A fast AI chip is useful only if developers can make their models run well on it. That often depends on less visible software: tools that translate programs into hardware instructions and move data between chips.
On 30 September 2026, DeepSeek said it had worked with Huawei to develop and open-source programming infrastructure for Huawei’s Ascend processors. According to Reuters’ account of DeepSeek’s official announcement, the released components include compute and communication libraries, while the partners described a system built around 128 Ascend 950 chips. This is a software and infrastructure announcement, not an independently verified result showing a complete system outperforming competing hardware. Reuters report
Why code matters to a chip
An AI model repeatedly performs operations such as multiplying large blocks of numbers and exchanging partial results among processors. A kernel is a small, carefully optimised program for one such operation. A communication library handles the traffic between chips. Without good versions of both, a large machine can spend too much time waiting, even if each processor is capable of high speed.
DeepSeek highlighted TileLang, a higher-level language meant to make these kernels easier to write while retaining performance. There is already a publicly visible Ascend adapter project for TileLang; a repository is evidence of available code, though it is not proof of broad production use or speed on every workload. TileLang Ascend project; Announcement context
The harder test comes after release
Software ecosystems become valuable when outside teams can install the tools, move existing workloads, diagnose problems and repeat performance results. Claims about a simpler programming model or a highly capable “supernode” need evidence across real training and inference tasks, along with information about reliability and power use.
DeepSeek and Huawei’s move is significant because it tackles the practical barrier between making processors and making them useful at scale. Open code may widen the pool of developers able to experiment. Whether it narrows the gap with established AI computing platforms will depend on measured performance and adoption.
Featured image: representative image by Brecht Corbeel / Unsplash. It does not depict the specific study, facility or equipment described.
Leave a Reply