Is DeepSeek finally bursting the AI bubble?
“Anything is possible on the stock market. Even the opposite.” Hungarian investor André Kostolany’s timeless observation feels especially true and relevant this week.
About three years ago, launched ChatGPT, a game-changing product that shook the tech world and paved the way for the emergence of a trillion-dollar artificial intelligence (AI) economy. This sparked a rush among tech giants and ambitious startups, all eager to capitalise on the AI boom, launching bold new solutions.
But this week, a Chinese AI startup disrupted the game in a major way.
Shares of Nvidia—a key player in the AI hardware market—took a massive hit, wiping out an estimated $592.7 billion in paper value on Monday.
The cause? DeepSeek, a Chinese startup that’s shaking up the AI landscape. Its new AI models claim to be on par (or exceed) compared to some of the top AI models in the US.
These models are being built at a lesser cost. Through ingenious engineering, DeepSeek has drastically reduced its computing power needs, sparking concerns about the future demand for Nvidia’s high-end chips in the West.
The company claims to have spent just $5.6 million to train its base AI model—a fraction compared to hundreds of millions, if not billions, poured into projects like Stargate, a venture specifically launched for AI infrastructure and backed by giants like OpenAI, Oracle, and SoftBank.
But what is DeepSeeek and how has it disrupted the AI space like never before?
What is DeepSeek?
DeepSeek was founded in late 2023 by Chinese hedge fund manager and entrepreneur Liang Wenfeng. Wenfeng, who also runs the AI-focused hedge fund High-Flyer, established the company in Hangzhou, China.
According to data from Sensor Tower, DeepSeek launched its services in the US in January, and its flagship AI assistant, built on top of the DeepSeek-V3 model, has already outperformed ChatGPT to become the top-rated free app on Apple’s App Store in the country.
To put things in perspective, AI models are dependent on advanced infrastructure for their performance. This includes high-performance computers, GPUs (Graphic Processing Units), and specialised AI chips to handle tasks such as deep learning and reasoning.
However, DeepSeek, with its new innovation, has eradicated the need for heavy computational power.
Tech investor Marc Andreessen has described this as “AI’s Sputnik moment.” This is mainly due to two underlying reasons—the cost-effectiveness of DeepSeek’s AI models and their ability to run efficiently by opting lower-cost hardware.
“The Sputnik description is a perfect analogy. While Russia was first to space, the West dominated after that, and through the nuclear arms race with Russia and the Strategic Defense Initiative under Reagan, the Soviet Union was bankrupted. While China will be near impossible to bankrupt, the West can collectively out-innovate them and hope that companies who wish to maintain their security will be hesitant to use an AI Chatbot from China – which will no doubt steal their data and be vulnerable to an attack as we are already seeing,” Dan Schiappa, Chief Product and Services Officer at Arctic Wolf tells YourStory.
“There is no doubt this will start an AI arms race. President Trump already highlighted the $500 billion investment coming to America with a partnership of SoftBank, Oracle and OpenAI, and many others will jump on board. In short, this will force an innovation boom,” he adds.
DeepSeek’s engineers, in their research paper, have revealed that they used approximately 2,000 Nvidia H800 chips—less advanced than the most AI chips—to train the model.
The team employed a collaborative approach with multiple specialised models working in tandem, allowing the less powerful chips to process data more efficiently.
The paper further suggests two possibilities—high-performing AI models may not require the most advanced chips, or Chinese companies can still acquire enough chips to meet their needs—or a combination of both factors.
DeepSeek holds another major advantage—the models are open-sourced. This means the code or technology is freely available for anyone to use, modify, and build upon. This makes it much easier for businesses to build their own solutions on top of them.
“Built by a Chinese hedge fund with just $5.6 million and older Nvidia H800 chips, it has delivered capabilities rivaling and, in some cases, surpassing Silicon Valley giants like OpenAI and Google. By challenging the traditional norms of exorbitant costs, cutting-edge hardware, and proprietary development, DeepSeek has redefined innovation with its open-source release and unprecedented accessibility,” says Jaspreet Bindra, Co-founder and CEO, AI&Beyond.
New breakthrough in reasoning
Reasoning is a fundamental aspect of human intelligence. It enables us with problem-solving, decision-making, and understanding complex systems.
Early AI models focused primarily on generating responses, often struggling with accuracy. Last year, OpenAI introduced the world’s first reasoning model, o3. Designed to tackle complex questions in science and mathematics, o3 employs a structured approach by breaking problems into smaller steps and testing multiple solutions behind the scenes before delivering a well-reasoned conclusion to the user.
To address the growing competition in reasoning models, Google launched “Gemini Flash Thinking” in December, while Chinese ecommerce giant Alibaba introduced an updated version of its Qwen chatbot (QwQ) built with similar reasoning capabilities.
However, DeepSeek stood out with its approach. It rolled out a preview of its reasoning model, R1, and v3, an advanced LLM with a 700GB size and 685 billion parameters—outpacing any model previously available for free download.
In comparison, Meta’s flagship LLM, Llama 3.1, released in July, features only 405 billion parameters.
Parameters are like the brain cells of the system. They help the model to learn and make decisions. These are numbers that the model adjusts during training to understand patterns, process information, and generate accurate responses.
DeepSeek’s v3 has 685 billion parameters, meaning it has more “brain power” to handle complex tasks compared to Meta’s Llama 3.1, which has 405 billion parameters. The more the parameters, the more the model can understand and generate more detailed and accurate responses.
Adding to its competitive edge, DeepSeek offers R1 at a significantly lower cost, charging just $0.55 per million input tokens, compared to OpenAI’s 01, which costs $15 per million input tokens. Input tokens are the small pieces of text that AI models read and process – it can be a word, part of a word, or even punctuation.
Disrupting the tech order
DeepSeek’s rapid rise has positioned it as a formidable competitor to established giants in the AI sector. DeepSeek’s R1 model now competes with the capabilities of competitors such as OpenAI’s GPT-4, Meta’s Llama, and Google’s Gemini.
The Chinese startup claims that DeepSeek-V3 and DeepSeek-R1 are both on par with OpenAI and Meta’s most advanced model.
“This model’s success challenges the dominance of Big Tech, breaks the US-China hegemony narrative in AI, and empowers developers globally with low-cost, high-performing tools. For India and other emerging markets, this represents a unique opportunity to build context-specific AI solutions, leveraging this democratized frontier to solve local problems,” Bindra adds.
The United States has long cited national security concerns as a reason for restricting the supply of high-power AI chips to China. The restrictions have raised doubts about the viability of some tech giants’ massive AI investments, with shares of several big tech players, including Nvidia, being hit.
For now, the Chinese AI has gained the attention of OpenAI boss Sam Altman, who called the model “impressive”.
“DeepSeek’s r1 is an impressive model, particularly around what they’re able to deliver for the price. We will obviously deliver much better models, and also it’s legit invigorating to have a new competitor! we will pull up some releases,” said Altman on X.
This again comes down to the launch of ChatGPT in late 2022, which triggered a race among Chinese tech companies to rapidly develop their own AI-powered chatbots. Despite ongoing efforts by the US government to restrain the growth of China’s AI industry, DeepSeek has altered the narrative of AI powerplay for now.
“But mostly, we are excited to continue to execute on our research roadmap and believe more compute is more important now than ever before to succeed at our mission,” Altman added.
