DeepSeek is everywhere, but why are we hearing so much about it lately: there might be two main reasons: the chip-making giant Nvidia shed almost $600bn (£482bn) of its market value on January 20th 2025 – the most significant one-day loss in US history, and the other event is DeepSeek AI dominated the apple store by getting 2 million downloads within just few weeks of release, surpassing the AI pioneer ChatGPT. Could these incidents be part of a bigger trend?

Surprisingly, yes. The advent of DeepSeek’s cost-effective AI model that operates using low-power chips created turbulence for Nvidia, a dominant player in the AI chip market, raising questions regarding the future demand for high-power GPUs. Since DeepSeek AI matched capabilities with OpenAI’s GPT-4, Meta’s Llama, and Google’s Gemini, there are also questions concerning the impact on the revenue projections of these top leaders under the AI market umbrella. 

What is DeepSeek AI? From where did they emerge?

DeepSeek is a Chinese AI company founded in 2023 by Liang Wenfeng. It did not emerge out of nowhere. The AI company has consistently released various versions throughout the last year. Still, the latest model, R1, turned heads in the generative AI world because of its low cost of operations, open-source version, and groundbreaking reasoning capabilities that match popular AI giants OpenAI, Meta, and Google.

But how did DeepSeek take the internet by storm? Top 5 reasons

The Low cost of operations

The number one reason for the hype is that AI pioneers are spending millions if not billions,  dollars on AI development; meanwhile, the Chinese company has claimed that its operation cost is $5.6 million, a fraction of their investment. This has prompted top investors in AI companies to question whether they are bleeding money when such remarkable results can be achieved with relatively modest investments.

This prompts another question: how did they make this possible at such a low cost? The reason is that popular AI players like ChatGPT use a supervised learning method to train their LLMs, which, done on a large scale, requires significant investments in computing power, data collection, and energy consumption. DeepSeek was able to achieve this because they used a model called the reinforcement model, which does not require huge data collection and computational power. Reinforcement learning is a self-learning model that uses trial and error to generate final responses. This way, DeepSeek could deliver promising results at an affordable price.

Open-source model

Contrary to leading AI companies with a closed AI model setup, DeepSeek has made its R1 model publicly available in the popular AI development platform Hugging Face by securing MIT licenses, which gives accessibility to commercial users without any restrictions. This means anyone can use, train, and build the model. 

This initiative has reopened closed doors for startups worried about funds and high investment costs and ensures a promising AI environment to achieve their growth. Making R1 open can significantly help companies to have a much lesser API cost, which is approximately 98% lower than OpenAI’s O1 model.

Low power chips

What makes this even more interesting is that DeepSeek has matched reasoning capabilities with ChatGPT- 4 and Claude 3.5, outweighing the US National Security Policy restricting access to high-power AI computer chips. It is reported that DeepSeek achieved such capabilities using low-power alternative Nvidia H800, which has many limited capabilities, compared to Nvidia H100, which is used by OpenAI. This fact has reignited the innovation spark in many countries where high-power hardware is not readily available.

Benchmark AIME 2024

As per Artificial Intelligence in Medicine (AIME) 2024, The R1 model has performed better than OpenAI’s O1 model in specific benchmarks such as MATH-500, which focuses on math problems, and SWE-bench, which is based on programming tasks. Though OpenAI excels in Natural Language Processing, DeepSeek, the reasoning model, performs better in maths, physics, coding, and more. Their reasoning capabilities are strikingly good, and Sam Altman, the CEO of OpenAI, has called DeepSeek “an impressive competitor.”

Chain of thought process

Another transformative feature of DeepSeek is its thought-chain (TOC) approach, wherein the model does not give an immediate response but instead thinks through and displays its thinking process step by step, like how a human would feel when faced with a complex query. This unique feature makes the multi-step reasoning process highly transparent, allowing users to interrupt, analyze, and critique until it arrives closer to their desired output.  

It’s not just an underdog tale; there’s more to it.

Environmental impact

The AI industry is energy-intensive to the extent that top AI leaders are considering investing in nuclear power plants to support AI data centers. AI, being such a high energy-demanding sector, is massive in terms of investments and is also said to create tension in power grids worldwide. To be precise, it is not a future problem anymore, as it has already started disrupting power flow in countries like Saudi Arabia, Malaysia, and Ireland, Bloomberg said. This massive surge in energy demands is a setback for countries considering AI space as a new venture. Existing market leaders are trying hard to meet their ends by balancing AI development and environmental impact, especially keeping up with their decarbonization goals.

DeepSeek claims to be more energy efficient than other AI veterans for the following reasons. DeepSeek has optimized hardware by using low-power AI chips and distilled versions that can be easily downloaded on laptops, which require very little computational power; since it uses reinforcement learning, the energy cost for training is less. All these reasons collectively make it an energy-efficient alternative. Even though we are unsure of the exact numbers, DeepSeek is said to have lesser energy consumption if not optimistic.

AI democratization

Even though DeepSeek’s future is not guaranteed, it has revolutionized AI space by breaking stereotypes that only organizations with deep pockets and colossal infrastructures can sustain in AI space. An open-source model like DeepSeek has opened doors for aspiring startups, researchers, and even governments like India, who are passionate about innovation yet stumbled down by the above barriers. With its mixture of expert design, optimized algorithmic efficiency, and novel architecture, DeepSeek has proved that entering the AI realm does not always need an expensive price tag.

Is DeepSeek’s AI revolution too good to be true?

Though it is reported that the cost of operations for the DeepSeek R1 model is $5.6 million, this is the cost associated with the training cost for the final model v3; this leads to questions regarding the overall cost structure concerning DeepSeek. A few have raised concerns about privacy policy and bias as well.

Wrapping Up

DeepSeek R1 has been called AI’s Sputnik moment, and what’s interesting is that the statement came from Marc Andreessen, a prominent figure in Silicon Valley. DeepSeek is not another AI model; it is a disruptive force in shaping the future of the AI world. While there are still questions regarding the model’s sustainability, one major thing DeepSeek has accomplished is for prominent AI leaders to rethink their strategy. Whether DeepSeek is to lead the AI race or fall out just like its predecessors, one thing remains unshaken: that AI race is yet to begin.