What does DeepSeek mean for AI?

Yesterday was an extraordinary day in the world of tech and AI. 

Leading chipmaker and darling of AI, Nvidia, went from the most valuable company in the world to third in a day, posting the a $600bn loss in its market capitalisation, the biggest market loss in history, with $1tn projected to be lost across US tech stocks.

This followed reports that DeepSeek, the Chinese AI company, had developed a ChatGPT rival at a tiny fraction of the cost of its US peers, although these claims have been disputed. Alongside winning on price, reports state that this model requires far less compute power, dashing the market-leading need for Nvidia’s chips. 

Its latest large language AI model, R1, has achieved comparable performance to its US rivals OpenAI and Meta. This is a free-to-use, open-source model which is currently topping the app store.

Whilst AI models capable of reasoning already exist, DeepSeek’s model appears to have delivered innovation on a scale that puts a PhD-level AI reasoning model in billions of pockets around the world — something people in AI like to call “intelligence too cheap to meter.”

Across multiple benchmarks, this tech meets or beats what already exists, but for less. The model was trained in under 3 million GPU hours, equating to just over a $5m training cost. This compares to estimates that Meta’s last major AI model cost $60-$70m to train, meaning the output has been delivered for 5% of the cost.

The model is also proving to be ultra efficient not only in training, but also for users. Commentators have said that it wouldn’t be long before we will be seeing lite versions of the model on tiny, powerful computers such as low-cost credit card-sized Raspberry Pi. 

This has enormous implications for the world of AI, the investment community and for businesses.

It raises serious doubts about the reasoning behind U.S. tech companies’ decision to pledge billions towards AI investment as less GPUs will be needed for LLMs than first thought, flipping the narrative on its head.

These developments could also significantly benefit smaller businesses who don’t have deep pockets for investment, offering access to previously out of reach AI tech.

Following the Government's recent 50-point action plan on AI earlier, this moment will also force UK officials to reconsider whether new mechanisms and projects are needed to excel in this new normal.

It has been a bruising moment for the Biden model, which had put export controls on chips moving from the USA to China, dashing expectations that the US can maintain its lead in AI. Far from stifling Chinese innovation, this moment highlights that Washington may have sparked it.