Claude completed the kernel code adaptation for AMD MI355X in just 3 days, reducing the timeline from 6–12 months to 72 hours, with performance only 8%–12% behind the CUDA solution.Article author, source: AIBase
The competitive landscape for AI hardware and underlying software ecosystems is undergoing a profound structural shift. Recently, major industry news emerged that Anthropic’s Claude Mythos successfully wrote nearly 100% of the底层 kernel code for AMD’s MI355X chip in just three days, with almost no human intervention.
This breakthrough has reduced the chip adaptation cycle—from originally requiring dozens of senior engineers 6 to 12 months—to just 72 hours. Real-world testing shows that the performance of the ported model is within only 8% to 12% of NVIDIA’s CUDA-optimized solution, fully meeting commercial deployment standards.
The barrier to cross-chip adoption has been significantly lowered, bringing diversification to the computing power supply chain.
For a long time, NVIDIA’s dominant position in the AI computing market has not been due solely to the raw hardware performance of its chips, but rather to the extensive software ecosystem built over years — a moat constructed by countless engineers through lines of code. For other chip manufacturers, the high cost of adaptation and the lengthy software porting timeline have remained insurmountable barriers.
Claude’s exceptional efficiency in writing low-level code has directly created a breach in NVIDIA’s software barrier, significantly lowering the barrier to cross-chip compatibility. Based on this technological advancement, Anthropic has now partnered with AMD to secure up to 2 gigawatts of computing power.
Advantages and limitations coexist:Multi-GPU collaboration remains a future focus
Although this technological breakthrough has caused widespread shock in the industry, multiple analyses have maintained an objective and rational perspective.
From an application perspective, this breakthrough currently primarily covers the inference pre-filling stage. In more challenging areas such as multi-GPU coordination for ultra-large-scale distributed training, NVIDIA continues to maintain a clear lead thanks to its deep technical expertise. Therefore, this advancement has not rapidly overturned the CUDA ecosystem, but it has undoubtedly introduced new variables into the computing power market, accelerating the global computing industry’s evolution toward greater openness and diversity.


