Ascend Chip

An Ascend chip is a highly specialized computer chip made by the Chinese technology conglomerate Huawei. It is not meant to run an operating system or programs like Word. Instead, it performs one single kind of task extremely fast: gigantic quantities of simple calculations with strings of numbers, all at once and in parallel. These are precisely the calculations that underlie everything today called artificial intelligence, i.e., programs that learn from examples instead of following fixed rules. Without such specialized chips, a system like ChatGPT would not be technically affordable. The Ascend series is Huawei’s attempt to become independent of the West in this market.

China’s way out of the chip blockade

The market for AI chips has for years been dominated by a single company: the US firm Nvidia. Almost all of the world’s major AI models were trained on Nvidia hardware. Since 2022, the US government has banned the export of the most powerful of these chips to China. The reason is political: Washington wants to prevent China from developing militarily usable AI on a large scale.

This created a real problem for Chinese companies. Anyone who is not allowed to buy top-tier chips cannot train top-tier models. Ascend steps directly into this gap. The chips are China’s most serious attempt to build its own foundation for AI. Whether this succeeds will help determine how strongly the technology world splits into two separate blocs.

That is why Ascend chips regularly turn up in economic news. Every report about a new Ascend generation is at the same time a report on the effectiveness of US sanctions. And it moves stock prices, because investors read into it how secure Nvidia’s lead still is.

What happens inside

A normal main processor is a jack-of-all-trades. It works through tasks one after another and is very flexible in doing so. An Ascend chip is the opposite: it contains thousands of very simple computing units that all start working at the same time. You can picture it as the difference between an excellent individual craftsman and a factory hall with a thousand workers. For a complicated one-off piece, the craftsman is better; for a thousand identical parts, the factory hall is better.

The calculations involved are so-called matrix multiplications, meaning the mass multiplication and addition of tables of numbers. Huawei has developed its own architecture for this purpose called Da Vinci. On top of that comes a software toolkit called CANN, through which programmers control the chip. An Ascend chip rarely works alone: for training large models, hundreds or thousands of them are linked together into a cluster.

A common misconception is that a chip’s computing power alone is decisive. The surrounding software is just as important. With its CUDA platform, Nvidia has built up a huge ecosystem over almost twenty years that developers know and trust. Huawei’s tools are younger and considered more cumbersome. On top of that comes manufacturing: because Huawei is not allowed to buy the most modern machines from the West, the chips are harder to produce in large quantities.

Where Ascend is used

As a private individual, you will not be buying an Ascend chip directly. It sits inside data centers, in cabinet-sized servers, often as part of a complete system called Atlas. Chinese government agencies, universities, banks, and telecom companies run their AI applications on it. Huawei’s own language model Pangu also runs on this hardware.

In the news, the term is usually encountered in one of three contexts. First, in connection with new expansion stages such as the Ascend 910B or 910C, whose performance is compared to Nvidia models. Second, in sanctions debates, when Washington deliberates further restrictions. Third, when a Chinese AI company announces that it is switching from Nvidia to Ascend.

For investors, this is a tangible issue. China was long a billion-dollar market for Nvidia. Every percentage point that Ascend gains there is missing from the US company’s balance sheet. However, anyone reading such reports should be cautious with manufacturer claims: announced peak figures say little about how many chips can actually be manufactured and delivered.