In the race to create powerful artificial intelligence (AI) models, a long-running assumption among investors was that the United States had a massive advantage over its rival, China.
This is because it was generally thought AI innovation would require vast economies of scale. And the United States, so the thinking went – replete with vast wealth and resources – would be best positioned to achieve this. Nvidia for instance – the world’s biggest producer of AI hardware – is US based and among the three biggest public limited companies in the world. The US also has a massive “first mover” advantage in the development of new technology.
Its three biggest tech companies, for instance, have a market capitalisation greater than China’s entire stock market. And the US has invested heavily in AI in recent years. Only last week, the Trump administration announced a staggering $500 billion worth of private investment in the infrastructure required to generate the computational power required to run large AI models.
Then – last week – this assumption was shattered when DeepSeek, a Chinese AI company, released RI. The large language model (LLM) was created at a fraction of the cost of competitors such as OpenAI’s ChatGPT, Google’s Gemini and Microsoft’s Azure, while maintaining similar levels of computational power. The Hangzhou-based company was founded less than two years ago by Liang Wenfeng, a co-founder of a hedge fund which uses AI to identify patterns in stock prices.
In 2021, it is believed he began buying Nvidia chips to develop AI models as a hobby, using money from his hedge fund. His company is focused purely on research rather than commercial products and the DeepSeek assistant and code can be downloaded for free. Its models are also cheaper to operate than those of US competitors. Last week, the company released its research paper, Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, which outlines how DeepSeek achieved its breakthrough by applying reinforcement learning techniques to refine a standard language model. Thus, techniques once jealously guarded by US tech giants are now available to anyone.
At first, industry insiders in the west treated DeepSeek with caution. Perhaps, they reasoned, the company was a propaganda arm of the Chinese state, aimed at undermining American AI dominance. Or perhaps the apparent costs of training the language model (a mere $5.5 million) were a gross underestimation.
But the full implications of RI’s release soon sent shockwaves through the global economy. An astonishing $1 trillion was wiped off the US tech index in a single day. Nvidia, meanwhile, suffered the biggest single day market value loss in US history, falling by almost $600 billion. Venture capitalist Marc Andreessen described DeepSeek’s release as a “Sputnik movement” for AI, a reference to the Soviet Union’s launch of its first satellite, which kicked off the “space race” with the US President Donald Trump, meanwhile, called the launch a “wake-up call” for the US tech industry.
What are the implications? Well, the assumption that the development of sophisticated AI models will demand immense economies of scale no longer makes sense. DeepSeek proves that much smaller models, when properly trained, can compete – or even exceed – the computational power of those built by tech giants with vast amounts of capital. That will be very good for competition long-term, with more investment likely to flow into smaller start-ups and away from the behemoths. DeepSeek’s open-source model, meanwhile, could, according to the tech analyst Jen Zhu Scott, also have a democratising effect. AI, she argues, ought not to be a zero-sum game between China and the US, but used for the better of humanity.
DeepSeek’s open-source philosophy would help achieve this. And the growth of smaller, but no less powerful AI models, could also be good for the environment. In a March 2024 piece for the Atlantic, the American journalist and data scientist Karen Hao cited a study published in the Netherlands that showed how, by 2027, AI servers’ electricity demand could equal the entire annual consumption of Argentina. Training models like DeepSeek’s, by contrast, require far less energy.
That said, we may wish to treat this story with some caution. OpenAI, for instance, claimed on Wednesday that its proprietary model was used by DeepSeek to train R1, in effect piggybacking on its capital-intensive work. If that is true, the claim that training the language model cost $5.5 million will be proved inaccurate. Nevertheless, DeepSeek has challenged the notion that the US was well ahead of China in the development of powerful AI models. The release of Chinese firm Alibaba’s powerful new AI model attests to this too. This will have geopolitical implications, with the US government likely to seek new, more stringent methods to prevent chips falling into the hands of Chinese rivals. US companies will also have to rethink their business models.
Why, for instance, would anyone pay for their products when open-source alternatives are available for free? The answer lies in the development of even more powerful AI models, an accelerating global race which the arrival of DeepSeek will do nothing to abate.
