Is DeepSeek a genuine disruptor?

Many businesses complain that they lack resources, but the smartest innovators tend to see scarcity as an opportunity, asking: how can we make more efficient use of what we have at our disposal? The high-pressure steam engine was invented by the engineer Richard Trevithick in Cornwall in the early 19th century, spurred by the high cost of imported coal, of which the region had no reserves. This invention led to the development of the locomotive and the railway network, transforming economic development across the world.

Are we witnessing a similar development in AI? When the reported efficiency gains became public of two large language models developed by the Chinese startup DeepSeek – the R1-zero, and the R1, with performance comparable to OpenAI ChatGPT – this prompted a colossal loss of stock market value of US tech companies, a huge $1 trillion in a single day, Monday 27 January. The value of 英伟达 alone fell by around $600 billion on the day. The implications were that demand for US large language models (LLMs), and some of their suppliers, as well as energy companies, would fall substantially.

Engineers at DeepSeek encountered a relative scarcity of supply following export ban by the US of the highest-performing chips made by Nvidia. So they sought ways to develop intelligent reasoning within their models using lower computing power.

DeepSeek is a relatively young company, founded by Liang Wenfeng, and funded by the hedge fund High Flyer Capital Management. He hired some of the smartest IT engineers from top Chinese universities, and rewarded them well.

It reportedly spent just $5.6 million based on 2.8 million hours of Nvidia H800 GPU (graphics processing unit) training the LLM. This compares with around 31 million hours on Nvidia H100 GPU to train Meta’s Llama 3 405B. In 2023 Nvidia modified its most advanced chip, the H100, to enable it to be legally exported to China as the H800.

Some of DeepSeek’s claims have been questioned. It is suspected that it had amassed a stockpile of higher-grade Nvidia chips before the export restrictions, and western experts are skeptical that the training costs really were as low as reported. Separately, OpenAI, which developed the rival ChatGPT, alleges possible theft of its intellectual property by DeepSeek.

What is acknowledged by western experts is the ingenious design. DeepSeek inventors have developed a ‘Mixture-of-Experts’ configuration, in which the relevant specialist sub-network is deployed only when needed. This explains the substantially lower computing power needed. DeepSeek have used open source computing, so some of their claims can be checked.

Accordingly, while some in the US dispute the scale of the efficiency gains, it may still be a disruptive innovation. The economist Nouriel Roubini, in his note on the subject, claimed that the efficiency gain was around two to three times, not the 20 times initially reported. Nonetheless, this is still enough to transform the economics of deploying AI and increasing demand.

The entry into the market of DeepSeek R1 and R1-zero challenges the assumptions, held firmly until just a week or so ago, that progress in AI capability depends upon huge scaling up of computing power, and that the US had an unassailable lead in the technology. It poses a challenge for US policy-makers: the intention of curbing exports of the most powerful chips was to curb Chinese development of AI – but it has delivered the opposite.

US tech giants have wasted no time in responding. Meta announced it has set up four ‘war rooms’ to investigate the technical achievements of DeepSeek, with a view to competing on cost and efficiency in its Llama technology.

The efficiency gains will not only be for Chinese providers and users of AI. Accordingly, the selling of US tech shares was almost certainly overdone, especially in the case of Nvidia. The availability of cheaper and more efficient AI models looks set to increase aggregate demand for LLMs generally.

It is probably only a matter of time before US competitors develop their own smarter, more efficient AI models. And while the focus is on China-US rivalry, there may be a surprise disruptor from somewhere else in the world. This is the beauty of human inventiveness: it never stops, and it often surprises.