DeepSeek has emerged as a competitive player in the AI model training landscape, spending approximately $5.58 million to develop its V3 model. This expenditure is surprisingly low when compared to US competitors, who typically invest tens to hundreds of millions on advancing high-level models.
In May 2026, DeepSeek made a significant move by permanently reducing the prices of its V4-Pro model by 75%. Remarkably, the cost of cached input decreased to around RMB 0.025 per million tokens. This aggressive pricing strategy positions them favorably against other players in the AI market.
Other Chinese companies are also joining the trend toward lower prices. 01.ai, for example, is offering inference services at approximately 14 cents per million tokens, which establishes a new low for global API pricing. The price differentials have allowed Chinese AI models on OpenRouter to experience a notable 5x increase in volume, primarily driven by these cost advantages over their US counterparts.
How is this significant reduction being achieved? Chinese developers have innovatively designed sparse mixture-of-experts (MoE) architectures. This technique has enabled them to cut down parameter activation from an astounding 671 billion to only 37 billion, resulting in a substantial 90-97% decrease in compute costs at the inference level.
In addition to architectural advancements, many Chinese teams are utilizing lower-precision training methods like FP8. This methodology effectively lowers the computational demands for each calculation. As a case in point, DeepSeek's R1 reasoning model was trained for a mere $294,000, relying on 512 H800 chips over 80 hours.
Historically, the landscape has shifted since 2023, when US export restrictions limited Chinese companies' access to high-end Nvidia hardware such as the H100 and its successors. As a workaround, Chinese developers have been utilizing the H800, a downgraded chip that complies with export regulations.
Major Chinese industry players pushing the boundaries of efficiency include Alibaba’s Qwen, Moonshot AI’s Kimi, Zhipu AI’s GLM, and ByteDance’s Doubao, alongside DeepSeek’s innovations.
For investors, this significant decrease in training costs raises questions about the traditional capital expenditure barrier surrounding US AI leaders. The prospect of achieving frontier-level AI performance with training expenses under $6 million—compared to well over $100 million—suggests that the competitive edge of US firms may be diminishing.
Moreover, the reduction in inference costs will have direct implications for the crypto and Web3 ecosystem. It will lower the operational costs for AI-driven decentralized applications, oracle networks, and on-chain analytics tools.
The advances Chinese developers are making, particularly the 97% compute reductions fostered by sparse MoE architectures, signal a critical shift in the market dynamics. Investors should take note, as market behavior often aligns with prevailing price trends.