Semiconductor analyst Jukan says he has turned very bearish on Nvidia after a leaked investor call in which DeepSeek founder Liang Wenfeng argued that AI powered coding tools are quickly dismantling Nvidia's software advantage. CUDA, the programming platform that lets developers run AI software on Nvidia chips and that has kept rivals locked out for over a decade, is not the real moat anymore, in Wenfeng's telling. He says DeepSeek already trained its V3 model on Nvidia GPUs while cutting its reliance on Nvidia's software stack, using its own compiler and a programming tool called TileLang instead. If that toolchain gets ported to Huawei's chips, Wenfeng argues, the software gap for Chinese chips could largely close within about a year, leaving hardware efficiency, where Huawei's chips still trail by roughly four times and about two years, as the only real bottleneck left. Huawei's 950 SuperNode system is already positioned to substitute for Nvidia's GB200 and GB300 racks in some workloads. Jukan ties this to a bullish case for AMD too, arguing the same coding AI tools will speed up development of ROCm, AMD's open source answer to CUDA, especially since AMD's recent investment tie-up with Anthropic already has it using Claude Code for chip design work.
Why it matters · If software tools, not scarce hardware, were the main thing keeping Chinese chipmakers and AMD locked out of Nvidia's ecosystem, then AI coding assistants closing that software gap would remove one of Nvidia's biggest competitive advantages years faster than expected, reshaping who captures the profits of the AI buildout.
Worth asking · If AI coding tools really can port DeepSeek's Huawei-ready software stack across in about a year, is Nvidia's CUDA moat cracking on a timeline the market hasn't priced in yet?