The chip technology war has created a multidimensional impact on Pinduoduo’s data center construction, which can be broken down into five major dimensions:
1. Reshaping Policy Background
1.1. Mandatory Domestic Substitution
• The new regulations from the U.S. BIS restrict the export of equipment below 14nm, forcing Pinduoduo to switch 30% of its servers to domestic chips (such as Haiguang 7285) in 2023.
• Participation in the Ministry of Industry and Information Technology’s “Xinchuan Cloud” pilot program, gaining additional points in the government procurement catalog.
1.2. Data Sovereignty Legislation
• The “Cybersecurity Review Measures” require core business data to be stored in self-controlled facilities.
• Forced to abandon the Singapore backup data center plan, shifting to the Guizhou “East Data West Computing” hub.
2. Intensified Supply Chain Risks
2.1. Chip Supply Crisis
• The Nvidia A800/H800 sales ban has delayed the expansion of AI training clusters by 6-8 months.
• Urgent procurement of Huawei Ascend 910B (with a performance loss of about 25%).
2.2. Domestic Equipment Pain Points
• The yield of the Zhongwei Semiconductor etching machine is 18% lower than ASML, increasing server depreciation costs.
• The latency fluctuation of Changxin Memory’s DDR4 chips exceeds 30ns, necessitating a reconstruction of the caching mechanism.
3. Challenges in Technical Substitution
3.1. Heterogeneous Computing Adaptation
• The self-developed TDDL database has restructured the query optimizer for Kunpeng chips (with an investment of over 300 person-months).
• The Cambricon MLU270 accelerator has a throughput in recommendation scenarios that is only 62% of the A100.
3.2. Fragmentation of Software Ecosystem
• TensorFlow’s support for Ascend NPU is lagging, forcing the use of the PaddlePaddle framework.
• The lack of an open-source RISC-V toolchain has led to the establishment of an in-house compiler team (with annual costs exceeding 200 million).
4. Transmission of Cost Pressures
4.1. Surge in Direct Costs
• The purchase price of domestic servers has increased by 40% year-on-year (compared to DELL PowerEdge).
• The PUE of liquid cooling retrofitting has risen from 1.25 to 1.38 (due to the efficiency issues of domestic refrigerants).
4.2. Explosion of Hidden Costs
• The expenditure on retraining technical personnel has surged from 5% to 15% of the IT budget.
• The complexity of multi-architecture operations has tripled the time required for fault diagnosis.
5. Pinduoduo’s Strategies for Breaking the Deadlock
5.1. Technology Layering Strategy
• The core transaction system employs a “Haiguang + Ascend” dual redundancy.
• Edge nodes utilize a mixed architecture of RISC-V + LoongArch.
5.2. Policy Arbitrage Model
• Obtaining tax reductions through the “New Infrastructure for Agriculture” project (saving 470 million in taxes in 2023).
• Collaborating with the Chinese Academy of Sciences to develop integrated storage and computing chips (receiving funding from the Ministry of Science and Technology’s 02 special project).
5.3. Scenario-Based Dimensionality Reduction
• The agricultural product traceability scenario has switched to a lightweight model (reducing parameter volume to 1/10).
• Utilizing the “100 Billion Subsidy” data to train small sample learning algorithms (reducing chip dependency).
Currently, the domestic production rate of its data center has reached 58% (up from only 12% in 2021), but achieving full-stack self-control will still require a 3-5 year effort.