chinese companies need millions of AI chips to develop and operate related services. It is estimated that NVIDIA alone sold 1 million H20 chips tailored for the Chinese market in 2024.

In September this year, Huawei publicly released its three-year vision aimed at weakening the dominance of competitors, breaking its previous tradition of keeping its plans and capabilities confidential.

Huawei's rotating chairman Xu Zhijun released a series of Ascend chips - 950, 960, and 970 - planned to be launched in phases until 2028. The sources said that these chips are a natural evolution of the long-standing Ascend 910 series, which remains Huawei's cash cow. A company representative has not yet responded to the request for comment.

Huawei has led a wave of Chinese chip companies competing to develop accelerators, while NVIDIA has been largely excluded due to government controls. Chinese regulators recently accused NVIDIA of violating the Anti-Monopoly Law, causing several tech giants to stop purchasing its AI chips designed for the Chinese market.

NVIDIA founder Huang Renxun has tried to assure Chinese authorities of the safety of its products, but it is unclear when or if this will happen. NVIDIA said in its recent earnings call that sales of its H20 chips tailored for the Chinese market were zero last quarter.

The sources said that despite reports that Chinese companies are trying to triple their semiconductor output next year, this target is not realistic. They pointed out that one issue is the yield rate, the proportion of chips available after they are produced, which is not satisfactory.

Huawei's current top artificial intelligence processor packages two bare chips into a single chip set - an advanced packaging technology that theoretically enhances product performance. However, the sources said that this process is extremely challenging and has contributed to ongoing supply shortages, allowing competitors like Cambricon to fill the gap.

But the sources added that Huawei successfully increased its output this summer. Over the past year, Chinese companies including Shanghai Micro Electronics (SMEE) have made significant progress in gear technology. The sources said that Huawei currently plans to launch a chip in late 2026 to replace the Ascend 910C, initially called the Ascend 910D.

They added that Huawei's goal is to achieve a sales volume of 100,000 units for this new chip set, which uses a more aggressive design that compresses four bare chips into a single chip set. Huawei said this month that it plans to release a chip named 950DT in late 2026.

The sources said that Huawei will deploy approximately 1.6 million bare chips across two chips next year, and 1 million bare chips in 2025. According to the sources, Huawei informed its customers earlier this year that, in addition to reserving about 100,000 chips for its own cloud computing department and some possibly government-related projects, the company is prepared to sell 200,000 Ascend 910C chips by the end of the year.

Currently, Huawei's artificial intelligence chips lag behind NVIDIA in single-chip performance. According to Bernstein, the performance of the upcoming Ascend 950 is only 6% of NVIDIA's next-generation VR200 super chip. So far, Huawei's major clients, including Alibaba and Tencent, have mostly used their best semiconductors for inference, i.e., running artificial intelligence models after training.

With the help of SMIC, Huawei is able to produce chips using outdated 7nm technology. The bare chips of the 910 product line use an enhanced version of SMIC's 7nm technology. In contrast, NVIDIA's latest Blackwell architecture graphics processing units (GPUs) are produced by TSMC at a 4nm node - about two generations ahead.

The report stated that Huawei is exploring next-generation Ascend chips beyond its currently released models, which may be an unknown quantity, planned to use SMIC's 7nm production line with better performance. According to one source, the performance of this new chip may be sufficient for major clients to train artificial intelligence algorithms, not just run them.

However, the source said that technical challenges in manufacturing could delay the commercialization of the product to 2027 or later. Huawei stated in a public speech at the beginning of this month that it plans to launch a chip named Ascend 960 within this time frame and another chip named Ascend 970 in 2028.

This month, Xu Zhijun proposed a series of solutions to overcome technological bottlenecks: brute force cracking, networking, and strategic support. One of the technologies he outlined allows Huawei to connect up to 15,488 Ascend series AI chips using its self-developed, super-node-oriented interconnection protocol UnifiedBus. This new technology was also officially launched on the same day.

Sources: RFI integrated reports from media such as Bloomberg

Original: https://www.toutiao.com/article/7555644358226100786/

Statement: This article represents the views of the author. Please express your opinion by clicking the 【up/down】 buttons below.