DeepSeek V3 Update Leads a New Paradigm in AI: Computing Power and Algorithm Dance Together

DeepSeek recently released a major update of version V3 on the Hugging Face platform—DeepSeek-V3-0324. This model has 685 billion parameters and has made significant improvements in code capabilities, UI design, and inference capabilities.

At the recently concluded 2025 GTC conference, Huang Renxun highly praised the achievements of DeepSeek. He pointed out that the market's previous belief that DeepSeek's efficient model would reduce the demand for high-performance chips was incorrect. Huang Renxun emphasized that future computing demands will only increase, not decrease.

As a representative work of algorithm breakthroughs, the relationship between DeepSeek and Computing Power has sparked thoughts on the roles of Computing Power and Algorithm in the development of the AI industry.

The Mutual Promotion of Computing Power and Algorithm

In the field of AI, the enhancement of Computing Power provides the foundation for more complex Algorithms to run, enabling models to process larger-scale data and learn more complex patterns. At the same time, the optimization of Algorithms can utilize Computing Power more efficiently, improving the utilization efficiency of computing resources.

The symbiotic relationship between Computing Power and Algorithm is reshaping the AI industry landscape:

1. Technical route differentiation: Some companies are committed to building ultra-large computing power clusters, while others focus on optimizing algorithm efficiency, forming different technical schools.

2. Industry Chain Restructuring: Some enterprises have become dominant players in AI Computing Power through ecosystems, while cloud service providers have reduced deployment thresholds through elastic computing power services.

3. Resource allocation adjustment: Companies seek a balance between hardware infrastructure investment and efficient algorithm development.

4. The Rise of Open Source Communities: Open source models enable the sharing of Algorithm innovations and Computing Power optimization results, accelerating technological iteration and diffusion.

DeepSeek's Technological Innovations

The success of DeepSeek is inseparable from its technological innovations. Below is a brief explanation of its main technological innovations:

Model Architecture Optimization

DeepSeek adopts a combination architecture of Transformer and MOE (Mixture of Experts), and introduces a Multi-Head Latent Attention (MLA) mechanism. This architecture functions like an efficient team, with the Transformer handling regular tasks while the MOE acts like a group of experts, calling upon the most suitable expert for specific problems. The MLA mechanism enables the model to flexibly focus on important details, further enhancing performance.

Training Method Innovation

DeepSeek has proposed the FP8 mixed precision training framework. This framework can dynamically select the appropriate computing power based on the different needs of various stages during the training process, ensuring model accuracy while improving training speed and reducing memory usage.

Improvement of Inference Efficiency

During the inference phase, DeepSeek introduced Multi-token Prediction (MTP) technology. Compared to traditional single-token prediction, MTP technology can predict multiple tokens at once, greatly accelerating the inference speed while reducing the inference cost.

Reinforcement Learning Algorithm Breakthrough

DeepSeek has developed a new reinforcement learning algorithm called GRPO (Generalized Reward-Penalized Optimization). This algorithm optimizes the model training process, achieving a balance between performance improvement and cost reduction while minimizing unnecessary Computing Power.

These innovations have formed a complete technological system that comprehensively reduces the computing power requirements from training to inference. This allows ordinary consumer-grade graphics cards to run powerful AI models, significantly lowering the threshold for AI applications and enabling more developers and enterprises to participate in AI innovation.

Impact on High-Performance Chip Suppliers

There are opinions that DeepSeek bypasses certain hardware layers, reducing dependence on high-performance chips. In fact, DeepSeek optimizes algorithms by directly manipulating the underlying instruction set. This approach tightens the binding of DeepSeek to the hardware ecosystem, while the lowering of AI application barriers may expand the overall market size.

However, the algorithm optimization of DeepSeek may change the market demand structure for high-end chips. Some AI models that originally required top-tier GPUs to run may now be able to run efficiently on mid-range or even entry-level graphics cards.

Significance for China's AI Industry

The algorithm optimization of DeepSeek provides a technological breakthrough for the Chinese AI industry. Against the backdrop of limited supply of high-end chips, the idea of "software compensating for hardware" alleviates the dependence on imported high-end chips.

Upstream, efficient algorithms reduce the pressure of computing power demand, allowing computing power service providers to extend hardware usage cycles and improve return on investment through software optimization. Downstream, the optimized open-source models lower the threshold for AI application development. Many small and medium-sized enterprises can develop competitive applications based on the DeepSeek model without the need for a large amount of computing power resources, which will give rise to more AI solutions in vertical fields.

The Far-Reaching Impact of Web3+AI

Decentralized AI Infrastructure

DeepSeek's algorithm optimization provides new momentum for Web3 AI infrastructure. The innovative architecture, efficient algorithms, and lower Computing Power requirements make decentralized AI inference possible. The MoE architecture is inherently suitable for distributed deployment, where different nodes can hold different expert networks without a single node needing to store the complete model, significantly reducing the storage and computational demands on a single node while enhancing the model's flexibility and efficiency.

The FP8 training framework further reduces the demand for high-end computing power, allowing more computing resources to join the node network. This not only lowers the threshold for participating in decentralized AI computation but also enhances the overall computing efficiency and capability of the entire network.

Multi-Agent System

1. Intelligent Trading Strategy Optimization: By analyzing real-time market data, predicting short-term price fluctuations, executing on-chain trades, and supervising trading results through the collaborative operation of multiple agents, it helps users achieve higher returns.

2. Automated execution of smart contracts: The coordinated operation of intelligent agents for monitoring, execution, and result supervision of smart contracts enables the automation of more complex business logic.

3. Personalized Portfolio Management: AI helps users in real-time to find the best staking or liquidity provision opportunities based on their risk preferences, investment goals, and financial situation.

DeepSeek innovates through Algorithm breakthroughs under Computing Power constraints, paving a differentiated development path for China's AI industry. Lowering application thresholds, promoting the integration of Web3 and AI, reducing dependence on high-end chips, and empowering financial innovation, these impacts are reshaping the digital economy landscape. The future of AI development will no longer be just a Computing Power competition, but a competition of collaborative optimization between Computing Power and Algorithm. On this new track, innovators like DeepSeek are redefining the rules of the game with unique wisdom.