LIVE Web3 News

ARTICLE INFORMATION

DeepSeek AI’s large language model disrupts industry with low-cost, high-efficiency approach

by ICN.live
May 14, 2025

Amira Khalil

DeepSeek AI’s large language model is redefining expectations in the artificial intelligence space.

Launched in January 2025, DeepSeek’s R1 model has quickly become one of the most-discussed names in AI. The company behind it, Hangzhou-based DeepSeek, is funded by hedge fund High-Flyer and led by founder Liang Wenfeng. Despite being new, it already competes with giants like OpenAI and Meta.

DeepSeek AI large language model stands out for its affordability and performance. Its V3 model was reportedly trained for just $6 million. That’s a fraction of the $100 million OpenAI spent to develop GPT-4. What’s more, DeepSeek used fewer AI chips, some even weaker ones, due to U.S. trade restrictions. Still, they managed to match the results of much more expensive models.

DeepSeek challenges AI norms with cheaper, smarter training

One technique behind this cost-cutting success is the mixture of experts (MoE) layers. These allow different parts of the model to be activated only when needed, saving power. DeepSeek also released its R1 model under an open-weight license. This means other developers can study and use its architecture, though some conditions apply.

The company’s training was done on China-compliant Nvidia chips in a custom-built cluster. Known as Fire-Flyer 2, the cluster housed 5,000 GPUs. By managing parallelism and optimizing compute use, DeepSeek reduced the resource load significantly.

Recruitment also plays a role in DeepSeek’s edge. It attracts top Chinese AI researchers and also hires from non-traditional backgrounds. That diversity enhances its model’s knowledge and performance.

DeepSeek AI large language model sends shockwaves across the tech world

The effects of DeepSeek’s rise are already being felt. Nvidia’s stock dropped dramatically, wiping out $600 billion in value—the biggest single-company market cap loss in U.S. history. Investors saw DeepSeek’s success as a challenge to the hardware giant’s dominance.

Meanwhile, DeepSeek’s open-weight strategy and rapid development have forced rivals to rethink their approach. The startup has also sparked new conversations about AI access, hardware independence, and how to build smarter, not just bigger.

With more cost-efficient tools, DeepSeek AI’s large language model proves you don’t need to be a tech titan to innovate in AI. Whether this leads to a broader wave of affordable AI remains to be seen, but DeepSeek has certainly changed the game.

What is DeepSeek and how is it different from other AI companies?

DeepSeek is a Chinese AI company founded in 2023 that develops large language models (LLMs). It’s backed by the hedge fund High-Flyer and led by Liang Wenfeng. What sets DeepSeek apart is its ability to deliver GPT-4-level AI models at a fraction of the cost. Its DeepSeek-R1 model, for instance, was reportedly trained for only $6 million. The company also emphasizes efficiency, using a technique called mixture of experts (MoE) and training with fewer or less powerful chips. Unlike many Western firms, DeepSeek shares its model parameters openly under an ”open-weight” license, allowing others to learn from and build on its work.

How did DeepSeek train its AI with such low cost?

DeepSeek kept costs down by using efficient architecture designs like mixture of experts (MoE), and by training on locally available chips. Instead of using high-end GPUs that are difficult to access due to U.S. export restrictions, DeepSeek relied on weaker, export-approved versions and optimized training across fewer units. It also utilized a proprietary GPU cluster (Fire-Flyer 2), carefully managing resource use. This strategy enabled DeepSeek to reduce both energy consumption and computational demand, without compromising on model performance.

Why did Nvidia’s stock drop due to DeepSeek’s success?

Nvidia’s stock fell dramatically—losing $600 billion in market value—because DeepSeek demonstrated that high-performance AI can be achieved without relying on the most powerful Nvidia GPUs. By successfully training competitive models using fewer and weaker chips, DeepSeek signaled a potential shift away from Nvidia’s expensive hardware. Investors reacted strongly to this disruption, fearing that Nvidia’s market dominance in AI chips might be undermined. DeepSeek’s achievement shows that software innovation and training strategy could become more important than raw hardware power.

Is DeepSeek AI large language model open source?

DeepSeek’s model is considered ”open-weight” rather than fully open-source. This means that while the model’s parameters are publicly available for research and development, there may be usage restrictions that differ from traditional open-source licenses. For example, the company might limit commercial usage or redistribution of its model in certain ways. However, this open-weight approach still allows the community to explore the model architecture and test it, encouraging transparency and collaboration in the AI field.

MORE FROM SPONSORED

LIVE Web3 News

ARTICLE INFORMATION

DeepSeek AI’s large language model disrupts industry with low-cost, high-efficiency approach

DeepSeek challenges AI norms with cheaper, smarter training

DeepSeek AI large language model sends shockwaves across the tech world

FEATURED

ICN TALKS EPISODES

Unlock Crypto's Most Exclusive Token Launch Services

Bimonthly Iron Newsletter