TORONTO — Chinese startup DeepSeek’s cut-price AI products have sparked crises of confidence in parts of the tech sector and the public markets. Here’s what you need to know.

What is it? DeepSeek, based in the southeastern Chinese city of Hangzhou, is an offshoot of a quantitative-trading hedge fund called High-Flyer. It develops large language models (LLMs), the systems that underpin generative AI tools. 

This month, the startup launched an assistant app that can respond to questions and summarize documents like ChatGPT does; it topped the iOS App Store over the weekend. Last week, DeepSeek released a new LLM called R1, which it claims performs just as well as models from leading Silicon Valley AI firms OpenAI and Anthropic on tests that require the taker to do some reasoning to figure out answers rather than just retrieving information. 

So what? DeepSeek’s models are open source, meaning developers can take and remix them as they please. And its prices are a lot cheaper than OpenAI, Anthropic or other providers, which offer restricted access to their technology via plug-ins. 

DeepSeek also claims it’s spent a lot less money and processing power to train its LLMs than other model makers. For example, it says its V3 model, released last month, cost US$5.58 million to set up; Anthropic CEO Dario Amodei said last June that firms like his spent about US$100 million each on current leading models, and had new ones in development that would run over US$1 billion.

Is bigger still better? AI advances long tracked so-called scaling laws, which suggested that LLMs would get better as they got bigger from being trained with more data and, above all, more processing power. Developers have already started to hit the limits of that approach. 

But in the meantime, major U.S. tech firms have committed hundreds of billions of dollars between them for new data centres to provide the compute power needed to keep growing their models. All those new facilities will need to be filled with hardware, which has sent AI chipmaker Nvidia’s stock skyrocketing.

But if DeepSeek can compete with OpenAI and Anthropic at a fraction of the cost—and it’s still an if—then perhaps all that spending isn’t worth it. Other developers like Toronto’s Cohere are also focusing on smaller models, arguing that specialized and higher-quality products are a better business.  

How are markets responding? Tech stocks tumbled over the weekend and through Monday, ahead of the sector’s earnings season. Nvidia’s share price dropped as much as 18.2 per cent, a US$560-billion fall in market cap that’s the largest ever recorded. Amazon, Google, Microsoft and Meta—each of which is building LLMs and spending heavily on data centres—all traded down.

How do tech bosses feel? Some of Silicon Valley’s loudest voices are warning that U.S. AI dominance has been challenged. 

Financier Marc Andreessen, whose firm has backed a lot of leading AI names, called the release of DeepSeek R1 a “Sputnik moment” for the sector; the Soviet Union’s October 1957 satellite launch set off the space race. Just as the geopolitics of that moment drove the U.S. government to spend huge sums on technology, tech firms have called for Washington to help them beat China in AI.