China’s AI breakthrough: Through hardware adaptation and algorithm optimization, the training cost has been reduced by 20%. For example, the cost of 1 trillion tokens has been reduced from 6.35 million to 5.08 million, accelerating the popularization of technology.
Core technology breakthrough: Ant Group Ling team’s innovative practice
Ant Group AI
1. Empirical data of 20% cost reduction
- Model parameters and training costs:
The Ling team led by Ant Group’s Chief Technology Officer He Zhengyu developed two hybrid expert models (MoE), Ling-Lite (16.8 billion parameters) and Ling-Plus (290 billion parameters). - Training cost comparison:
- The cost of training 1 trillion tokens using high-end GPUs (such as NVIDIA H100/H800) is 6.35 million yuan.
- After adopting a low-end hardware system, the cost dropped to 5.08 million yuan, a decrease of 20%.
- Performance benchmarking: The model’s performance in benchmarks such as English comprehension and mathematical reasoning is comparable to top models such as Qwen2.5-7B and DeepSeek-V2.5, and even better in some tasks.
Technical comparison: Collaborative breakthrough with DeepSeek
2. Technical Implementation Path
- Hardware Adaptation:
- Through the compatible technical framework of heterogeneous computing units and distributed clusters, the model can be seamlessly switched between different hardware.
- Experiments show that the training effect of the hybrid expert model with a parameter scale of about 300 billion on low-end devices is not significantly different from that on high-end devices.
- Algorithm Optimization:
- Dynamic Sparsification MoE Architecture: Only activate expert modules in specific fields to reduce inference costs by 30%-50%.
- Model Quantization and Pruning: Convert floating point numbers to integers to reduce memory usage and computing requirements.
3. Open Source and Application Prospects
- Open Source Plan: Ling-Plus and Ling-Lite will be open source to support scenarios such as medical image analysis and financial risk control.
- Industry Landing: It has been applied to industrial quality inspection (such as Foxconn iPhone motherboard production), logistics optimization (SF Express business planning) and other fields.
Technical comparison: Collaborative breakthrough with DeepSeek
1. DeepSeek’s “low-cost-high-performance” path
- Architecture innovation:
- Dynamic sparse MoE: Only 37 billion activation parameters can approach the performance of the world’s top models.
- Long context processing: Supports 128K tokens of ultra-long text, and significantly improves logical coherence.
- Training cost:
- DeepSeek-V3 has a total of 671 billion parameters, and the cost of complete training is only 5.576 million US dollars (about 37 million RMB), which is much lower than similar models.
2. Domestic chip replacement acceleration
- Hardware adaptation: Ant Group uses a mixture of domestic chips and NVIDIA chips in training to challenge the monopoly of high-end GPUs.
- Industry impact: Promote the penetration rate of domestic chips such as Huawei Ascend and Baidu Kunlun in AI training.
Policy and ecological support: the formation of Beijing AI innovation highland
1. Computing infrastructure construction
- Beijing artificial intelligence public computing platform:
- A leading ultra-large-scale computing cluster has been built in China, and the capacity is continuously expanded to support low-cost model training.
- As a national computing power scheduling demonstration sample, it provides infrastructure support for technological breakthroughs.
2. Talent training and introduction
- “AI100 Youth Pioneer” program:
- Select 65 AI research and industrialization talents under the age of 40, including Bao Fan of Shengshu Technology and Zeng Guoyang of Mianbi Intelligence.
- Haidian gathers more than 10,000 AI scholars, including more than 100 AI2000 global top scholars.
3. Industrial policy support
- “Zhongguancun AI Beiwei Community”:
- Provide office space with up to three years of rent-free, 1,000 talent apartments and supporting incubation services.
- The goal is to create a “global young AI talent habitat” and a trillion-level industrial cluster.
Industry applications and future challenges
1. Typical application scenarios
- Agriculture: Henan Yunfei Technology uses DeepSeek to predict crop pests and diseases, increasing production by 18%.
- Manufacturing: Geely Automobile introduces DeepSeek to optimize the intelligent interactive system, and Foxconn improves production efficiency.
- Logistics: SF Express predicts order fluctuations through AI models to reduce operating costs.
2. Technical challenges and ethical supervision
- Energy consumption issues: AI data centers are expected to account for 2.5% of the total power consumption of the whole society, and breakthroughs in energy-saving technologies such as liquid cooling servers and photonic computing are required.
- Data privacy: The proportion of synthetic data is expected to reach 45%, and the risk of “data pollution” needs to be prevented.
- Regulatory policies: The “Interim Measures for the Management of Generative AI Services” promotes model transparency and balance between commercial interests.
