Gate News, April 27 — DeepSeek has delayed the release of its V4 model to fine-tune its software stack to suit Huawei’s Ascend chips, reflecting broader initiatives in Beijing: as access to advanced foreign semiconductors becomes increasingly restricted, efforts are being made to build a domestic AI supply chain. DeepSeek’s V4-Pro model matches performance benchmarks set by OpenAI and Anthropic in major tests, while its API costs are significantly lower: $1.74 per million input tokens, a sharp decrease compared with Western competitors. The company says that V4-Pro’s compute efficiency is 27% higher than its V3.2 predecessor, using significantly less compute power in a 1 million token context. DeepSeek previously demonstrated cost efficiency with its R1 model, which the company says required less than $6 million in development investment. Market reaction reflects a shift toward adopting domestic chips. Chinese AI companies MiniMax and Zhipu (Knowledge Atlas Technology) saw their share prices each fall by about 8%, while chip manufacturers benefited: Semiconductor Manufacturing International Corporation rose 9%, and HuaHong Semiconductor climbed 15%.