With the pace of news regarding Artificial Intelligence, it seems you blink, and you’ve already missed a week’s worth of developments. One such development was an underdog story of a Chinese Business that killed the American monopoly on AI. It was an arrogant monopoly, to be accurate.
Sam Altman, CEO of OpenAI, during an Economic Times Conversations event in June 2023, responding to a question about the feasibility of creating a substantial AI model with limited resources, claimed, “It’s totally hopeless to compete with us on training foundation models.”
The notion of training a large language model at a low cost was touted as practically impossible. Given the confidence, it must have been a shock for him when the release of a new AI model on the 27th of January sent seismic shockwaves throughout the industry.
NVIDIA, the golden boy, destined to gallop with tech revolutionaries to an AI-dominated future; the $2 trillion chipmaker was practically synonymous with the AI boom. Yet the news made shares of Nvidia plummet by 17% in a single day, cutting nearly $600 billion from its market cap and leading to the largest single-day loss in US history.
The culprit: a Chinese AI startup named DeepSeek, with its flagship model, DeepSeek-R1, rivalled the performance of OpenAI’s ChatGPT and its reasoning engine but cost just 2% as much to run. Here’s how a little-known startup from Hangzhou disrupted the status quo and what it means for the future of technology.
Founded in May 2023 by former hedge fund manager Liang Wenfeng, DeepSeek emerged from High-Flyer Quant, a firm known for algorithmic trading. Unlike US rivals reliant on venture capital, DeepSeek was bankrolled entirely by High-Flyer, freeing it from shareholder pressures. Liang’s vision was audacious: democratize AI by slashing costs and prioritizing open-source models.
In a smart move, before the Biden administration started limiting exports of AI chips to China in 2021, Liang began buying thousands of NVIDIA graphics processors with the goal of stockpiling 10,000.
The startup’s DeepSeek-V3 model, released in late 2024, introduced Multi-Head Latent Attention (MHLA) and a Mixture-of-Experts (MoE) architecture, activating only 37 billion of its 671 billion parameters per task. This design drastically reduced computational demands while maintaining performance. By January 2025, its successor, R1, outperformed OpenAI’s o1 in math benchmarks (91.6% vs. 89.3% on MATH) and offered a 128,000-token context window.
Unlike proprietary models (e.g., OpenAI, Anthropic), DeepSeek’s AI open-source framework enables global collaboration, empowering startups and academia to build on its technology. This has spawned over 700 derivatives and integration with platforms like Microsoft Azure and AWS. Techniques like MHLA reduce memory usage to 5–13% of traditional methods, while FP8 mixed-precision computation and PTX programming optimize GPU utilization.
But the real bombshell was cost. DeepSeek claimed it trained R1 for just $6 million using 2,000 NVIDIA H100 chips, compared to OpenAI’s billions. While skeptics argue this figure excludes R&D and hardware investments, the message was clear: AI didn’t need the crazy amount of capital that was flooded into it.
With a rigid dedication to algorithmic efficiency, the startup sidestepped the energy-intensive demands of rivals. For example, its models use reinforcement learning (RL) pipelines that learn autonomously through feedback loops, reducing reliance on labelled data. Meanwhile, techniques like Multi-Token Prediction—predicting several tokens during training—boosted performance without requiring more GPUs.
The startup’s open-source ethos further rattled Washington. By releasing model weights publicly, DeepSeek invited global collaboration—a stark contrast to OpenAI’s guarded approach. Despite US-China tensions, Microsoft’s rapid integration of R1 into Azure AI Foundry underscores the model’s allure.
DeepSeek’s application, which briefly topped US app store download charts, faces bipartisan scrutiny over data privacy. A leaked database exposing user chats and API keys in January 2025 amplified fears of Chinese surveillance. Besides privacy, the app is also censored according to Chinese likings, with sensitive questions regarding Taiwan or Uyghurs often being declined or ignored.
A cold war had always been brewing between China and the West in earlier years. The West would lead in education and research with their workers, primarily aimed at advanced manufacturing and engineering, while China leveraged its manpower and focused on simple manufacturing and resource extraction. The dynamic is now shifting with the success of Chinese education and academia fostering innovation. China is able to compete with advanced manufacturing while keeping costs low. Its developments in electric vehicle production are another recent success. It is a testament to Chinese innovation and aptitude in science and technology.
Even with the technological limitations due to the trade restrictions for microchips, the Chinese solution was creative and cost-effective. The decision to make it open source is also commendable. While OpenAI had initially kept an open-source model, it abandoned it to secure funding.
Given their rather antagonistic relations, China and America will end up battling over the monopoly. Similar to the space race during the Cold War, China launched its “Sputnik,” and now, can America respond by putting a man on the moon?
If you want to submit your articles and/or research papers, please check the Submissions page.
The views and opinions expressed in this article/paper are the author’s own and do not necessarily reflect the editorial position of Paradigm Shift.
The author is studying Economics at the National University of Science and Technology (NUST) with a keen interest in financial affairs, international relations, and geo-politics.
