Pricing is a central part of DeepSeek’s strategy. The company is charging $3.48 per million output tokens for V4-Pro and $0.28 for V4-Flash—far below the roughly $25 to $30 charged by leading U.S. providers for comparable usage. DeepSeek said costs could fall further as Huawei increases production of its Ascend 950 AI chips, which were used to train the model.
Markets reacted unevenly to the release. Shares of Semiconductor Manufacturing International Corporation, which manufactures Huawei’s AI processors, rose 10% in Hong Kong trading, while rival Chinese AI firms MiniMax and Knowledge Atlas each dropped more than 9%. The more muted global response contrasts with the company’s earlier releases, which triggered a broad sell-off in U.S. tech stocks.
DeepSeek first drew international attention in late 2024 with its V3 model, which it said was trained on roughly $5.6 million worth of compute—an amount AI researcher Andrej Karpathy described as a “joke of a budget.” That release, followed by its R1 reasoning model, challenged assumptions about the cost of building competitive AI systems and helped spark a wave of open-source model development across China. The company’s approach also pushed competitors, including OpenAI, to release their own open models.
The new V4 models arrive in a more crowded field. Chinese developers including Alibaba, Moonshot AI, MiniMax, and Knowledge Atlas have all launched high-performing open-source systems this year, intensifying competition. At the same time, DeepSeek is reportedly seeking new funding from Tencent and Alibaba that could value the company at $20 billion, with talent retention cited as a key motivation.
Hardware strategy remains a critical factor. DeepSeek worked closely with Huawei to ensure compatibility with its Ascend processors, and Huawei said Friday its chips would offer “full support” for the models. That alignment reflects ongoing constraints from U.S. export controls, which limit Chinese firms’ access to advanced Nvidia hardware. Those restrictions have pushed local developers to optimize for efficiency, contributing to lower operating costs.
The company’s reliance on domestic chips also underscores a broader shift away from U.S. semiconductor suppliers. Nvidia CEO Jensen Huang recently warned that “the day that DeepSeek comes out on Huawei first, that is a horrible outcome for ,” highlighting concerns about the geopolitical implications of China’s growing AI self-sufficiency.
At the same time, DeepSeek continues to face scrutiny from U.S. officials and competitors. Policymakers and companies including OpenAI and Anthropic have accused Chinese developers of using “illicit” model distillation or unauthorized access to advanced chips. On Thursday, White House science advisor Michael Kratsios alleged “industrial-scale campaigns” to copy U.S. technology, claims China’s foreign ministry dismissed as “groundless” and “a smear against the achievements of China’s AI industry.”
Despite those tensions, DeepSeek’s latest release underscores a narrowing performance gap between Chinese and U.S. AI systems—paired with a pricing model that could intensify pressure on incumbents already raising rates to manage demand.
This analysis is based on reporting from Fortune.
Image courtesy of Deepseek.
This article was generated with AI assistance and reviewed for accuracy and quality.
