When a stone hits a lake, it's not the 'plop' that captivates but the ripples it leaves behind.

DeepSeek's emergence in the AI industry is akin to such a stone, sparking a chain reaction in just over half a month: numerous companies have adopted it, cloud providers and smart computing centers have rolled out DeepSeek's inference services, and underlying ecological cooperation has intensified.

This scenario deepens our understanding of DeepSeek's impact, extending beyond performance and cost to application implementation, compute power optimization, hardware-software collaboration, and the rise of open-source ecosystems... An industrial revolution is in the making.

Here are 10 trends we've observed:

1. AI Competition's Main Stage: Shifting from Model to Engineering Innovation

Previously, large model competition centered on parameter scale and architecture optimization. However, DeepSeek, through extreme engineering innovation, has boosted training efficiency by 2-3 times and inference efficiency by 5-7 times, significantly reducing development and deployment costs. AI competition now hinges not just on model architecture but also on optimizing training and inference efficiency with limited compute power. Compute power optimization and engineering innovation are becoming core competitiveness.

2. Hardware-Software Collaborative Innovation, Driving a 'Chip-Framework-Model' Closed Loop

DeepSeek's deep integration with Ascend AI demonstrates that collaborative hardware-software optimization can substantially enhance compute power utilization and inference performance. For instance, Ascend has improved DeepSeek's inference throughput by over 30% through parallel inference optimization and expert load balancing. This indicates that chip vendors are evolving from mere hardware providers to deep participants in model optimization and framework design, propelling AI computing towards full-stack integration.

3. Compute Power Structure Evolving to 'Pre-training + Post-training + Inference'

DeepSeek's compute power demands have expanded from pre-training alone to post-training and efficient inference, complicating and diversifying the compute power architecture. Additionally, Ascend covers the entire DeepSeek process from pre-training to fine-tuning and reinforcement learning. Large models are transitioning from technological advancements to a parallel path of 'technological advancements + engineering innovation', with the compute power structure shifting from 'pre-training' to 'pre-training + post-training + inference', and compute power demands continually growing.

4. All-in-One Solutions Accelerate, Penetrating Various Industries with Large Model Services

As of February 12th, over 30 hardware and software partners have launched Ascend-based DeepSeek all-in-one solutions, supporting rapid deployment of the DeepSeek model series and meeting high-concurrency production environment needs across language understanding, image analysis, and knowledge reasoning scenarios. This 'out-of-the-box' approach will expedite large model penetration into various industries.

5. Domestic AI Ecosystem Breaking Through CUDA Barriers, Towards Independence and Controllability

While international AI giants rely on the CUDA ecosystem, domestic AI models like DeepSeek have achieved native adaptation on Ascend. Launched with Ascend support, DeepSeek has been quickly deployed by 70% of domestic customers and partners. The domestic AI ecosystem is rapidly developing, accelerating future hardware-software integration, achieving true independence and controllability, and enhancing China's AI industry's global competitiveness.

6. Large Model Popularization Ushers in the 'Hundred Models, Thousands of States' Paradigm

DeepSeek has significantly reduced model fine-tuning costs through engineering optimization, making it easier for companies to customize their models. Currently, DeepSeek's industry adaptation covers multiple sectors like internet, finance, government, and education, fostering the 'hundred models, thousands of states' AI landscape.

7. Declining Large Model Training Costs Fast-Tracks AI Startups

DeepSeek has cut training costs by 60% and enhanced model performance by 20%, lowering barriers for AI startups. With declining compute costs, more AI startups will emerge, diversifying AI innovation and driving the industry into a faster development phase. For instance, after jointly launching the DeepSeek model series with the Ascend Community, Mole Community saw over 20,000 downloads within a week, making it China's most popular open-source model community.

8. AI Computing and Large Model Ecosystems Trending Towards Greater Openness

While some AI giants enclose their underlying technologies, DeepSeek, Ascend, and others adhere to open-source ecosystems. MindSpore, an open-source AI framework, has become China's fastest-growing AI framework, accounting for 30% of the market share. In the future, open-source ecosystems will offer greater advantages in industrial implementation and technological innovation, steering AI computing and large models towards greater openness.

9. Edge AI and Large Model Integration Births New Application Modes

Ascend's end-edge-cloud collaborative capabilities enable DeepSeek to operate efficiently not just on cloud computing but also on edge devices. For example, through Ascend's algorithm optimization, DeepSeek has improved inference throughput by 20%~35% in industrial scenarios. Large models will no longer be cloud-confined but will penetrate terminals like smartphones and IoT devices, opening up broader AI application avenues.

10. Compute Power as a Service (CaaS) Becoming AI's Mainstream Business Model

A stable and reliable compute power base remains crucial for model innovation. Ascend is currently assisting over 20 provinces and 25 cities in China to deploy DeepSeek models on public compute power service platforms, with local support available in over 70% of regions nationwide. AI computing will shift towards a 'Compute Power as a Service' model, allowing companies to focus on model applications without the underlying compute power architecture's complexity, thereby fostering industry-wide innovation.

Summary

DeepSeek's rise is not just an innovative revolution for large models but also a comprehensive upgrade of AI infrastructure, computing models, and business ecosystems. Compute power optimization, full-stack integration, domestic ecosystem independence, AI application industrialization, edge intelligence... The ripples created by DeepSeek are spreading across upstream and downstream industries, reshaping the global AI industry's competitive landscape.