Huawei is trying to make thousands of its AI chips work as one system, while DeepSeek is helping develop software to make that hardware useful. A new report describes the partnership and its promise, alongside the power, cooling and software challenges that make replacing Nvidia more than a chip-counting exercise.
DeepSeek Watch analysis
What happened
The Hani reports that Huawei’s Atlas 950 SuperPoD is designed to link up to 8,192 Ascend AI chips. Huawei demonstrated a working 1,024-chip network in July. In September, DeepSeek and Huawei unveiled a framework for coordinating computing and communication across 128 Ascend chips, as well as infrastructure for using the open-source TileLang programming tool.
The report cites IDC figures showing Nvidia held 34% of AI accelerator shipments to Chinese data centres through June, compared with Huawei’s 28%. Those shipment shares do not measure computing performance or sales, the report cautions. Read the Hani report.
Why it matters
The effort is not simply to build a faster chip. Linking large numbers of less powerful processors could help Chinese developers build AI systems with domestic hardware, while software work aims to reduce dependence on Nvidia’s CUDA platform. DeepSeek’s role matters because chips need a working software ecosystem before they become useful infrastructure.
Scale brings its own bill. The Hani notes that connecting many chips raises power use, cooling and communication costs, and that stable performance in large commercial systems remains difficult. A rising shipment share is a sign of momentum, not proof that the replacement job is done.
Our read
This is a credible account of a serious effort to build a domestic AI-computing stack, not evidence that Nvidia has been neatly pushed aside. The telling detail is the partnership between hardware and software: China’s alternatives need both to improve together. The next test is whether they can deliver reliable performance at a cost operators can live with. Silicon does not become an ecosystem by sheer force of optimism.
What to watch
- Whether Huawei demonstrates larger Ascend systems in sustained commercial use.
- Whether DeepSeek’s framework and TileLang work expand beyond the reported 128-chip setup.
- How power, cooling and communication costs compare as systems grow.
Discussion spark: If domestic AI chips have lower individual performance but can be linked at scale, what should count most when judging whether they can replace Nvidia: benchmark speed, operating cost or a mature software ecosystem?
Sources and evidence
- DeepSeek and Huawei drive China’s AI ascent, but Nvidia remains a hurdle – 한겨레 (17 November 2026, 07:50 UTC)
not affiliated with or endorsed by DeepSeek