Citrini’s view: Bullish on AMD, bearish on NVIDIA. Coding AI is eroding NVIDIA’s competitive moat from the software side, marking the end of its CUDA moat.

Citrini analyst Jukan, citing recent core views from DeepSeek founder Liang Wenfeng, pointed out that AI-driven code generation and high-level programming languages like TileLang are rapidly lowering entry barriers to the CUDA ecosystem. While DeepSeek uses NVIDIA GPUs to train its V3 model, it has significantly reduced its reliance on NVIDIA’s software ecosystem via its self-developed compiler and TileLang environment. Earlier, Liang projected that porting TileLang and DeepSeek’s compiler to Huawei chips would largely resolve China’s chip ecosystem issues in about a year, with production capacity being the only remaining bottleneck. Liang quantified the China-U.S. chip gap: hardware efficiency is roughly four times lower, and there is a roughly two-year time lag. He also revealed that DeepSeek is working closely with Huawei, expecting to obtain around 16,000 Huawei AI chips, and the Huawei 950 SuperNode can replace the workloads of NVIDIA’s GB200/GB300. Analyst Jukan characterized this as "the end of CUDA’s moat" and holds a highly bearish outlook on NVIDIA. Jukan added that this line of reasoning is precisely one reason for being bullish on AMD: advances in coding AI will also naturally accelerate the development of the ROCm ecosystem, helping narrow its gap with CUDA. When AMD recently invested in Anthropic, it announced it would actively use Claude Code for chip design and software engineering. Overall, advances in AI programming tools are systematically eroding NVIDIA’s competitive barriers from the software side. China’s chip ecosystem issues will be rapidly resolved thanks to code generation capabilities, while AMD will benefit from ROCm’s accelerated growth. The CUDA moat NVIDIA relies on to retain developer loyalty is facing a two-pronged attack, and catching up in hardware efficiency and production capacity is only a matter of time.