China’s DeepSeek and why is it freaking out the AI world?

 

China's DeepSeek is an AI company that has recently gained attention for its highly efficient and powerful AI models, which some analysts believe could rival OpenAI. The company's latest model, DeepSeek V3, is one of the largest open-weight AI models, boasting 671 billion parameters—making it larger than Meta's Llama 3.1 405B. What’s surprising is that DeepSeek managed to train this model using only $5.5 million and 2,048 Nvidia H800 GPUs over two months, a fraction of the cost and hardware used by companies like OpenAI and Google, which typically rely on 16,000+ high-end GPUs for training frontier models.

DeepSeek has also introduced DeepSeek-R1, a reasoning-focused AI model, aimed at competing with OpenAI’s advanced models. Notably, it offers significantly lower pricing than OpenAI—charging just $2.19 per million tokens, compared to OpenAI's $60 per million tokens. This drastic price difference has led to concerns about the profitability of Western AI companies, as DeepSeek could potentially force them to lower prices.

The company's unexpected efficiency in developing such powerful AI despite U.S. chip restrictions has also raised questions. DeepSeek claims it has only a limited supply of Nvidia GPUs, but some experts suggest that the company may have access to tens of thousands of high-end chips despite U.S. export bans. This has led to growing concerns in Washington and Silicon Valley about China’s increasing AI capabilities and whether U.S. attempts to slow its progress are failing.

Meta and other tech firms are reportedly studying DeepSeek’s models closely, trying to understand how they were trained so efficiently. Meanwhile, the AI industry is reacting to DeepSeek’s rise, with Nvidia’s stock dipping 15% amid fears that its high-end AI chips might not be as essential for training massive models as previously thought.

However, DeepSeek’s models remain subject to Chinese government regulations, meaning they avoid politically sensitive topics and adhere to China’s core socialist values. This limits their usability in certain applications compared to models from OpenAI or Anthropic.

Why is DeepSeek "freaking out" the AI world?

1. Low-cost AI development – Training at a fraction of OpenAI or Google’s costs.

2. Comparable or better performance – Its models challenge top-tier Western AI.

3. U.S.-China AI rivalry – Raising concerns about China's AI advancements despite U.S. restrictions.

4. Potential market disruption – Lower pricing could force competitors to reduce costs.

5. Uncertainty over GPU access – Questions about how DeepSeek acquired high-end hardware.