China's DeepSeek releases an open model on par with OpenAI, and Nvidia stock plunges

By early 2025 the rules of the game seemed clear: the best AI models were made by a handful of US companies that spent billions of dollars on them and kept their work closed.

On 20 January 2025, the Hangzhou-based Chinese lab DeepSeek released DeepSeek-R1. It is a reasoning model: before answering, it works through a problem step by step, like a student solving a sum on scrap paper. DeepSeek said that in math, programming and logic tasks R1 matches OpenAI's o1, one of the strongest models at the time.

The key point: R1 was released fully open, under the MIT license (named after the Massachusetts Institute of Technology), one of the most permissive licenses there is. Anyone can download the model, run it on their own computer and use it however they like, including to train other models.

A week later, on 27 January, amid the frenzy around DeepSeek, shares in Nvidia, the leading maker of AI chips, fell about 17%. In a single day the company lost roughly $600 billion in value, more than any company had ever lost in one day on the US market.

Why it matters

DeepSeek called into question two things the industry believed. The first was that world-class models are within reach only of whoever has the most money and chips, all the more so as the US restricts sales of its most powerful chips to China. The second was that the best models would be closed.

For investors, it was the first serious reason to ask whether the world was overpaying for data centers. For policymakers in Washington, it raised the question of whether the chip restrictions were working. The answers are still disputed: some argue controls should be tightened, others that they should be rethought.

For the curious

API (application programming interface, programmatic access to the model) pricing: $0.55 per million input tokens (cache miss) and $2.19 per million output tokens; a token is a fragment of a word that the model works with. Alongside R1, DeepSeek released six smaller models "distilled" from it, meaning trained to reproduce its answers. The company says the 32-billion and 70-billion-parameter versions are comparable to o1-mini.