Techub News, citing CryptoBriefing, reported that Chinese AI developers are lowering training and inference expenses through innovations including sparse MoE architecture. According to the report, computing costs have been reduced by as much as 97%.
Low-cost inference expands usage on OpenRouter
DeepSeek was highlighted as one of the examples. The training cost of its V3 model was about $5.58 million. Its V4-Pro model cut prices by 75% in May 2026, bringing inference costs down to RMB 0.025 per million tokens.
The report also noted that Chinese companies including 01.ai are offering extremely low API pricing. This cost advantage has driven a fivefold increase in the use of Chinese AI models on OpenRouter, indicating that cheaper inference services are expanding model usage across developer platforms.
For the Web3 ecosystem, lower inference costs will reduce operating expenses for decentralized AI applications and on-chain analytics tools. Projects that rely on repeated model calls can benefit from lower API and inference pricing in their operating cost structure.