>>106285914
periodic reminder that all "DeepSeek failed to train on Huawei" news are suspect, at least people at DS tell me not to trust insane rumors. The journalist doesn't even understand the stage of Huawei's chip development, there literally was no platform for DS to do a V3-scale training run until very recently and it's implausible that they'd have been forced to debug prototype hardware with a novel architecture just to make Xi happy. She also apparently doesn't know about V3-0324, R1-0528, Prover-V2 and other evidence that they've been continuing their normal development cycle, or about actual Ascend use by DeepSeek (for inference).

China is not that cartoonishly retarded and DeepSeek doesn't share anything with the press. I think the most likely hypothesis of the source is cope from a disgruntled Huawei employee or a competitor like Alibaba.

The next model is most likely V4 with NSA, coming out in a matter of weeks. I predict it'll have been trained on normal Nvidia GPUs, for 3+ months, and retards will claim that this is a second run after the "failure" to train "R2" on Ascends by May to "capitalize on hype". All of this is layered confusion of Western and Chinese hypemongers and journalists.
You can trust me or FT, your choice.