DeepSeek, the Chinese company that wants to challenge the US in the global artificial intelligence race, is sending shockwaves across the AI world with its cost-effective strategy.
The technology company, based in the eastern Chinese city of Hangzhou, was founded in May 2023, about four months after OpenAI's ChatGPT exploded into the scene. It remained relatively quiet and unknown – until last week, when it launched its DeepSeek R1 large language model.
What happened next was an unexpected and meteoric rise that has shaken up the AI community at large – and raising the question if AI could really be more affordable.
How good is DeepSeek?
According to tests, DeepSeek's R1 model matches, if not surpasses, OpenAI o1, which is the US company's best available offering.
And, to top it all, it's free, or, technically, almost free.
While ChatGPT, Google Gemini and Microsoft Copilot have varying prices and tiers, DeepSeek lets you sign up with just your email address for the right to use its V3 and R1 models, which were launched in December and last week, respectively.
According to its website, V3 “achieves a significant breakthrough in inference speed”, and “tops the leaderboard among open-source models and rivals the most advanced closed-source models globally”.
Meanwhile, R1 “demonstrates remarkable reasoning capabilities” and “naturally emerges with numerous powerful and intriguing reasoning behaviours”, says the DeepSeek website.
Sure, there are free versions of the top generative AI models, but that won't give you access to their most premium features.
And that's where DeepSeek has an advantage over others: its free-to-use model gives you access to everything they have on offer – as long as you use them personally or locally on your computer.
If you want to use DeepSeek on an API – or application programming interface, which basically is mainly used by developers and connects them to other data sources – you will have to pay for tokens.
Tokens in AI are the basic unit to calculate the length of text and can include punctuation marks and spaces, and vary from one language to another. The bigger the token count, the more comprehensive the generated results are.
Is DeepSeek really cost-efficient?
As it stands, yes. You need to pay just $0.14 and $0.55 per million tokens for V3 and R1, respectively.
How is DeepSeek able to do this? The company claims to have spent just $6 million – and only two months – to train its model. By comparison, it has been reported that Meta Platforms used $60 million for its Llama model, while OpenAI splashed out more than $6 billion, with both taking longer as well.
DeepSeek charges a comically low amount. But it isn't a laughing matter to others.
How has DeepSeek shaken up the market?
DeepSeek is trying to rival the biggest names without spending too much. The company also managed to unseat ChatGPT on top of Apple's App Store.
China has always tried to undercut the rest of the world with their more affordable products, with the most notable examples of these being smartphones and cars.
The stock market also felt the DeepSeek effect on Monday morning. Nasdaq futures dropped 2 per cent, while the S&P 500 shed 1 per cent. In Asia, technology stocks fell, wary of the effect DeepSeek may have on AI investments worldwide.
China's biggest chipmaker, Semiconductor Manufacturing International Corporation, declined 2.5 per cent, while Japan's Disco and major Nvidia partner Advantest gave up 2.9 per cent and 8.1 per cent, respectively.
Nvidia, the leader of the AI chip revolution, was also poised to open down from its Friday close.
“Futures … are looking pretty bad this morning … hammered by the news that the Chinese start-up DeepSeek could run its latest AI model on less advanced chips – and could disrupt the US tech’s global dominance,” Ipek Ozkardeskaya, a senior analyst at Swissquote Bank, said in a research note on Monday.
Still, it is too early to declare DeepSeek as the new leader in the AI race, let alone if it will actually be able to challenge the leaders.
“Keep calm and breathe … this week’s Big Tech earnings will certainly give a stronger conviction to those looking for a fresh direction,” Ms Ozkardeskaya said.
Microsoft, Meta, Tesla and Apple are due to announce their fourth-quarter earnings this week.
