Frontier | Chou Xiaoxin of AXERA: High-computing-power chips will become the main growth engine for enterprises next year

In today's automotive industry, where cost reduction is increasingly emphasized, the balance in the minds of car manufacturers is gradually tilting towards domestic chips with higher cost - effectiveness. Aixin Yuanzhi has spotted this development opportunity. Before the Beijing Auto Show, Aixin Yuanzhi, together with companies such as Qianli Technology and Jieyue Xingchen, established the "Qianli Alliance". Meanwhile, the company will release the M97, a high - computing - power assisted driving chip, in the third quarter of this year.

Qiu Xiaoxin, the founder of Aixin Yuanzhi, publicly stated at the Beijing Auto Show that "the high - computing - power product has successfully completed tape - out and is entering the normal development cycle". It will be gradually introduced into multiple vehicle models in 2026, becoming an important engine for performance growth. Currently, traditional terminal computing products still account for over 80% of Aixin Yuanzhi's revenue, while the revenue from in - vehicle and edge computing businesses accounts for a relatively low proportion. Qiu Xiaoxin expects that this situation will change in the next three years.

Qiu Xiaoxin is quite confident in the performance of M97. She said, "The current mainstream high - level assisted driving chips have a relatively big problem, which is insufficient bandwidth. After we realized this problem, we significantly increased the bandwidth when designing the M97."

The effective computing power of an assisted driving chip is determined by two factors. One is the number of computing units, and the other is the bandwidth, that is, whether the data reading and writing speed is fast enough. "Even if the computing power is 2000T, if the DDR bandwidth is insufficient, the chip cannot fully utilize its 2000T computing power. A significant advantage of our chip is that it has a very high DDR bandwidth, which can fully utilize the computing power. At the same time, our chip has good power consumption control, which is also an advantage."

In order to fully utilize the computing power of the chip, Aixin Yuanzhi "made it clear during the design phase to be one - generation ahead. The company also selected a technology one - generation ahead of the current mainstream chips in the market."

The consideration behind this is that the chip manufacturing process has a significant impact on performance. Take the DDR frequency as an example. "For a 7 - nanometer chip, the maximum frequency can reach 6400, and the bandwidth is limited by this 6400. If it is a 5 - nanometer or 4 - nanometer chip, it can reach frequencies of 8533, 9600, or even over 10,000. The higher the manufacturing process of a high - computing - power chip, the better its performance. After the manufacturing process is improved, the chip area will be significantly reduced. When the chip area is reduced, the yield rate will also be significantly improved."

As for why Aixin Yuanzhi designed the computing power of M97 to exceed 700T, Qiu Xiaoxin thought like this: Currently, the evolution direction of the assisted driving technology route is still uncertain, but from end - to - end to the current VLA and world models, they are all about understanding the surrounding physical world.

An important test for the chip is computing power. "So, the M97 starts by reaching the benchmark level of computing power and also increases the DDR bandwidth. In fact, this is to provide sufficient effective computing power to support more advanced large models such as VLA in vehicles. In the past, it was end - to - end + VLM. VLM is used for various semantic understandings and runs in parallel with the end - to - end model. VLA attempts to combine them into one." Qiu Xiaoxin believes that high - computing power is the inevitable evolution direction of chips. "It won't work well to run this on two separate chips. The computing power on the same chip must be high enough."

While increasing the computing power, Aixin Yuanzhi also emphasizes cost control. The cost of the chip is reduced through design. "For example, in the general architecture, it inherits a set of architectures from cloud - based training all the way to in - vehicle chips, and the entire architecture design has a lot of redundancy. This is not necessary for in - vehicle chips."

A company's performance growth depends not only on products but also on strategies. Aixin Yuanzhi positions itself as "a relatively neutral third - party chip supply platform". Qiu Xiaoxin explained that the advantage of this positioning is that "at this time, car manufacturers can choose us or others. The most important thing is to give the decision - making power back to the car manufacturers." "Having choices is still very important", and "as an independent chip company, Aixin Yuanzhi needs to maintain a neutral and independent state."

In addition to the in - vehicle field, edge computing is also an area where Aixin Yuanzhi plans to make significant investments. In the second half of this year, Aixin Yuanzhi will release two edge computing products to better adapt to mainstream large models such as Qianwen in the market. Qiu Xiaoxin told 36Kr that building an ecosystem is very important for adapting to large models. One of the ways Aixin Yuanzhi builds the ecosystem is that "whenever others open - source a large model, we will immediately deploy the large model on our chips. After deployment, we will open - source it again and make it available on GitHub for partners to download. Since the underlying architecture of most large models is currently Transformer, this is beneficial for us to deploy various models as the underlying operators are relatively unified."

Regarding the application scenarios of edge computing chips, Aixin Yuanzhi has observed that AI Agents will be an area where the company can make great achievements. "In the future, when AI becomes more popular, water, electricity, gas, broadband, and computing power will become necessities in people's lives. How to connect this computing power into homes requires an AI Agent box, which may be part of a router."

Since the Agent has access to all the user's core information and needs to run 24/7, it must run locally. If it is connected to the cloud, there may be risks such as latency and privacy leakage. Aixin Yuanzhi's 8850 chip and the upcoming edge computing products can support large models to run on the edge side. Qiu Xiaoxin commented that AI Agents may be a ToC product with strong explosive power, and Aixin Yuanzhi is full of expectations for this field.