CANN

CANN is a software package from the Chinese company Huawei. It sits between two worlds: above are the programs developers use to build artificial intelligence, below are the specialized chips that do the computing for it. At Huawei, these chips are called Ascend. They don’t understand ordinary programs but need instructions in their own language. This exact translation is what CANN performs. The name is an abbreviation for Compute Architecture for Neural Networks.

Huawei’s answer to Nvidia's lead

The market for AI chips has been dominated for years by the US company Nvidia. Its real strength isn’t just the hardware. It’s the software called CUDA, which developers use to control Nvidia chips. Nearly all AI tools worldwide are tailored to CUDA. Anyone who buys a competing chip first has to laboriously adapt these tools.

This is exactly where CANN comes in. It’s meant to make switching to Huawei chips as painless as possible. This is especially important for China, because the US has restricted the export of powerful AI chips there since 2022. Chinese companies therefore need domestic alternatives. But a good alternative always consists of chip and software together.

In August 2025, Huawei announced it would largely open up CANN. The source code is meant to become viewable and modifiable for developers. The idea behind this: the more people contribute, the faster the gap to CUDA closes. Whether this will succeed remains open. Nvidia has around a twenty-year head start in building its developer community.

From model to instruction on the chip

An AI model is usually built in a framework, meaning a construction kit for neural networks. Well-known examples are PyTorch or the Chinese MindSpore. In this construction kit, one describes which computational steps should happen in sequence. This description is still abstract and not intended for any particular chip.

CANN takes in this description and restructures it. One part of it, the graph compiler, views the computational steps as a network of nodes. It merges steps, removes unnecessary ones, and rearranges them. After that, ready-made computational building blocks come into play, so-called operators. Each operator is a highly optimized small program for a typical task, such as a matrix multiplication.

You can picture CANN as a translator with an editorial mandate. It doesn’t just translate word for word. It shortens, rearranges, and chooses the fitting phrasing in the target language for each sentence. Only this way does a model become a workflow that truly utilizes the chips. Poor translation here means: the chip is fast, but idle half the time.

Where the name appears in the news

CANN is directly encountered only by developers running AI on Huawei hardware. These are mainly Chinese data centers, cloud providers, and universities. You’ll never see the name in mobile apps or on websites. It’s hidden beneath the surface, much like an operating system beneath a program.

In business news, however, CANN comes up regularly. Whenever the chip dispute between the US and China is discussed, the term is mentioned. Analysts use it as a gauge: how far along is Huawei really? Reports of new CANN versions or of companies switching from Nvidia to Ascend are seen as indicators of this.

A common misconception is equating CANN with the Ascend chips. CANN is pure software and does no computing itself. Equally mistaken is the assumption that switching is just the push of a button. In practice, developers report missing features and less sample code than with CUDA. In the end, building the hardware was the smaller task for Huawei.