In a recent interview show, NVIDIA CEO Jen-Hsun Huang said that if DeepSeek’s new model is launched on the Huawei platform, it will be a terrible day for the United States. Huang Renxun said that this means that AI models are optimized to perform best on Chinese AI hardware, and when these models spread around the world, they will promote Chinese technology to become the world standard.

There have been reports before that DeepSeek V4 will support the Huawei Ascend platform, and there are even reports that V4 only supports the Huawei platform. However, this is impossible. The DeepSeek V4 released today clearly supports both Huawei and NVIDIA platforms.

DeepSeek stated in the article that DeepSeek V4 will complete the verification of the fine-grained expert parallelism (EP) solution on both NVIDIA GPU and Huawei Ascend NPU.

This means that what Huang Renxun was most worried about has not happened - DeepSeek V4 will not only support Huawei or other domestic AI platforms, but will still support NVIDIA's CUDA ecosystem, and it will also be the first support.

Of course, tonight we still have to pay attention to whether DeepSeek V4 will immediately include support for Huawei's Ascend platform. Will this impact NVIDIA's stock price like last year? If the stock price is also stable, Lao Huang can have a good sleep.

In addition, there is another issue worth paying attention to this time with DeepSeek V4, that is, the price of V4 Pro has increased 6-8 times compared with the previous V3.2. The output of one million Tokens is 24 yuan, the cache miss in the output of one million Tokens is 12 yuan, and the cache hit is 1 yuan, which is higher than Domestic large models such as Kimi 2.6 or GLM-5.1 also have high and low prices, and they are not as cost-effective as before.

There are many factors that affect the cost. After all, V4 Pro is a big guy with 1.6 trillion parameters, and the computing power cost is bound to be impossible to be very low. Moreover, the official API page also mentions that due to the high-end computing power, the service throughput of the current V4 Pro model is still limited.

However, DeepSeek also made it clear that after the Ascend 950 super node is expected to be launched in batches in the second half of the year, the Pro price will be significantly reduced, so surprises will have to wait until the second half of the year.