At the World Economic Forum in Davos last week, some mentioned Hangzhou-based DeepSeek and its recently released R1 model as a prime reason for countries such as the US to be doubling down on AI advancements. On tech chat boards, engineers had begun comparing its programming performance to leading models from the likes of OpenAI and Microsoft Corp. Its product quietly rose through the ranks of top performers on a UC Berkeley-affiliated AI leaderboard.
Then, within the past 36 hours, interest in the startup exploded. Silicon Valley heavyweights including investor Marc Andreessen and AI godfather and chief Meta Platforms Inc. scientist Yann LeCun began piling into the conversation, with Andreessen calling DeepSeek’s model “one of the most amazing and impressive breakthroughs” he’s ever seen.
By the end of the weekend, DeepSeek’s AI assistant had rocketed to the top of Apple Inc.’s iPhone download charts and ranked among the top downloads on Google’s Play Store, straining the startup’s systems so much that the service went down for more than an hour. The company was eventually forced to limit signups to those with mainland China telephone numbers — but claimed the move was the result of “large-scale malicious attacks” on its services.
The fallout from the seemingly overnight surge in interest around DeepSeek was swift, and severe: The company’s AI model, which it claims to have developed at a fraction of the cost of rivals without meaningfully sacrificing performance, drove a nearly $1 trillion rout in US and European technology stocks as investors questioned the spending plans of some of America’s biggest companies. The share plunge in AI chipmaker Nvidia Corp. alone wiped out a record $589 billion in stock-market value from the world’s largest company on Monday.
Some stocks, including Nvidia, later erased some losses in after-hours trading.
By Monday, it was clear the overwhelming interest in DeepSeek’s services was taking a toll on the company’s system. “Currently, only registration with a mainland China mobile phone number is supported,” the startup said on its status page. DeepSeek did not specify whether the signup curbs are temporary or how long they will last.
It was the company’s longest major outage since it started reporting its status. Unlike some rivals, DeepSeek’s assistant shows its work and reasoning as it addresses a user’s written query or prompt. Reviews on Apple’s app store and on Alphabet Inc.’s Android Play Store praised that transparency.
Founded by quant fund chief Liang Wenfeng, DeepSeek’s open-sourced AI model is spurring a rethink of the billions of dollars that companies have been spending to stay ahead in the AI race.
“While it remains to be seen if DeepSeek will prove to be a viable, cheaper alternative in the long term, initial worries are centered on whether US tech giants’ pricing power is being threatened and if their massive AI spending needs re-evaluation,” said Jun Rong Yeap of IG Asia.
Like all other Chinese-made AI models, DeepSeek self-censors on topics deemed politically sensitive in China. Unlike ChatGPT, DeepSeek deflects questions about Tiananmen Square, President Xi Jinping or the possibility of China invading Taiwan. That may prove jarring to international users, who may not have come into direct contact with Chinese chatbots earlier.
The initial success provides a counterpoint to expectations that the most advanced AI will require increasing amounts of computing power and energy —- an assumption that has driven shares in Nvidia and its suppliers to all-time highs.
The exact cost of development and energy consumption of DeepSeek are not fully documented, but the startup has presented figures that suggest its cost was only a fraction of OpenAI’s latest models. That a small and efficient AI model emerged from China, which has been subject to escalating US trade sanctions on advanced Nvidia chips, is also challenging the effectiveness of such measures.
“The US is great at research and innovation and especially breakthrough, but China is better at engineering,” computer scientist Kai-Fu Lee said earlier this month at the Asian Financial Forum in Hong Kong. “In this day and age, when you have limited compute power and money, you learn how to build things very efficiently.”
For its part, Nvidia — the biggest provider of chips used to train AI software — described DeepSeek’s new model as an “excellent AI advancement” that fully complies with the US government’s restrictions on technology exports. The startup’s work “illustrates how new models can be created” using a technique known as test time scaling, the company said.
Nvidia’s statement appeared to dismiss some analysts’ and experts’ suspicions that the Chinese startup couldn’t have made the breakthrough it has claimed. The company also pointed out that inference, the work of actually running AI models and using it to process data and make predictions, nonetheless requires a lot of its products.
“Inference requires significant numbers of Nvidia GPUs and high-performance networking,” the company said.
(Updates to reflect after-hours trading in sixth paragraph)
