A Chinese-based AI model has taken the tech world by storm, seemingly without build up or warning, as DeepSeek climbs to the top of the Apple App Store in the UK, US, and China.

The AI app overtook LLM starlight ChatGPT in popularity, offering the same, or sometimes better, performance as OpenAI’s o1 model on a range of benchmarks, with a price tag at just a fraction of competitors.

DeepSeek claims that its AI model, R1, which was released in mid-January, costs only $5.6 million to train, a drop in the water compared to the hundreds of millions used to train rival models like OpenAI’s ChatGPT and Meta’s LlaMA.

Potentially key to the company’s success is what competitors hoped would keep it back – a lack of access to advanced chips, the sale of which was prohibited to Chinese companies by US sanctions.

DeepSeek founder Liang Wenfeng had a stockpile of Nvida A100 chips, which are included in the sanctions, which he used to launch his company.

He then paired these with cheaper, less advanced computer chips that were available in China to create the cheaper AI offering. In order to work on the available GPUs, which operate at half the speed of the top products banned in China due to sanctions, DeepSeek had to reconfigure the training process of its R1 model.

Further, DeepSeek worked on speeding up calculations and decreasing memory usage without compromising the accuracy of responses. All this, on less powerful computer chips, significantly decreased DeepSeek’s overall spend on developing and using its resulting AI model.

The sanctions, which were intended to keep American companies at the top of the AI game, shifted the strategy of Chinese-based companies, encouraging collaboration and resource pooling, the MIT Technology Review suggested.

Despite its success, however, Wenfeng says that the American sanctions are creating a bottleneck of supplies.

Recommended reading

• Gartner: AI Cyber-Attacks New Top Risk for Enterprises
• Is the UK Lagging Behind in Global AI Race?
• Gartner: Global AI Chips Revenue to Total £56BN in 2024

Still, creating an AI chatbot that can compete with – and even beat – ChatGPT needing less computational power and less upfront training costs can lead to an AI revolution.

While American companies aim to lobby for further data centres and chip protections to protect their AI dominance, the underdog company could lead to cheaper AI for all, and even disrupt the current AI supply chain.

Making a low-cost AI model can make investors question the exorbitant amount of capital invested in rival AI companies promising a return on investment.

Companies heavily entrenched in the AI race in the US saw their shared drop on Monday, including Nvidia, Microsoft, and Meta.

Tech analysts and leaders have a variety of reactions to DeepSeek’s sudden surge to the top. While many have voiced concerns on how this can affect US competitors, Meta Chief AI Scientist Yann LeCun said that, since the DeepSeek’s model is opensource, “everyone can profit from it.”

Further, showing that it is possible to make a more efficient and cheaper AI model will benefit the AI industry as a whole, Garry Tan, CEO of Y Combinator argued on X: “If training models get cheaper faster and easier, the demand for inference (actual real world use of AI) will grow and accelerate even faster, which assures the supply of compute will be used.”

While it may be good for AI advancement overall, it is understandable that massive corporations are concerned on how DeepSeek’s success – both technologically and in the market place – could affect their economic prosperity and plans.

Afterall, it was just last week that US tech giants and foreign investors banded together to announce an AI infrastructure plan – The Stargate Project – with $500bn in investment.

If a successful AI model can trump US AI companies currently dominating the scene, it draws into question their strategic priorities when it comes to business and AI, especially amid rising environmental concerns around energy and resource demands.