Nice summary. This is only one potential response to the training challenge in China. It should be viewed along with distributed training and heterogenous hardware platforms as part of a set of responses to US controls and the focus on scaling. Even without controls Chinese firms should be taking a different road, playing a different game. See my latest SS on distributed computing and other innovations in China's AI sector...
Hmm I don't know if it can really be called a real deepseek moment. Trading off that much precision might kneecap an entire generation of models. And software-led AI growth is just not realistic. We can see from GPT-5, everyone coalescing/plateauing at the same area that there are limits to using existing hardware to train new models. Hardware must grow commensurately
Good article! I have no doubt about the R&D capabilities of DeepSeek, Huawei, or Cambricon. But at this point in time, it seems that China’s AI industry, both hardware and software, is still following behind the U.S. (which is already no small achievement). Take MXFP8 as an example: Nvidia’s Blackwell already supports it, and even supports its own trickier NVFP4 format. Meanwhile, Chinese hardware vendors haven’t fully mastered FP8 yet. So I can only interpret DeepSeek’s statement this time as an attempt to leverage nationalism for momentum. Given that hype and exaggeration are common in the AI industry, this isn’t something to criticize. But for me personally, it does affect my feelings toward this great company.
I am a little hesitant on over pivoting on this as gpt-oss leveraged mxfp4 pretty heavily as well. Training stability on these micro scaling formats is hard, but it certainly helps with the bandwidth constraints.
This reminds me of the Bible-size circuit boards that the USSR used in its missiles because it could not access the leading edge US equipment and materials that enabled finer circuit traces that reduced weight & volume.
This might be partially OT, but I was/am really impressed by the just-launched Orin Thor Boards/SoCs from Nvidia; this is Nvidia's hardware development platform for Nvidia's strong push in AI/Robotics. The ecosystem they built around that is (IMHO) already truly remarkable.
Back "On Topic" : I would like to know and learn more about similar efforts in China.
Nice summary. This is only one potential response to the training challenge in China. It should be viewed along with distributed training and heterogenous hardware platforms as part of a set of responses to US controls and the focus on scaling. Even without controls Chinese firms should be taking a different road, playing a different game. See my latest SS on distributed computing and other innovations in China's AI sector...
Hmm I don't know if it can really be called a real deepseek moment. Trading off that much precision might kneecap an entire generation of models. And software-led AI growth is just not realistic. We can see from GPT-5, everyone coalescing/plateauing at the same area that there are limits to using existing hardware to train new models. Hardware must grow commensurately
Good article! I have no doubt about the R&D capabilities of DeepSeek, Huawei, or Cambricon. But at this point in time, it seems that China’s AI industry, both hardware and software, is still following behind the U.S. (which is already no small achievement). Take MXFP8 as an example: Nvidia’s Blackwell already supports it, and even supports its own trickier NVFP4 format. Meanwhile, Chinese hardware vendors haven’t fully mastered FP8 yet. So I can only interpret DeepSeek’s statement this time as an attempt to leverage nationalism for momentum. Given that hype and exaggeration are common in the AI industry, this isn’t something to criticize. But for me personally, it does affect my feelings toward this great company.
I am a little hesitant on over pivoting on this as gpt-oss leveraged mxfp4 pretty heavily as well. Training stability on these micro scaling formats is hard, but it certainly helps with the bandwidth constraints.
This reminds me of the Bible-size circuit boards that the USSR used in its missiles because it could not access the leading edge US equipment and materials that enabled finer circuit traces that reduced weight & volume.
This might be partially OT, but I was/am really impressed by the just-launched Orin Thor Boards/SoCs from Nvidia; this is Nvidia's hardware development platform for Nvidia's strong push in AI/Robotics. The ecosystem they built around that is (IMHO) already truly remarkable.
Back "On Topic" : I would like to know and learn more about similar efforts in China.