DeepSeek May Be The Best Chinese New Year Gift To China This Year, But Can It Sink Its US Rivals?

Over the past week, the global artificial intelligence (AI) industry has been fixated on one thing and one thing only: The emergence of, until now, a little-known Chinese start-up called DeepSeek.

- Tan Dawn Wei

BEIJING -

This privately held AI newbie, founded only in 2023, appears to have come out of nowhere and done the unthinkable with its open-source large language models.

When it dropped its V3 general-purpose model the day after Christmas in 2024, the AI industry was dumbfounded that DeepSeek could rival the performance of chatbots built by AI giants like OpenAI and Google, which its researchers said it did with far less money and computing resources.

Last week, its reasoning model, the R1, stunned the world by beating OpenAI's ol model on certain benchmarks such as advanced mathematics and programming tasks.

Tech nerds everywhere who put it to the test came away impressed, triggering a massive sell-off of US chipmaker Nvidia's shares that wiped nearly US$600 billion (S$812 billion) off its market value, the biggest one-day loss in US stock market history.

A narrative of triumph over adversity has quickly seized this instant success story, a nose-thumbing moment for the Chinese who are forced to innovate in the face of crippling US tech sanctions.

It is a narrative shared by Chinese tech giant Huawei, which rattled the Biden administration when it unveiled its home-made advanced chip in its Mate 60 smartphone in 2023.

DeepSeek says it used only about 2,000 Nvidia H800 chips to train its chatbot, while other leading AI companies need almost 10 times more than that.