Future trends: From technical independence to ecological reconstruction
- Technical direction:
- Reinforcement learning fusion: Optimize the reasoning path through unsupervised reinforcement learning and reduce invalid calculations.
- Multimodal large models: such as Emu3 of Zhiyuan Research Institute, which integrates text, image, and video data to improve industrial prediction capabilities.
- Industry landscape:
- China’s AI has turned from a “follower” to a “competitor”, 7 laboratories have launched cutting-edge reasoning models, and the open source ecosystem has gradually matured.
- It is estimated that the scale of China’s AI chip market will reach 153 billion yuan in 2025, a year-on-year increase of 42%.
Conclusion
China’s AI has achieved a breakthrough of significantly reducing computing costs by 20% through hardware adaptation, algorithm optimization, and policy support.
Ant Group and DeepSeek’s technical practice not only challenges the monopoly of high-end chips, but also promotes AI from laboratories to the core of the industry.
In the future, with the popularization of low-cost models and the rise of domestic chips, China’s AI is expected to trigger a deeper efficiency revolution in the fields of medical care, manufacturing, logistics, etc., and accelerate the reconstruction of the global AI ecosystem.
About All AI Applications
AllAIApplications.com focuses on all AI applications, AI tools, global AI ideas and AI product advertising services. We are committed to AI technology research, AI idea exploration, AI for business, AI for education, AI for cross-border e-commerce and other projects. If you have this idea, please contact us to discuss it together.
If you need to place AI ads on our website, or you are a traditional smartwatch manufacturer, you can also contact us at any time.
FAQs
Through hardware adaptation + algorithm optimization dual path implementation:
Hardware: Use domestic chips (such as Huawei Ascend 910, Ali Hanguang series) to replace some NVIDIA high-end GPUs, combined with heterogeneous computing unit scheduling.
Algorithm: Use Mixed Expert Model (MoE) to dynamically activate sub-modules in specific fields and reduce redundant calculations; with model quantization and pruning technology, further reduce costs.
Cost of training 1 trillion tokens: from 6.35 million yuan to 5.08 million yuan, a decrease of 20%.
Energy consumption: Reduced by 15%, hardware investment reduced by 20%.
Case: Ant Group Ling-Plus model completed 9 trillion token pre-training under five hardware configurations, saving nearly 20% of costs.
Medical: Disease prediction (such as Henan Yunfei Technology’s agricultural pest prediction), image analysis.
Manufacturing: Foxconn uses AI to optimize iPhone production processes.
Finance: Ant model is used for risk control and fraud detection.
Logistics: SF Express reduces operating costs through AI planning.
Computing power infrastructure: Beijing’s artificial intelligence public computing power platform expands to support low-cost model training.
Talent training: The “AI100 Young Pioneers” program attracts global scholars, and Haidian gathers more than 10,000 AI talents.
Industry support: Zhongguancun AI Beiwei Community provides rent-free office space to promote trillion-level industrial clusters.
Energy consumption: AI data centers consume 2.5% of the total electricity in the society, and breakthroughs in energy-saving technologies such as liquid cooling and photonic computing are needed.
Data privacy: Synthetic data accounts for 45%, and the risk of “data pollution” needs to be prevented.
Model generalization: Improve cross-scenario adaptability and avoid overfitting in a single field.
Market size: The market size of China’s AI chip is expected to reach 153 billion yuan in 2025, a year-on-year increase of 42%.
Technical direction: Reinforcement learning fusion and multimodal large models (such as Emu3 of Zhiyuan Research Institute) improve industrial forecasting capabilities.
Ecological reconstruction: Low-cost models accelerate efficiency revolutions in medical care, logistics and other fields, and promote global AI ecological changes.
Benchmark test: The Ant model performs close to NVIDIA H800 in tasks such as English comprehension and mathematical reasoning, and is better in some scenarios.
Training efficiency: Through algorithm optimization, the training effect of low-end hardware is no different from that of high-end GPUs.
Ant Group: Ling-Lite (16.8 billion parameters) and Ling-Plus (290 billion parameters) models have been open sourced to support medical, financial and other fields.
Technical tools: Open source framework DLRover and reasoning framework Flood help developers build an application ecosystem.
Technology autonomy: Challenge Nvidia’s computing power hegemony and promote the application of domestic chips such as Huawei Ascend and Baidu Kunlun.
Supply chain security: Reduce dependence on Western technology and accelerate the process of domestic substitution.
Policy level: The “Interim Measures for the Management of Generative AI Services” promotes model transparency.
Technical level: Secure federated learning, model harmlessness optimization (such as Ling-Plus balancing security and usefulness).
