DeepSeek V4-Pro Price Cut Permanent: Domestic AI Completes Full-Stack Loop
On May 31, DeepSeek V4-Pro model API’s 2.5 discount promotion ended, but the official price was adjusted to one-quarter of the original pricing. This means the 2.5 discount has been permanently retained. Lower LLM costs will further activate SME AI innovation.
Core Price Adjustment Details
| Item | Details |
|---|---|
| Original Price | Original pricing (specific number not publicly disclosed) |
| Discounted Price | One-quarter of original pricing |
| Effective Date | May 31, 2026 |
| Previous Promotion | 2.5 discount ended, but equivalent discount permanently retained |
This is not a simple price cut. DeepSeek simultaneously completed full-stack underlying architecture refactoring.
Full-Stack Heterogeneous Refactoring: Changing Engines at 10,000 Meters
DeepSeek completed full-stack heterogeneous refactoring from CUDA, rewriting over 200 core compute units across the entire stack. Industry insiders describe the difficulty as “changing an aircraft’s engine at 10,000 meters altitude.”
Key achievements:
- Ascend NPU listed in verification matrix: DeepSeek-V4 is the first to list Huawei Ascend alongside NVIDIA GPUs for official support
- Full Huawei product line support: All Ascend super-node products receive official support
- Domestic compute ecosystem expansion: Shifting from “policy procurement-driven” to “general-purpose AI infrastructure market”
This marks the official entry of the “domestic LLM + domestic compute” closed loop into large-scale commercial deployment.
Edge AI: ModelBest and Ternary LLM
ModelBest, in collaboration with Tsinghua University, released BitCPM-CANN, China’s first fully open-source ternary LLM trained end-to-end on domestic compute platforms (Huawei Ascend).
Significance of ternary LLM:
- Lower power consumption: Suitable for edge deployment
- Domestic compute throughout: From training to inference without reliance on overseas chips
- Open-source: Lowers industry barriers
On the NVIDIA side, Vera, the first CPU dedicated to agents, has officially entered production and delivery, with single-core performance up 50%. Edge AI competition is accelerating.
Industry Impact
- LLM costs continue to fall: DeepSeek’s pricing strategy forces competitors to follow
- Domestic compute market shifts from policy-driven to market-driven: Ascend moves from “being procured” to “being chosen”
- SME AI innovation barriers lower: API cost reductions directly lower trial-and-error costs
- Dual-track landscape forms: High-end scenarios use NVIDIA, domestic replacement scenarios use Ascend, expanding choice
Key Assessment
DeepSeek’s price cut is not a price war; it is the result of cost structure changes. After full-stack heterogeneous refactoring, dependence on a single vendor is reduced and bargaining power increases.
This is a positive signal for the entire domestic AI industry chain. Not a question of “whether it works,” but a question of “whether the cost is competitive.” The latter now has an answer too.
Sources: CSDN, 2026-05-27; DeepSeek Official Announcement; Huawei Ascend Launch Event