Deepseek said on Monday that it will temporarily limit the registration of new users “due to large -scale malicious attacks” into their services, although current users can still log in as usual.

This Chinese artificial intelligence startup has attracted significant attention over the past few weeks, when becoming a competitor that has been emerging quickly with Openai’s chatgting, Google Gemini and other top AI tools.

Also on Monday, Deepseek surpassed Openai’s top position on the free application charts downloaded in the US on Apple’s App Store, replacing Chatgpt with Deepseek’s own AI assistant. This has contributed to a major sell -off in the global technology stock market.

The attention for the company, founded in 2023 and has just released the R1 model last week, spread to technology analysts, investors and developers. They said that this attraction, along with the fear of being dropped in the race that is changing day by day, may be completely grounded, especially in the booming period of AI born, where the giants are attacked. Turmeric and the startups are racing to not be left behind in a market that is expected to achieve more than $ 1 trillion revenue in the next decade.

According to the report, Deepseek comes from a AI research unit of the Chinese venture capital fund in April 2023, with the goal of focusing on large language models and towards general artificial intelligence ( AGI) – A branch of AI is designed to achieve or overcome human intelligence on many tasks. This is also the goal that Openai and the opponents are all aggressively pursuing.

The attraction around Deepseek began to explode last week before the company launched R1 – theoretical model that could compete with O1 of Openai. In particular, R1 is an open source, allowing any developer to use. This model quickly climbed to the top of the application rankings and industry assessment tables, with many compliments for its outstanding performance and theoretical ability.

It is worth noting that Deepseek’s models have been developed despite the US three -time tightening of chip exports to China in the past three years. The estimates of R1 training costs are still different, but according to Jefferies analysts, a recent version of R1 has a training cost of only about 5.6 million USD (based on the GPU rental price is 2 USD/hour on H800). This figure is less than 10% compared to the cost of training LLAMA model of Meta.

Although there is no exact specific figure, the reports agree that Deepseek’s model is developed at a much lower cost than rivals such as Openai, Anthropic, Google and other companies.

The success of DeepSeeK is asking many big questions for AI industry, including whether huge capital calls and dollar valuation in the industry are really necessary or not, and whether this industry is facing. With the risk of bubbles breaking.