DeepSeek open-sources Huawei Ascend kernels to break CUDA lock-in
Portable TileLang and DeepGEMM libraries run the same code on Huawei or Nvidia, as a 128-chip Ascend supernode ships and a court narrows AI's fair-use cover.

Copy markdown
Write once, run on Huawei
DeepSeek open-sourced an Ascend SDK — TileLang-Ascend for writing kernels, plus DeepGEMM-Ascend and TileKernels — that keeps the same Python APIs whether you target an Nvidia or a Huawei backend. It's the first credible way to port inference and custom kernels off CUDA without a rewrite, and it's free on GitHub.
A 128-chip Ascend supernode
Alongside the SDK, DeepSeek and Huawei detailed a "supernode" that wires 128 Ascend 950 processors into one system, tuned for the new toolkit. Non-Nvidia compute at real scale is arriving — worth tracking if you deploy in China or want leverage on GPU price and supply.
AI training loses its fair-use shield
The Third Circuit affirmed that ROSS's use of Westlaw headnotes to train a legal-research AI was not fair use — the first US appellate ruling on fair use for AI training, though the opinion is still sealed. It sharpens the legal risk of training or fine-tuning on scraped copyrighted data and strengthens the case for licensed corpora.