Ascend

Ascend is a family of computer chips from the Chinese technology conglomerate Huawei. These chips are not intended for ordinary office work, but for the enormous amounts of computation that artificial intelligence requires. When a language program like ChatGPT learns or generates a response, billions of simple calculation steps run behind the scenes. Ascend chips are precisely tailored to this. They sit in servers in large data centers, meaning halls full of computers that are used over the internet. The name stands for an entire series: the best-known models are called Ascend 910B and Ascend 910C.

The comparison almost always comes down to the same company: Nvidia. The US company dominates the global market for AI chips. Ascend is an attempt to break this dependency.

China’s answer to the Nvidia monopoly

Since 2019, the US has banned the export of its best AI chips to China. The reason is political: Washington does not want China to catch up militarily or economically with this technology. Chinese companies are therefore not allowed to buy Nvidia’s most powerful models at all. This deprives them of exactly the tool used to build modern AI.

Ascend fills this gap. For Chinese internet companies, universities, and government agencies, it is often the only available option in this performance class. That is why these chips regularly appear in economic news. Every report about a faster Ascend is also a report on how well US export controls are still working.

This is relevant for investors because Nvidia’s revenue in China depends on it. If Nvidia loses customers there to Huawei, it loses out on a market worth billions. Conversely, every setback for Ascend shows just how hard it really is to replicate these chips.

What’s inside an Ascend chip

AI calculations consist almost entirely of multiplying and adding large tables of numbers. An Ascend chip contains a very large number of simple computing units for this purpose, which work in parallel. Huawei calls this design Da Vinci. You can picture it like a factory hall with thousands of identical workstations. Each individual one is unspectacular, but together they accomplish enormous amounts of work.

A second component is the memory located directly next to the compute chip. It must supply numbers quickly, otherwise the computing units sit idle. This fast specialized memory is a weak point for Ascend, because worldwide it is produced by only a few companies and is likewise subject to export restrictions.

Then there is the software. With CUDA, Nvidia has had a programming environment for years that virtually every AI developer knows. Huawei’s counterpart is called CANN and is considerably less widespread. A chip is only as useful as the programs that run on it. This is where the bigger gap actually lies, not merely in raw computing power.

Ascend in headlines and products

You won’t be able to buy an Ascend chip directly. It is built into server systems, such as Huawei’s Atlas line, and sold to data centers. Anyone using Chinese cloud services may well be computing on Ascend without realizing it.

In the news, the term typically comes up in three contexts. First, in connection with new US export rules, where Ascend serves as evidence that China is catching up. Second, regarding production issues: the chips are manufactured by the Chinese company SMIC, which uses older machinery and therefore has high defect rates. Third, in connection with Chinese AI models that are deliberately trained on domestic hardware.

A common mistake is confusing Ascend with Kirin. Kirin refers to Huawei’s smartphone chips, while Ascend refers to data-center chips for AI. Both come from the same corporate division, but they serve entirely different purposes.