There is a recent rumor circulating in the industry: Anthropic's Claude Mythos took only about 3 days to write nearly all of the underlying kernel code for the AMD MI355X chip with almost no manual intervention.

This statement currently comes from second-hand reports. We have not seen complete official data, so it is more suitable to be viewed as a signal to be verified.

According to reports, this type of chip adaptation originally took dozens of senior engineers 6 to 12 months, but this time it was reduced to about 72 hours. Compared with the NVIDIA CUDA optimization solution, the performance gap after transplantation is said to be between 8% and 12%, which is considered to be enough for commercial standards.

If true, it hits the hardest part of Nvidia.

Nvidia's advantage in the AI ​​computing power market has never been just the chip itself, but the CUDA software ecosystem accumulated over many years and built by a large number of engineers line by line. For other chip manufacturers, high adaptation costs and long software migration cycles have always been obstacles that cannot be overcome. If the model can greatly automate this part of the work, the threshold for changing chips will be much lower.

But the scope must be clearly stated. The progress mentioned in the report currently covers the pre-filling process of reasoning. Multi-card collaboration in ultra-large-scale distributed training is much more difficult, and NVIDIA is still clearly ahead in this area. In other words, the inference side may loosen up first, and the training side cannot do without CUDA in the short term.

Also mentioned along with this matter is the cooperation between Anthropic and AMD for the supply of computing power of up to 2 gigawatts.

What will really change the situation is that model companies will begin to have the ability to gnaw on the underlying code of the chip themselves, and the degree of being bound by the software ecosystem when selecting chips will become smaller. As for the specific figures of 3 days and 8%, we have to wait for more formal disclosure.