Business and economy·October 20, 2026, 10:52

DeepSeek and Huawei reportedly developing AI chips together

AI-generated from the sources listed, with automated source checks.

According to a report from The Information, relayed by Huawei Central, DeepSeek and Huawei are collaborating on chips for DeepSeek's models. An analysis citing SemiAnalysis argues that parts of DeepSeek's latest model, V4, were designed with Huawei's Ascend chips in mind.

AI-generated image

Chinese AI company DeepSeek and technology group Huawei are working together to design AI chips tailored to DeepSeek's models. That is according to the news outlet The Information. The claim comes here from the technical news site Huawei Central, which relays the report. The report itself is behind a paywall.

Joint work on chips and software

According to Huawei Central's account, DeepSeek shares the computational problems that arise when large language models run. Huawei uses this to improve the performance of its Ascend chips and to strengthen its CANN software. CANN is Huawei's alternative to Nvidia's CUDA, the software many AI developers use today to control Nvidia chips.

Huawei Central also writes that work is under way to improve memory bandwidth and the connections between chips in the next generation of Ascend. Bandwidth indicates how quickly data can be moved in and out of memory. It is often a bottleneck for large AI models.

The two companies will also, according to Huawei Central, build an open ecosystem with local AI models. That is a plan, not something that has happened yet. Huawei Central also notes that Ascend 960DT is expected to be announced in the first quarter of 2027, and that Ascend 960PR is expected to be released in the third quarter of 2027.

V4 and demand for Ascend

An analysis dated June 17, 2026, on the Substack newsletter China AI Dispatch looks at how DeepSeek V4 relates to Huawei's chips. The analysis cites SemiAnalysis, which argues that parts of V4's architecture were designed around Huawei Ascend from the start. This includes the attention mechanism, the quantization in the mixture-of-experts, and the communication between experts in the cluster. Quantization, in short, means that numbers in the model are made smaller and easier to compute with.

The analysis states that Ascend 950DT has 144 GB of memory at 4 TB/s. According to it, Huawei engineers said the training could not have taken place without the Ascend 950 supernode.

The analysis also writes that ByteDance has been allocated half of the Ascend 950 production line, and that Alibaba and Tencent have each ordered hundreds of thousands of units. According to the analysis, China Mobile has ordered 776 Ascend nodes for procurement in 2026-2027. DeepSeek's token share on Vercel AI Gateway rose to 17 percent in the month V4 was launched.

Caveats

The information about the collaboration relies on an intermediary and on anonymous sources in The Information's report. The Substack analysis is one individual's analysis, and the figures in it have not been independently confirmed. The analysis is also mostly about V4 and inference, meaning when a finished model is used, and not about the chip design collaboration itself.

Sources

More on this topic

Get the week's AI news in your inbox

Choose your level, topics and length. One email a week, unsubscribe at any time.

Subscribe to Promptly Newsletter