DeepSeek Goes Viral: Will The AI Boom Surge Again?

Recently, DeepSeek has dominated the tech world’s attention. Developed by the AI startup under High-Frequency Trading firm Qifan, its large-scale model soared to the top of global app store rankings. This breakout success ignited investor enthusiasm, driving AI-related stocks such as DeepSeek concepts and cloud computing indices up over 10% post-Lunar New Year, with computing and media sectors also leading gains.

DeepSeek’s breakthroughs lie in cost efficiency and reasoning capabilities. Its V3 model significantly improves training cost and computational efficiency, while the R1 model pioneers a novel approach by optimizing FP8, MoE, MLA, and PTX techniques. This maximizes computational resource utilization, slashes costs (just 1/30 of OpenAI’s equivalent), and delivers exceptional performance in math, coding, and natural language tasks—comparable to OpenAI’s o1 model.

Traditionally, AI models relied on RLHF (Reinforcement Learning from Human Feedback), where human-curated responses guided the model’s learning, akin to a teacher correcting mistakes. However, R1 disrupts this by eliminating human feedback (HF) and adopting pure reinforcement learning (RL), allowing the model to learn through trial and error independently.

R1 defines two reward functions:

1. Correctness Reward – The model validates answers through external tools, rewarding itself only when they are verifiably correct.
2. Logical Process Reward – Even if an answer is incorrect, a well-reasoned thought process still earns rewards, fostering strong logical reasoning skills.

This approach enables R1 to refine problem-solving strategies autonomously, reducing reliance on labeled data while closing the gap between open-source and proprietary AI models. It also narrows China’s AI technological gap with the U.S.

DeepSeek’s rise signals a more diverse AI landscape, with cost reductions accelerating commercial AI applications. This is expected to drive exponential growth in inference-side computing demand, benefiting inference chips and edge computing.

Meanwhile, the AI startup ecosystem is shifting. Companies dependent on expensive GPU clusters face financial risks, potentially flooding the second-hand GPU market. Smaller AI firms will likely favor cost-effective, lower-end GPUs. The industry is also witnessing a structural transition—compute demand is moving from training to inference, a trend expected to continue. Instead of building foundation models from scratch, more companies may adopt open-source solutions like DeepSeek, challenging NVIDIA’s dominance in training hardware.