on US vs China compute question.
Keep in mind that frontier US AI labs have a pretty large compute advantage over China, at least in terms of training cluster compute.
In overall compute, China actually has plenty inside China and outside of China (like with ByteDance, Ali & Tencent in their overseas Data centers). But the problem is that the big tech in China that has large Nvidia clusters are not the most innovative AI shops in China. Whereas Zhipu and Moonshot are probably limited in their Nvidia cluster. Even now, AI researchers in China don't like to use Ascend, they like Cuda.
Now that might be changing since with LLM coding, you can now get kernels for AI chips designed really well (like Qwen did with M890). But there is a gap between when OpenAI and Anthropic got their compute and funding vs when the DeepSeek, Zai and Moonshot got theirs. So that's why you can potentially see a case where American ones just iterate faster. It's unclear to me the implication of that, since the only purpose I've seen in general purpose LLM so far is cyber warfare and replacing software engineers.
But this could become problem. Based on the current pace of development by Ascend (+ other domestic chip options and AI supply chain) and China's willingness to hammer Japan's midstream electronic supply chain, you could see some real shift in compute available to Chinese frontier lab and American ones.
But again, DeepSeek is unlikely to get big gains from Ascend until second half of this year. At which time, it can run more experiments and hopefully speed up releases.
If you look at the release cycles, DS is actually pretty slow. DS4 is structurally the best architecture, but legit under trained. Both GLM-5.2 and Kimi 2.7 are a mile ahead in software engineering work. So, I hope they really speed things up later this year. At least do releases every 2 months consistently, instead of 3 or 4.
