DeepSeek is developing its own inference ASIC, roughly a year into development and pre-tapeout, engaging with design houses, foundries, and memory suppliers. This follows a pattern of frontier AI labs like OpenAI and Anthropic evaluating or designing custom silicon to optimize inference costs. China's export controls block DeepSeek from leading-edge foundries, forcing reliance on trailing nodes and domestic memory suppliers like CXMT, which targets HBM3 mass production by late 2026, a 1-2 generation gap behind HBM4.