Chinese AI lab DeepSeek went substantially mainstream this week after its chatbot app climbed to the top of the Apple App Store charts (and Google Play as well). The compute-efficiently trained AI models raised doubts among Wall Street analysts-and technologists-about whether the U.S. would maintain dominance in the AI race and whether demand for AI chips could hold.
But where from and how did DeepSeek come into the international limelight so fast?
Trader origins of DeepSeek
DeepSeek is financed by High-Flyer Capital Management, a Chinese quantitative hedge fund trading with an AI twist.
AI fan Liang Wenfeng co-founded High-Flyer in 2015. Wenfeng was said to have been doing some trading since his student days at Zhejiang University. High-Flyer Capital Management was set up in 2019 as a hedge fund focused on building and deploying AI algorithms.
In 2023 High-Flyer set up DeepSeek as a stand-alone lab for AI tool research, separate from its trading-oriented deep-end activities. With High-Flyer as one of its investors, the lab went on to spin off its own company of the same name, DeepSeek.
DeepSeek has been setting up its data center clusters from day one for model training. But like other AI companies in China, DeepSeek suffered from the U.S. export bans on hardware. One of the company’s new models, for instance, was trained using Nvidia’s H800 chips, a less-powerful version of the chip-H100-that is allowed to be sold to U.S. companies.
The technical team in DeepSeek is believed to be predominantly young. The company purportedly aggressively recruits doctorate AI researchers from top Chinese universities. DeepSeek also hires workers from other fields, supposedly without any comp background, to broaden the tech knowledge spectrum and understanding of many subjects, according to The New York Times.
Models of DeepSeek that are powerful
DeepSeek announced the first versions — DeepSeek Coder, DeepSeek LLM, DeepSeek Chat — in November last year 2023. It was last spring that the AI section started focusing on the startup when the company launched its next generation of models dubbed DeepSeek-V2.
It was general-purpose text-and-image analysis system DeepSeek-V2 that scored fairly well on several AI benchmarks but was far lower in running costs than other comparable systems at that time. This forced the main domestic competitors of DeepSeek, such as ByteDance and Alibaba, to introduce lowered usage prices for some of their models and the rest completely free.
DeepSeek-V3 only attracted more notoriety onto DeepSeek in December 2024.
In DeepSeek’s internal benchmark tests, DeepSeek V3 outperformed both downloaded open-public models like Meta’s Llama and “closed” models that could be accessed via an API only like OpenAI’s GPT-4o.
Equally stellar is DeepSeek’s R1 “reasoning” model. DeepSeek claims on major benchmark comparisons that R1 performed similarly to OpenAI’s o1 model.
As a reasoning model, R1 fact-checks itself with this self-corrective process that enables it to dodge a few of the traps into which most models will fall. Reasoning models take a little longer, although, indeed, they will be closer to seconds-multiple of minutes later than a typical non-reasoning model to arrive at a solution. The upside of such models is that they are generally more dependable in areas like physics, science, and math.
There is, however, a downside with R1, DeepSeek V3, and all other DeepSeek models fueled by the same Chinese law. The models are under China benchmarking by the internet regulator so that they could ensure that responses “incorporate core socialist values.” So under DeepSeek’s chatbot application, for example, R1 would refrain from answering questions about Tiananmen or Taiwan.
A disruptive technique
If DeepSeek has a business model, it is not quite clear what that model is, exactly. The company prices its products and services well below market value—and gives others away for free.
The very wording of Depth Seek makes its business seem to be derived from tremendous efficiency breakthroughs. Some experts dispute the figures the company has supplied, however.
In all respects, therefore, developers have taken to DeepSeek’s models, which aren’t open source as the phrase commonly understood. Still, they are indeed available under permissive licenses allowing commercial use. As per Clem Delangue, CEO of Hugging Face, the platform hosting DeepSeek’s models, Hugging Face developers built more than 500 “derivative” models of R1 that collectively boast 2.5 million downloads.
Described as “upending AI” and “over-hyped,” DeepSeek’s success over the much more significant and established players drew this response. At least partly responsible for Nvidia’s stock price tumble of 18% on Monday, the company’s success was one of the causes for prompting OpenAI CEO Sam Altman’s public response.
That is why Microsoft extended the news that DeepSeek accesses its Azure AI Foundry service. Created by Microsoft, it includes all its AI enterprise services in a single umbrella. When asked to speak about DeepSeek effects during Meta’s first-quarter earnings call regarding AI spending, CEO Mark Zuckerberg said that investment on AI infrastructure is going to remain a strategic advantage for Meta.
However, on the other side, there are many banning DeepSeek, entire countries and governments included. DeepSeek is even banned from being in use on government devices in New York state.
As for DeepSeek’s fate, it remains uncertain. Improved models can be expected, but the US government appears to be getting more and more cautious over what it views as harmful foreign influence.
TechCrunch has a newsletter that focuses specifically on AI! Sign up here to get it in your inbox every Wednesday.
This story was first published on January 28, 2025, and will be updated continuously as more information becomes available.
Naijaeyes Report
Join Our Social Media Channels:
WhatsApp: NaijaEyes
Facebook: NaijaEyes
Twitter: NaijaEyes
Instagram: NaijaEyes
TikTok: NaijaEyes
