/socialsamosa/media/media_files/2025/01/28/bxSdweIcsiFeyrPpOgEW.jpg)
When the coal-powered industrial revolution struck, engineers marvelled at steam engines that promised efficiency, only to discover a counterintuitive reality — as coal use became more efficient, demand for coal skyrocketed. This puzzling outcome, first observed by economist William Stanley Jevons in the 19th century, would later be called ‘Jevons Paradox.’ Now, fast forward to the 21st century, where the principle can be applied not with coal but with artificial intelligence. Enter DeepSeek, a Chinese AI model that is suddenly the talk of the town, promising to do more with less — less computing power, fewer chips and lower costs.
The relatively unknown AI research lab from China released the open-source model on January 20. According to its research paper, DeepSeek-R1 supposedly beats models like OpenAI o1 on several math and reasoning benchmarks. The startup’s engineers have explained that they used only a fraction of the specialised computer chips that other A.I. companies rely on to train their systems.
In simple terms, this model is giving Western AI giants a hard time. Its app reached the top spot in app stores around the globe over the weekend, preceding OpenAI’s ChatGPT. According to analytics firm Sensor Tower, the model has seen more than 80% of its total mobile app downloads in the past seven days, wherein it saw nearly 300% more app downloads than Perplexity.
Worse, the startup is disrupting financial markets, with chip maker Nvidia seeing its shares dip nearly $600 billion. Its stock closed down 17% on January 27, marking its worst single-day percentage loss since March 2020. This model has left the tech industry baffled as tech giants like Meta have reportedly created ‘war rooms’ to understand DeepSeek’s training methods.
While it is said that the model performs better than ChatGPT, DeepSeek’s newly-released Janus Pro multimodal AI model can apparently also outperform Stable Diffusion and DALL-E 3.
Its efficiency isn’t just a technological triumph, it could be a catalyst for larger disruptions across the AI ecosystem, from the hardware industry to global geopolitics, given that the US currently dominates the AI market, and this could cause further tensions between the US and China.
When was DeepSeek launched?
The startup was launched in 2023 by Liang Wenfeng, co-founder of quantitative hedge fund High-Flyer. Liang's fund announced in March 2023 that it would concentrate resources on creating a ‘new and independent research group, to explore the essence of Artificial General Intelligence (AGI).’ High-Flyer reportedly owns patents related to chip clusters used to train AI models of around 10,000 A100 chips.
DeepSeek was officially launched after years of development by a coalition of Chinese tech firms operating under the umbrella of Beijing’s AI-forward policies. China’s government-backed AI initiatives have long sought to establish dominance in this space.
At its core, it is a generative AI model designed for versatility. From natural language processing to complex data analysis, it claims to offer robust multilingual capabilities, optimised for Asian languages and a focus on integration with Chinese apps and platforms.
Why is there hype around DeepSeek?
The buzz around DeepSeek stems from three main factors – efficiency, accessibility and timing. The claim that it requires fewer chips to operate has sparked conversations about cost-effectiveness and scalability. The startup claims it costs less than $6 million to develop as compared to OpenAI’s GPT-4 which reportedly costs over $100 million to train.
In a world where AI is both expensive and resource-intensive, such a model could democratise AI adoption in emerging markets. It has also launched at a time when global tech giants are being criticised over data privacy, bias and monopoly.
Despite its market losses, Nvidia believes that DeepSeek is an “excellent AI advancement and a perfect example of Test Time Scaling”. The chip maker notes that new models can be created using that technique, leveraging widely available models and computing that is fully export control compliant.
It also pointed out that despite DeepSeek’s cost-efficiency, it still relies heavily on Nvidia GPUs to produce outputs after they are trained.
President Donald Trump highlights DeepSeek’s rise as a “wake-up call” for U.S. technology firms. Trump has emphasised the need for U.S. industries to focus on staying competitive in the global AI race and has discussed imposing higher tariffs on foreign computer chips to incentivise domestic manufacturing. His proposed tariff-based strategy wants to ensure that US manufacturers would shift production to avoid such taxes. However, it is unclear whether these measures would be feasible.
Trump’s push for US-based manufacturing appears to have a common consensus amongst tech giants. OpenAI recently urged the U.S. government to support the development of U.S. AI, to avoid Chinese models surpassing them in capability. In an interview with The Information, OpenAI’s VP of policy Chris Lehane also expressed concerns about High Flyer, DeepSeek’s parent company.
However, OpenAI’s CEO Sam Altman has also acknowledged its progress, calling its R1 model impressive for its cost-effectiveness. Altman noted that U.S. companies would deliver “even better models” while appreciating the healthy competition. Meanwhile, Trump has pointed out that DeepSeek’s innovations could lead to reduced costs in AI development, contrasting with the massive spending planned by companies like OpenAI and SoftBank, which are gearing up to invest hundreds of billions in AI data centres.
Does it really need fewer chips to compute?
While DeepSeek has claimed to require fewer chips and could reduce operational costs significantly, making AI accessible to more users, there is scepticism that these claims might be overstated or come with trade-offs in performance during complex tasks. Regardless, even the perception of efficiency poses a challenge to companies like Nvidia, whose business model thrives on high chip demand. AI chips are costly, and Nvidia has stated that its AI chip, Blackwell B200, might cost anywhere between $30,000 and $40,000 per unit.
The company’s growth has been fueled by the increasing demand for high-performance GPUs, which are essential for training and running large AI models by OpenAI and Google. If DeepSeek’s approach proves viable on a larger scale, it could reduce dependency on Nvidia’s hardware.
Additionally, Nvidia's business thrives on the assumption that advanced AI requires massive computational power and large volumes of GPUs. DeepSeek’s rise might lead to a future where AI development might not need massive investments in hardware.
If DeepSeek’s efficiency proves scalable, it might drive demand for new categories of low-power chips, pressuring manufacturers to innovate. But this could take years with heavy investments.
As of now, the AI model still trails behind in certain aspects, given that it abides by Chinese laws and refuses to answer certain geopolitical questions.
Much like Jevons Paradox, DeepSeek’s claim of efficiency may lead to greater AI consumption while leading to either global progress or a trigger for new conflicts with governments and companies.