DeepSeek Heralds the Golden Age of Bootstrapped Entrepreneurship

I’m publishing this series to discuss a topic that I follow closely - cloud stocks, trends, strategy, acquisitions, and more. Please subscribe to my Cloud Stock Analysis series to never miss an article.

Early this year, AI development firm DeepSeek took the tech world by storm with its cost-effective AI models. It reached 100 million users in just 20 days while ChatGPT reached the same milestone in two months. The cherry on the cake is that it took a fraction of the cost. That’s truly inspiring for bootstrapping entrepreneurs!

DeepSeek’s History

Founded by Liang Wenfeng in 2023, DeepSeek is based in Hangzhou, China. Prior to founding DeepSeek, Liang had co-founded High-Flyer, a quantitative trading fund in 2016. By 2019, Liang had established High-Flyer as a hedge fund focused on developing and using AI trading algorithms. It had invested 200 million RMB ($28 million) in developing Fire-Flyer One, its proprietary deep learning training platform.?

By 2021, High-Flyer had surpassed the 100 billion RMB ($14 billion) milestone and was recognized as one of the “Four Kings of Quant-Investing”. It then invested 1 billion RMB ($140 million) in Fire-Flyer Two using around 10,000 Nvidia A100 GPUs.

However, the US restriction on chip sales to China that year meant that Liang’s team had to be innovative in maximizing the computing power of the chips they had. In 2023, Liang took the best people in this team to found DeepSeek with investment from High-Flyer. In November that year, it released its first model called DeepSeek Coder followed by the DeepSeek-LLM series of models.

In January 2024, it released two mixture of experts (MoE) models: Base and Chat. In April, it released?three Math models: Base, Instruct, and RL. It then released DeepSeek-V2 followed by the DeepSeek-Coder V2 series. DeepSeek V2.5 was released in September and updated in December 2024. In December 2024, it released the base model DeepSeek-V3-Base and the chat model DeepSeek-V3.?

On 20 January 2025, DeepSeek released the DeepSeek chatbot, based on the DeepSeek-R1 model, free of charge for iOS and Android. By 27 January, DeepSeek had surpassed ChatGPT as the most downloaded freeware app on the iOS App Store in the United States.

The Way Forward

DeepSeek claims to have used just 2,048 Nvidia H800s and $5.6 million to train a model with 671 billion parameters and 14.8 trillion tokens. GPT 4 was trained on over one trillion parameters with 13 trillion tokens at a cost of over $100 million.?

DeepSeek is also priced about 27 times lower than OpenAI. DeepSeek R1 is priced at $0.55/1 M input tokens and $2.19/1M output tokens while OpenAI o1 is priced at $15/1M input tokens and $60/1 M output tokens.

DeepSeek has several competitors such as OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini, Meta’s Llama3, and Perplexity AI. However, they are not as capital-efficient.

The arrival of capital-efficient AI models like DeepSeek will reduce the costs massively and can lead to the golden age of entrepreneurship. We could also see more DeepSeek type companies based in the US that will democratize AI.?

As AI becomes more accessible, the focus will shift from computing power to applying AI cost effectively for different verticals like healthcare.

Looking For Some Hands-On Advice?

If you’re looking to work with Sramana Mitra?in private,?please check?1-on-1 Consulting. If you are comfortable discussing in public, you can also come to our Free Mentoring Roundtables.

Photo Credit: Trong Khiem Ng/Flickr.com