DeepSeek v4 Pro 1.6T model post-trained by Huawei on 1000 Ascend 910C chips

China's competitiveness in AI, despite export controls, should surprise nobody.

AI workloads are very simple and massively parallel. Energy into the workload, trained artifact out. The bigger the artifact the more parallel you have to be and hence the more energy you have to use. (the SW engineering to make it possible is difficult but fundamentally tractable).

Because of this, it is possible for China to train competitive models even at a fraction of the power efficiency of the advanced USA chips. Energy is the issue more and more as the models get larger.

We have a serious problem with the cost of energy in the US. AI dominance is far from guaranteed. I DGAF but the policymakers who seem to care about this are not doing much to fix the situation. Like in EVs and solar, we are going to start getting lapped. Instead of export controls in the USA->China direction, we're going to start seeing import controls in the China->USA direction, for inference.

>As for the claim coming out of Shenzen, it carries no benchmarks, gives no figure for how long the run took, how it compared to the same job on Nvidia hardware, or how efficiently the 1,000-chip cluster was used. It’s ultimately just another addition to a series of dubious claims that have come from the Chinese state without anything to back them up; DeepSeek itself hasn’t commented.

As if OpenAI and Anthropic are giving us ball to ball commentary on how their training runs go. Deepseek did train it on domestic hardware, model might be out in public soon (open weights or not) and then anyone can see what is it about.