Citrini View: Bullish on AMD, Bearish on NVDA, Encoding AI Eroding NVIDIA's Moat from the Software Side, CUDA Moat is Coming to an End

BlockBeats News, July 24th, Citrini analyst Jukan quoted DeepSeek founder Liang Wenfeng's recent key point, pointing out that AI-driven code generation and advanced programming languages like TileLang are rapidly lowering the barrier to entry of the CUDA ecosystem. Although DeepSeek uses NVIDIA GPUs to train the V3 model, it has greatly reduced reliance on the NVIDIA software ecosystem through self-developed compilers and the TileLang environment. Liang Wenfeng previously predicted that if TileLang and the DeepSeek compiler were ported to Huawei chips, the ecosystem issue of Chinese chips could be basically solved within about a year, with the true bottleneck being only production capacity. Liang Wenfeng quantified the gap between Chinese and American chips—hardware efficiency is about 4 times higher, with a time lag of about two years—and revealed that DeepSeek is closely collaborating with Huawei, expecting to receive around 16,000 Huawei AI chips. The Huawei 950 SuperNode can replace the workloads of GB200/GB300. Analyst Jukan characterized this as the "end of CUDA's moat" and is extremely bearish on NVIDIA.

Jukan added that this logic is precisely one of the reasons to be bullish on AMD—advancements in AI coding will also naturally accelerate the development of the ROCm ecosystem, helping to narrow the gap with CUDA. When AMD recently invested in Anthropic, it announced that it would actively use Claude Code for chip design and software engineering. Overall, the advancement of AI programming tools is systematically eroding NVIDIA's competitive moat from the software side, the ecosystem issue of Chinese chips will be rapidly resolved due to code generation capabilities, AMD will benefit from ROCm acceleration, NVIDIA's CUDA moat that developers rely on is facing a double squeeze, and catching up in hardware efficiency and production capacity is only a matter of time.