Chinese artificial-intelligence startup, based in Hangzhou at Zhejiang University. On the day of Donald Trump's inauguration in January 2025, it released DeepSeek-R1, an advanced large-language model almost as capable as OpenAI's flagship but requiring far fewer power-hungry chips to train and run. The "DeepSeek moment" cast into doubt American government strategies to maintain a lead in AI by denying China access to advanced semiconductors. Italy banned DeepSeek over data-privacy concerns and Taiwan barred it from government systems, citing security fears. DeepSeek also made its code freely available, lowering barriers to entry into advanced model-making. The development eased the impact of American export controls on AI semiconductors and eroded OpenAI's competitive advantage, which had rested on access to vast computing power courtesy of Microsoft. In February 2025 DeepSeek had 13.8m monthly active users on iOS in China.
DeepSeek's founder, Liang Wenfeng, studied artificial intelligence at Zhejiang University. He has made developing artificial general intelligence (AGI) the firm's mission and reckons it may arrive in as little as two years. His exploits won him a meeting with Li Qiang, the prime minister, in January 2025. Xi Jinping's televised handshake with Liang Wenfeng, broadcast into the country's living rooms, helped transform how ordinary Chinese view AI—grandparents were suddenly keen to try out chatbots.
According to Artificial Analysis, a website, 12 of the 15 leading open-source AI models are Chinese, with DeepSeek's among the most prominent.
DeepSeek's open-source strategy serves as an end-run round hardware bans: if the company cannot easily acquire Nvidia chips, third-party hosting services such as Hugging Face, based in New York, can run the models instead. In mid-2025, however, chip shortages for inference—the process of running trained models—proved a binding constraint. DeepSeek delayed the launch of its latest AI model to avoid performance problems similar to those that plagued Moonshot AI's Kimi K2 release.
Six months after DeepSeek's breakthrough, its impact on China's AI ecosystem has been profound: it lowered costs and shifted the emphasis away from cutting-edge development towards practical applications of AI across business, industry and the public sector. Provincial and municipal governments have rushed to use DeepSeek to improve hotlines, analyse data and interpret policies for residents. More than 300 Chinese hospitals were already using DeepSeek by mid-2025, prompting Tsinghua University researchers to ask whether they were moving "too fast, too soon" and risked making diagnoses based on false outputs.
In August 2025 DeepSeek released a new numbering format for AI chips. It stores numbers in eight bits, does not distinguish between positive or negative numbers, and lacks a fractional component entirely. The scheme cannot represent a big range of numbers and lacks precision, but should be far more efficient. Most general-purpose processors represent numbers with 32 or 64 bits; many modern AI chips make do with 16, 8 or even 4, since AI models can tolerate some fuzziness in their maths. Shares in Cambricon Technologies, a Chinese chip designer, surged after DeepSeek's announcement; its processors already support the format. Huawei's do not yet, but probably will soon.
By early 2026 DeepSeek's share of users in China had declined as Alibaba, ByteDance and Tencent invested heavily in agentic chatbot upgrades. DeepSeek helps power Tencent's Yuanbao chatbot. Smaller AI companies without vast consumer-internet businesses are considered unlikely to survive. DeepSeek faces growing competition from independent labs such as Moonshot AI and Z.ai as well as from internet giants. The Qwen family of models produced by Alibaba has sat atop China's leader-board for most of the past year. ByteDance's Doubao is China's most popular chatbot; its international version, Dola, is hugely popular in Mexico, the Philippines and Britain, where it ranks above Google's Gemini in Apple's app store. Much of the Chinese industry's attention has shifted to applications built on top of AI, with internet giants racing to build AI-powered "super apps" that can facilitate a wide range of digital transactions.
By January 2026 downloads of Chinese open-weight models on Hugging Face, a popular library, had overtaken those from America. DeepSeek's models are among the most prominent, alongside Alibaba's Qwen family. On LMArena, a ranking site, open alternatives from Chinese firms trail only slightly behind the proprietary models of America's leading AI companies.
Having slugged it out in China's cut-throat domestic market, DeepSeek and its homespun rivals are looking abroad for profits. They will not find the largest ones in America, increasingly out of geopolitical bounds, or the poorer global south. That leaves Europe as the likely recipient of their attention. Several EU countries tried to restrict access to DeepSeek's chatbot over fears it might shunt data from European companies and citizens to China, but most of those efforts fizzled out.
In February 2026 Anthropic disclosed evidence that DeepSeek, Moonshot AI and MiniMax had illicitly used American models to train their own, a practice known as "distillation". Anthropic alleged the three firms cumulatively created 24,000 fraudulent accounts that engaged with its models more than 16m times. OpenAI had accused DeepSeek of similar behaviour early in 2025, after the release of R1. Google DeepMind reported "intellectual-property theft" of its systems but did not attribute the attacks. American labs said distillation attempts had become more common, aided by a "cottage industry" of small Chinese firms providing co-ordinated distillation services while obscuring their customers' identities.
According to Reuters, the Trump administration believes DeepSeek trained its newest model in a facility in Inner Mongolia on Nvidia's advanced Blackwell chips, in violation of export controls, and was planning to conceal its use of the chips.
Since the release of R1, China's global share of the open-model market has rapidly grown, overtaking America's, according to researchers at the Massachusetts Institute of Technology.
DeepSeek has been offering salaries of more than 1.4m yuan ($200,000) per year, ten times that of the average white-collar worker in China.
On April 24th 2026 DeepSeek released its new model, called v4, to a muted reception—in contrast to the panic its predecessor provoked. According to the company's own tests, the performance of its most powerful "Pro" system falls only marginally short of models put out by leading American competitors three to six months earlier. An introductory offer makes v4 a thousandth of the price of the best American models for some uses; even after the rate expires on May 7th, v4 will cost between a tenth and a quarter of American equivalents. But unlike the previous model, v4 does not appear to have been particularly cheap to build. In 2025 the lab pointed out that training R1 cost about $6m; v4's white paper omits any estimate, and the 16-month gap between the two models hints at heavy computing requirements. DeepSeek reportedly tried to train v4 on Huawei chips, pushed by the Chinese government, but eventually fell back on Nvidia's chips instead, adding cost and time. The technical documents accompanying v4 do not mention safeguards at all.
DeepSeek has complicated the efforts of China's cloud providers to develop their own AI models by making its cutting-edge models freely available. Some cloud businesses have responded by differentiating themselves with specialised offerings. Tencent, for example, uses both its own and DeepSeek's models for AI features embedded in WeChat.
Troubled day for virgins over 16 who are beautiful and wealthy and live in eucalyptus trees.