DeepSeek’s rise signals end of AI monopoly

In a remarkable turn of events, DeepSeek, a Chinese-developed artificial intelligence application, has rapidly ascended to the top of global tech discussions, shaking up the AI industry and posing unprecedented challenges to established giants.

Advertisement

Its popularity has directly impacted the stock prices of major US tech companies, including Microsoft, Meta, Nvidia and Alphabet. Collectively, these companies have experienced a loss exceeding $1 trillion in market value. Nvidia saw its shares drop by 17.5 per cent, resulting in a $600-billion decline in market capitalisation.

The President of the US, Donald Trump, described DeepSeek's advancements as a "wake-up call", underscoring the need for the US to be competitive in AI development. Industry leaders, including Microsoft CEO Satya Nadella and OpenAI CEO Sam Altman, acknowledged DeepSeek's achievements, with Altman calling the model "super impressive."

So, what is this DeepSeek? DeepSeek is an advanced artificial intelligence project developed in China. It is based on the large language model (LLM) that leverages deep learning and transformer-based architecture to generate human-like text, perform reasoning tasks and enhance AI-driven applications. It has challenged the dominance of western AI models, such as OpenAI's GPT-4 and Google's Gemini.

DeepSeek was founded by Liang Wenfeng in May 2023 as a spin-off from High-Flyer AI, a hedge fund managing agent which is designed to leverage AI algorithms for trading financial instruments. The company remained largely unnoticed until it released a paper introducing an innovative load balancer for connecting elements within its mixture of experts (MoE) foundation model.

Later, during the holiday season, DeepSeek-AI unveiled the architectural details of its DeepSeek-V3 foundation model. It features 671 billion parameters — of which only 37 billion are active per generated token — and was trained on a dataset of 14.8 trillion tokens. On January 21, 2025, DeepSeek launched DeepSeek-R1, featuring two additional reinforcement learning stages and two supervised fine-tuning stages to improve reasoning.

Large Language Models are massive transformer-based neural networks designed for solving the next-word prediction problem. The breakthrough for these models came in 2017, when the introduction of transformer architecture revolutionised text generation. These models are typically enormous, requiring substantial computational resources — often necessitating hundreds of thousands of graphics processing units (GPUs) — to undergo multiple training cycles, using vast amounts of text data sourced from across the internet. Through this iterative learning process, LLMs refine their ability to predict and generate text with increasing accuracy.

As AI research progressed, a race emerged among leading tech companies to develop ever-larger models, training them on increasingly extensive datasets to achieve superior performance. This race fostered the belief that AI supremacy depended on access to powerful computational infrastructure and larger models. Consequently, a market monopoly began to form, where only the wealthiest organisations with access to cutting-edge hardware could dominate AI advancements with their larger proprietary models.

However, DeepSeek has disrupted this narrative with its innovative model. Unlike conventional approaches that prioritise sheer computational scale, DeepSeek-R1 leverages optimised methodologies that enable high performance without the need for excessive hardware resources. For example, DeepSeek-V3 is trained on a cluster of 2048 Nvidia H800 GPUs, which achieves nearly the same or better accuracy as that of the traditional one.

This presents a formidable challenge to the existing AI monopoly, demonstrating that groundbreaking AI development is possible outside the constraints of infrastructure supremacy. Unlike traditional AI infrastructures that require immense power consumption, DeepSeek's model significantly reduces energy demands and carbon emissions. This environmentally conscious optimisation presents a sustainable alternative, reinforcing DeepSeek's role in advancing AI while minimising its ecological footprint.

On the other hand, by making everything open source and placing it in the public domain, DeepSeek challenges the notion that Chinese companies engage in data theft. This transparency compels users to reconsider their stance and explore the accessibility and potential of DeepSeek.

By January 27, DeepSeek had unseated OpenAI's ChatGPT as the most downloaded free app on the iOS App Store in the US. Its rapid rise has caught global attention, marking a pivotal moment in the evolution of AI technologies.

DeepSeek has shattered the myth that achieving AI proficiency requires absolute computational supremacy. The constraints imposed by limited access to high-profile GPUs due to international restrictions have, instead, driven Chinese engineers to innovate and develop creative solutions.

This again proves that necessity is the mother of invention. Indeed, the achievement of such capabilities is a technological breakthrough. And given that there is ban on the export of high-end Nvidia graphics processors to China, this is significant.

While it may seem paradoxical for a non-democratic country to lead efforts in democratising AI, DeepSeek is making significant strides in reshaping the AI landscape. Its approach not only enhances accessibility but also contributes to AI supremacy by optimising computational efficiency. Since its launch, DeepSeek has taken the digital world by storm. The sudden market disruption underscores the potential of Chinese tech innovations to challenge western dominance in AI.

Investors have expressed concerns about the competitive edge DeepSeek offers and the implications for other AI developers. OpenAI charges $2.50 per million input tokens for its GPT-4 model while DeepSeek offers a much lower rate of $0.14 per million tokens when the AI can use cached information. While costing only around $5-6 million — as against the hundreds of millions spent by major western AI labs — it's no surprise that industry users are beginning to question whether AI development is overpriced.

Could the next major breakthrough come from smaller teams with fresh ideas on training these systems?