In an industry currently defined by the soaring costs of compute and hardware, Chinese AI challenger DeepSeek has thrown a gauntlet that could reshape the domestic landscape. While major incumbents like Baidu and Tencent have recently signaled price hikes for their cloud-based AI services, DeepSeek has announced a staggering 75% discount on its latest V4 Pro model. The new pricing structure brings input costs down to a mere 0.25 RMB per million tokens, a move that starkly contrasts with the rising overhead of its peers.
This aggressive strategy comes at a time when the 'Big Three'—Alibaba, Baidu, and Tencent—are citing the global crunch in high-end AI chips and infrastructure as a reason for upward price adjustments. Since March, these giants have reported service fee increases ranging from 5% to 30%, citing the escalating costs of maintaining massive GPU clusters. DeepSeek’s move suggests a fundamental shift in strategy, prioritizing market penetration and architectural efficiency over immediate margin recovery.
Technologically, DeepSeek V4 achieves these cost savings through a proprietary 'Mixed Attention Architecture' comprising Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). By alternating these mechanisms, the model drastically reduces the floating-point operations and memory cache required for processing long-context prompts. Compared to its predecessor, the V4 Pro requires only 27% of the compute power and 10% of the memory cache, allowing DeepSeek to maintain high performance while undercutting competitors.
Crucially, DeepSeek is decoupling itself from the global Nvidia-dependent supply chain by deepening its integration with domestic hardware. The company has announced full 'Day 0' optimization for Huawei’s Ascend 950 and A3 super-node clusters, as well as Cambricon’s inference frameworks. This synergy between domestic chip design and model architecture is becoming a critical competitive moat for Chinese AI firms facing ongoing US export restrictions on advanced semiconductors.
When viewed against the global market, the price disparity is even more jarring. DeepSeek’s input price is over 700 times cheaper than the weighted average of overseas models like GPT-5.5 Pro. By positioning itself as a high-efficiency, low-cost alternative, DeepSeek is not just fighting a local price war; it is challenging the established economic model of Large Language Models (LLMs) that has traditionally favored those with the deepest pockets for hardware procurement.