DeepSeek's Counterattack: The Spring of Edge Computing is Here

In early 2025, the technology world witnessed a "shockwave" — Chinese startup DeepSeek, with only $5.576 million in training costs, created an AI large model comparable to OpenAI's GPT-3.

In contrast, OpenAI's equivalent model burned through $500 million, and Elon Musk's Grok3 used 200,000 Nvidia GPUs, costing over $6 billion. DeepSeek's cost was less than 1% of OpenAI's, yet it achieved similar performance, directly overturning the traditional "spending big to achieve miracles" AI training model.

Previously, the AI industry followed the logic of "stacking parameters, burning computing power, and spending money to capture markets," but DeepSeek's success proves a new possibility: "small models + efficient computation" is the future.

Netizens humorously say, "OpenAI burns money, DeepSeek burns brains." This revolution not only challenges the computing power hegemony of industry giants but also shifts edge computing to the forefront of AI development.

The Rise of Small Models: How DeepSeek Achieved Low-Cost, High-Performance AI

DeepSeek's success can be attributed to its optimized "small model strategy." This strategy, driven by three core technologies, has successfully redefined AI training and inference:

1. Computing Power Optimization: Maximizing GPU Efficiency

· Distillation Training: DeepSeek employs "distillation" to transfer the core capabilities of large models into smaller models, removing redundant computations while preserving high performance. This ensures that every ounce of computing power is maximized.

· FP8 Low Precision Computation: While traditional AI training uses FP16, DeepSeek adopts FP8, offering faster computation and lower memory usage. This even allows models to run on ordinary consumer devices.

· MOE (Mixture of Experts) Architecture: By intelligently assigning tasks, DeepSeek activates only the necessary computation units of the model, achieving over 70% effectiveness with only 40% of the computation power — a strategy akin to "AI portioning."

2. Low-Cost Inference: Making Edge Computing Feasible

DeepSeek not only optimized the training phase but also pushed for widespread edge computing adoption during inference:

· JanusFlow-1B Model: After just one week of training using 128 A100 GPUs, DeepSeek’s model can generate high-quality images and support WebGPU for browser-based inference, without relying on expensive cloud computing.

· Edge Computing Advantages: DeepSeek's small AI models run efficiently on mobile phones, industrial computers, and IoT devices in low-power, low-latency environments, aligning perfectly with the demands of 5G edge computing. This shift eliminates the need for heavy cloud support, making AI applications more flexible and cost-effective.

3. Open Ecosystem: Open Source + Low-Cost APIs, Capturing the Market

DeepSeek employs a "rural encircles the city" strategy by open-sourcing part of its model weights and offering extremely low-priced APIs (0.5 RMB per million tokens for input and 8 RMB for output). This drastically reduces the AI usage costs for enterprises and developers.

In comparison, OpenAI’s APIs are five times more expensive, and its models are closed, making deep customization challenging. DeepSeek's open strategy allowed it to quickly integrate into ecosystems like WeChat, Huawei Cloud, and Tencent Cloud, becoming widely used in innovative applications such as AI-based fortune-telling and lottery prediction.

The Mid-Life Crisis of AI Giants: The Dilemmas of Large Models

DeepSeek’s rise is not only a technological breakthrough but also a mirror reflecting the deep issues in the current AI industry. For years, AI giants relied on massive computing power, excessive spending, and closed ecosystems to maintain their technological lead. However, when DeepSeek broke these rules with "small models + extreme optimization," the limitations of large models became more evident.

1. Uncontrolled Training Costs

OpenAI’s "Interstellar Gateway" project has a budget of up to $50 billion, while DeepSeek achieved similar results with just a fraction of that cost. This has led the industry to question: is it really necessary to invest so much computing power in large models? As AI training costs grow exponentially, the performance improvements become marginal, and the "computing power black hole" is becoming an unbearable burden for the industry.

2. High Performance Doesn’t Always Mean Practicality

In fields like financial risk control and medical diagnosis, users need precise inferences, not lengthy AI analysis reports. In industrial inspection and autonomous driving, what matters most is sub-second response times, not the complex calculations of large models. Large models often act like a sledgehammer to crack a nut: expensive, cumbersome, and resource-intensive. DeepSeek's small models, on the other hand, offer better cost-effectiveness in these scenarios.

3. The Closed Ecosystem is Becoming an Obstacle

The APIs of AI giants like OpenAI and Google are expensive and highly closed, making it difficult for companies to customize and deeply apply AI technologies. In contrast, DeepSeek’s open-source and low-cost API strategy not only makes AI more accessible to developers but also ensures the circulation of technology across the industry, attracting more companies and small teams to the ecosystem. This open model rapidly expanded DeepSeek’s influence, forming a distinct development path from traditional AI giants.

DeepSeek's Rise: Three Key Transformations in AI Industry

DeepSeek’s success signals three key directional shifts in the AI industry:

1. From "Parameter Stacking" to "Efficiency Competition"

Traditionally, AI’s competitive focus was on the size of the model parameters — the larger the model, the more computing power it had. However, as technology matures, the core competitiveness of AI will shift from the quantity of parameters to "computing power output per dollar." This means the AI industry will focus more on how to efficiently utilize computing resources rather than simply pursuing larger models. DeepSeek’s extreme optimization strategy demonstrates that even small models can deliver immense performance potential with limited resources.

2. From "Cloud Computing" to "Edge Computing"

With the widespread use of 5G and IoT, computing demands are shifting from traditional cloud computing to edge devices. Small models with low power consumption and high responsiveness are becoming essential in various industries. DeepSeek's successful models demonstrate that small AI models can provide powerful intelligence while running directly on mobile phones, industrial devices, and IoT terminals, significantly reducing dependence on cloud supercomputing. This transition makes AI applications more flexible and real-time while aligning with the needs of edge computing in the 5G era.

3. From "Closed Ecosystems" to "Open Collaboration"

DeepSeek is challenging traditional giants like OpenAI and Google with its open-source and low-cost API strategy. This open approach not only lowers the barrier to AI technology but also fosters more innovation. The combination of open-source models and affordable APIs allows developers to access and customize AI technology at a lower cost, while pushing AI toward a more inclusive direction. This shift enables small and medium-sized enterprises and independent developers to participate in AI innovation and application, breaking the monopoly of large tech companies over AI technologies.

Conclusion: DeepSeek Leads the New AI Revolution, Edge Computing as the Biggest Winner

DeepSeek’s counterattack is akin to the "four taels of weight" maneuver in martial arts. While tech giants are still competing over who has the most computing power, DeepSeek uses its "small model + extreme optimization" strategy to create a crack in AI’s computing power hegemony. The "small vs. big" battle is just beginning, and edge computing is becoming the biggest beneficiary of AI transformation.

The edge computing community will continue to monitor DeepSeek’s breakthroughs, working alongside developers and enterprises to explore this emerging field and seize the opportunities brought by the AI wave. As AI technology evolves, small model applications will continue to expand, and the value of edge computing will be further unlocked. In this revolution, every developer, enterprise, and tech enthusiast is an important driving force for the future of the intelligent ecosystem.