they might not need more training infrastructure at this point

The article says "to run AI models" so it probably means both inference and training, not just inference.

Running refers to inference usually

If they can be used for inference, I would assume they can be used for training

They run during training too. Also, why would they get 160000 chips only for inference?

Training has a lot of extra functionality like calculating how to update the weights of the model during training to make it more performant on the dataset(backpropagation) and much more

Meanwhile inference is mostly running the weights of the model as they are. The model isn't being adjusted in any way. And Nvidia holds a strong grip on training libraries through Cuda

Nvidia does not have a monopoly on training.

meituan-longcat/LongCat-2.0 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

(huggingface.co)

Both the full training run and the large-scale deployment are built entirely on AI ASIC superpods. Pretraining spans millions of accelerator-days across more than 35 trillion tokens, with no rollbacks or irrecoverable loss spikes — demonstrating that we have the capability to conduct frontier-scale training on alternative hardware platforms.

LongCat 2.0 is of similar size to DeepSeek V4 Pro and "AI ASIC Superpods" are from Huawei.

Yes but no. Asics are very inflexible and have a hard time dealing with new functions. GPU's are general purpose and therefore new types of convolutions are much easier to implement and have the driver support for it.

There are frequent claims of random companies making NPU's TPU's etc but they always have software issues for training new models with the latest methods. If not Nvidia wouldn't be having such massive market share anymore.

That's all software, I don't think anyone is doing back prop in hardware. Your are right about cuda, but z.ai just demonstrated that most is evaporating quick

If deepseek bought those chips only for inference, they wouldn't be able to keep up in model development and wouldn't need inference anymore.

As Meituan uses these chips for training, Deepseek will also use these chips for training. You claiming the chips are somehow defective won't make Deepseek not use them.