Before last week, almost no-one had heard of DeepSeek. Yet this Chinese start-up has shocked the tech world — and global stock markets — with the release of R1, a new model that threatens to upend the economics of generative AI.
“DeepSeek-R1 is one of the most amazing and impressive breakthroughs I’ve ever seen — and as open source, [it is] a profound gift to the world,” Marc Andreessen, a Silicon Valley luminary and investor, posted on X last week.
What has Andreessen and others in Silicon Valley excited — and some worried — is not so much that DeepSeek appears to have caught up with US AI pioneer OpenAI in areas such as “reasoning”, but that it has done so at a fraction of the cost normally associated with training and using large language models.
As a result of its cost advantages — developed out of necessity as US sanctions crimp Chinese companies’ ability to buy advanced chips from firms such as Nvidia — DeepSeek could bring profound changes to AI. And, as Monday’s stock market rout shows, Wall Street is paying attention.
Necessity is the mother of invention and the US sanctions may, ironically, have spurred China to build technology that could disrupt the AI boom that has propelled the US stock market to record highs.
This shows that US government measures are unlikely to hold China back for long in the race to develop advanced AI tools. Heavy-handed restrictions failed to hobble Huawei for long financially, and they’re unlikely to hold China back meaningfully in AI. Whether that’s a good thing is open to debate.
Founded by billionaire Liang Wenfeng and funded by High-Flyer, a hedge fund he created, DeepSeek is making its AI models available on an open-source basis. This means anyone can freely download the software and start playing with it, provided they have a modern PC with a good graphics processing unit (GPU) and lots of memory.
US sanctions may, ironically, have spurred China to build technology that could disrupt the AI boom that has propelled the US stock market to record highs
Because it’s open source, anyone can modify the code for their own applications. This, coupled with its cost advantages, has meant that R1 has quickly become one of the most popular AI models on Hugging Face, a website where developers share AI code and collaborate on AI projects.
Proprietary alternatives from companies such as OpenAI and Google are offered for a fee: between $20 and $200 a month in the case of OpenAI.
Even if you want to use R1 for development purposes, its API costs — the fees you pay to use the company’s own computer resources — are less than 10% than those for o1, the equivalent model from OpenAI. Installing and playing around with it doesn’t require you to be a developer; only that you’re comfortable using a command-line interface.
The launch of DeepSeek-R1 came as US President Donald Trump unveiled a plan by companies, including OpenAI, SoftBank and Oracle, to invest $500bn in AI infrastructure by 2029. Because of the funding commitments involved, the plan — known as Stargate — has drawn comparisons with the Manhattan Project during World War 2 to build nuclear weapons.
Outside Stargate’s plans, big tech companies are already pouring billions of dollars into AI infrastructure. Meta Platforms alone has said it will spend up to $65bn in 2025 to build, among other infrastructure, a data centre housing at least 1.3-million GPUs.
But has DeepSeek suddenly thrown the wisdom of this eye-watering capital expenditure into doubt? It certainly raises questions about Nvidia’s valuation, which cratered nearly 18% on Monday as investors processed the significance of DeepSeek’s efficiency breakthrough.
If future AI models are going to be significantly more efficient at using GPUs, will the world need to spend nearly as much as planned on AI data centres and the Nvidia chips needed to run them? If not, there are deeply troubling implications for the AI-led stock market boom.
The release of DeepSeek-R1 shows that China is more than up to the task of challenging US leadership in AI. Just don’t ask the software about the Tiananmen Square uprising or who controls Taiwan — it won’t answer you, or will provide a censored and sanitised response. China is making stunning progress in AI; it’s just a pity about its authoritarian and censorial government.
* McLeod is editor of TechCentral
Would you like to comment on this article?
Sign up (it's quick and free) or sign in now.
Please read our Comment Policy before commenting.