Wenfeng asserts its domestic 950DT AI chip can perform all functions of NVIDIA's GB300, including latency-sensitive tasks, at an effective 1:4 exchange rate. The company believes NVIDIA's CUDA moat is rapidly dismantling due to AI's code-writing capabilities and new high-level languages like TileLang, which can quickly rewrite CUDA operators.