At the end of 2024, the Chinese AI upstart, DeepSeek, shook the industry with the release of its new reasoning model, R1. Claiming to rival the performance of OpenAI and Google at a fraction of the cost and computing power, this Hangzhou-based startup is worrying its Western counterparts.
Key innovations
Affordable, efficient, and fast, DeepSeek reportedly trained its latest R1 model using the DeepSeek-V3 base model. The V3 model cost just $5.6M – compared to the estimated $100 million cost of training OpenAI’s GPT-4 – and was trained using older technology in the form of 2,048 Nvidia H800 GPUs.
This innovative approach to training its reasoning model challenges the prevailing belief that AI requires massive budgets, numerous AI accelerators, and cutting-edge resources. It also undermines the traditional scaling law, which assumes AI models require ever-increasing compute and data resources.
Alongside signalling a shift towards cost-efficiency and leaner development models in the AI realm, DeepSeek is open-sourced, democratising access to advanced AI capabilities.
The impact on the market
Following R1’s release, there’s been a significant industry reaction to this less expensive and more sustainable app, resulting in a $1 trillion sell-off in global tech stocks.
According to Bloomberg, Nvidia stock fell by 17%, while DeepSeek’s new AI assistant app reached the top spot on Apple’s ‘Top Free Apps’ chart with more than 1.6 million downloads for the US, UK, Australia, Canada, Singapore and China.
This has led to broader scepticism about the high-cost AI strategies of Silicon Valley giants, with Nvidia’s market value dropping by nearly $600bn – a record loss. Some, however, like Nvidia, see it as technological breakthrough and an opportunity for increased demand for their chips.
Nvidia’s CEO, Jensen Huang, stated that DeepSeek’s R1 model was “an excellent AI advancement,” and highlighted how reasoning models will drive demand for their inference chips and GPUs – a rapidly evolving AI market where US-based Nvidia is a leading player.
Responses from industry leaders
In an attempt to understand this obscure Chinese-made AI app achieved this outstanding success, industry leaders, including OpenAI, Meta, and Google, are analysing DeepSeek’s methodology and reevaluating their strategies and investments.
As reported by Silicon Valley Capital Partners (SVCP), the CEO of OpenAI, Sam Altman told employees “the release marks a major landscape shift for the startup.” Concerns regarding AI security, however, remain at the forefront of industry leader minds with Gemini and ChatGPT advising caution.
According to Forbes, these AI behemoths cite potential security vulnerabilities and privacy risks, the current geopolitical climate, and its origins in China as reasons for DeepSeek being a less safe choice to download onto a personal device.
Regulatory, geopolitical, and future implications
Despite the tightening of US restrictions on advanced chip exports to China, as highlighted by the BBC, DeepSeek’s success showcases how resourceful methods can bypass such limitations.
In response, US authorities are reportedly investigating whether the Chinese AI upstart obtained and utilised restricted US-made Nvidia AI chips in the development of its models, as revealed by Tech Monitor.
While travelling in Florida, US President Donald Trump also commented on the accelerated rise of DeepSeek and how the US tech industry must dedicate their efforts to competing with Chinese companies.
“The release of DeepSeek AI from a Chinese company should be a wake-up call for our industries that we need to be laser focused on competing,” he said, as reported by NBC News.
Future implications
Crucially, while US technology firms currently advise caution with utilising the DeepSeek app, the success of this Hangzhou-based upstart demonstrates the exciting potential for the broader adoption of AI thanks to reduced costs.
It also emphasises and encourages the use of innovation from smaller players globally and reinforces the importance of efficiency over brute force in AI development.