- Premium Content It takes our newsroom weeks - if not months - to investigate and produce stories for our premium content. You can’t find them anywhere else.
Scott Shuey · · 6 min readChips don’t lie: Did China’s latest AI just beat the US?
“It is essential … that the world’s AI runs on American rails.”
That was from a White House statement released on January 13, 2025. It came three weeks after The Wall Street Journal published a story that said China was catching up to the top US artificial intelligence models faster than anyone had predicted.
Source: Shutterstock
The focus of the WSJ article was DeepSeek, which had launched a preview version of its AI model in November 2024. The Chinese startup dropped an app-based version of the model on Apple’s App Store and Google Play on January 10 – three days before the White House statement.
By January 27, DeepSeek had surpassed ChatGPT as the most downloaded free app on the Apple’s App Store in the US.
Is this a sign that America’s AI leadership is about to go off the rails?
US equity markets thought that might be the case on Monday, the day DeepSeek topped the App Store charts. The Nasdaq tumbled 3%, and Nvidia – the golden child of American AI – lost nearly US$588 billion or almost 17% of its market value.
It’s unclear what this means for the AI industry. But on the surface, the DeepSeek V-3 R2 large language model (LLM) seems to perform just as well as US-based OpenAI’s o1 model at a much lower cost.
Of course, there are questions about this feat.
DeepSeek, which was established by CEO Liang Wenfeng in 2023, is claiming its current LLM was trained without high-end chips and at a cost of only US$5.6 million. The company did this by using between 10,000 and 50,000 Nvidia A100 chips – the industry standard in 2023 – before the US imposed AI chip restrictions on China.
GPT-4 is estimated to have cost between US$41 million and US$78 million to train.
So how did DeepSeek keep the costs down? There are a couple of possible answers, but none of them are really satisfactory.
Blame the loopholes
The first is to rush to the same conclusion that Tesla CEO Elon Musk did: DeepSeek is lying.
It is quite possible that DeepSeek had accessed an AI cloud server somewhere outside China and trained its model on Nvidia’s H100 chips, which are generally considered the industry’s best.
Fresh fabs for faster FLOPS
But what if it’s real?
Stay ahead in Asia’s tech landscape
This is premium content. Subscribe to read the full story.
Its AI model development costs are over 7x lower than OpenAI’s ChatGPT-4, DeepSeek says. How did this Chinese AI startup do it?
We know this is not ideal. ⌛ Sign up in 20 seconds. Cancel anytime.
Our subscriber community includes professionals from these companies:
Stay updated on the go with our mobile app.
Get latest insights with smoother, more personalized experience through TIA mobile app.
