See Also: Live Webinar | AI-Powered Defense Against AI-Driven Threats

Hangzhou-based DeepSeek released its R1 model on Jan. 20, touting its performance as on par with the OpenAI o1 reasoning model. R1 rocketed to become the top download on the Apple App Store, surpassing OpenAI’s ChatGPT. It occupies an upper-tier position on Chatbot Arena, the AI benchmarking ranker.

DeepSeek warned users late Monday morning that it was the target of “large-scale malicious attacks,” leading to a slowdown in sign-ups. “Registration may be busy. Please wait and try again,” the company said. Chinese entrepreneur Liang Wenfeng founded DeepSeek in 2023 with funds from High Flyer, his quantitative hedge fund.

But what’s really rattling investors – who sent the Nasdaq composite down approximately 3% and the value of AI chip-designer Nvidia down more than 15% – is how much DeepSeek says it cost to develop the related V3 model: $5.6 million, compared to the hundreds of millions typically required by leading American AI companies (see: China’s DeepSeek Aims to Rival OpenAI’s ‘Reasoning’ Model). Anthropic CEO Dario Amodei in July 2024 pegged the cost of training AI models at around $100 million but said models under development would rack up costs of $1 billion.

Venture capitalist Marc Andreessen on Friday described R1 as “one of the most amazing and impressive breakthroughs I’ve ever seen.” The company’s technical paper says DeepSeek trained it “via large-scale reinforcement learning, RL, without supervised fine-tuning, SFT, as a preliminary step.” That approach allows the model to “explore chain of thought, CoT, for solving complex problems,” making it the first open model to find that large language model reasoning can be incentivized just through reinforcement learning and without supervised fine-tuning.

The cost efficiency is striking given the U.S. sanctions prohibiting the sale of advanced chips to Chinese entities. DeepSeek in a December 2024 paper said it needed only a fraction of the computing power to train the V3 model, of which R1 is a refinement. The company said it used a cluster of 2,048 Nvidia model H800 chips. Nvidia conceived H800 chips in 2023 as a sanctions-compliant version of its then H100 flagship chip. Company executives told reporters at the time that Chinese technology firms such as Alibaba, Baidu and Tencent deployed the H800 into their cloud computing offerings.

Some industry analysts sounded a skeptical note about DeepSeek’s publicity over costs. Analysts from Bernstein wrote Monday that DeepSeek’s total V3 training cost was unknown and higher than $5.6 million used for computing power, Reuters reported.

Y Combinator CEO Garry Tan wrote on social media that DeepSeek advancements will benefit the tech industry. “If training models get cheaper, faster and easier, the demand for inference – actual real-world use of AI – will grow and accelerate even faster, which assures the supply of compute will be used,” he said.

Meta Chief AI Scientist Yann LeCun urged the industry to look beyond geopolitical framing. For LeCun, DeepSeek’s success depicts the rising power of open-source innovation over proprietary models. “DeepSeek has profited from open research and open source, e.g., PyTorch and Llama from Meta,” LeCun wrote on LinkedIn. “They came up with new ideas and built them on top of other people’s work. Because their work is published and open source, everyone can profit from it.”