DeepSeek, a Chinese startup, is working to solidify its lead in the AI industry. The Hangzhou-based company, which sparked a $1 trillion-plus sell-off in global equity markets last month with its cost-effective AI reasoning model R1, is now fast-tracking the release of its successor, R2. Originally slated for an early May launch, insiders say the company is pushing to release R2 as soon as possible, although specific details remain undisclosed.
The new R2 model is expected to deliver improved coding capabilities and multi-language reasoning, extending beyond English. While DeepSeek has not commented on the accelerated timeline, competitors continue to assess R1—a model built with less-powerful Nvidia chips yet competitive with the far more expensive offerings of U.S. tech giants.
Read more: Liang Wenfeng, the founder and CEO of DeepSeek
Vijayasimha Alilughatta, COO of Indian tech services firm Zensar, noted that R2’s launch could be transformative for the AI industry, potentially prompting companies worldwide to speed up their own developments and challenging the market dominance of the few leading players.
The U.S. government is likely to be concerned about R2’s release, given its focus on AI leadership as a national priority. The move may further mobilize Chinese authorities and companies, many of which are already integrating DeepSeek’s models into their products.
DeepSeek, founded by billionaire Liang Wenfeng—who made his fortune with the quantitative hedge fund High-Flyer—has remained low-key, with Liang not speaking to the media since July 2024. Former employees describe DeepSeek as operating more like a research lab than a typical for-profit enterprise, eschewing the rigid hierarchies common in China’s tech industry. Liang’s approach, characterized by a flat management style and a willingness to work closely with young talent from prestigious institutions like Tsinghua and Peking University, has helped attract some of China’s best research minds.
While major Chinese tech companies rushed to develop consumer-facing versions of ChatGPT during the global AI boom, Liang intentionally focused on enhancing the quality of DeepSeek’s models rather than on heavy spending in app development. Both DeepSeek and its parent, High-Flyer, are reputed for their generous compensation, with senior data scientists reportedly earning as much as 1.5 million yuan annually—significantly more than industry competitors.
High-Flyer’s substantial, decade-long investments in research and computing power have been central to DeepSeek’s success. The quant fund, an early pioneer in AI trading, reinvested 70% of its revenue into AI research and spent 1.2 billion yuan on two supercomputing clusters in 2020 and 2021. The second cluster, Fire-Flyer II, comprised about 10,000 Nvidia A100 chips, which caught the attention of Chinese securities regulators. Despite initial concerns, authorities allowed the acquisition of these chips, a decision that would later prove crucial when the U.S. banned their export to China in 2022.
DeepSeek has built its competitive edge by utilizing cost-effective architectures, such as Mixture-of-Experts (MoE) and multihead latent attention (MLA), which significantly reduce computing costs. These techniques enable the AI model to activate only relevant areas of expertise for a given query and process multiple aspects of information concurrently. As a result, DeepSeek’s pricing is estimated to be 20 to 40 times lower than that of equivalent models from OpenAI, according to early February estimates from Bernstein analysts.
Even before R1 captured global attention, DeepSeek had already garnered support from Beijing. State media reported in January that Liang attended a meeting with Chinese Premier Li Qiang as the designated representative of the AI sector, positioning DeepSeek ahead of more established tech firms. The cost-competitiveness of DeepSeek’s models has led dozens of Chinese companies and state entities to adopt its technology. At least 13 Chinese city governments and 10 state-owned energy companies have integrated DeepSeek’s solutions, while major tech giants like Lenovo, Baidu, and Tencent have also incorporated the startup’s models into their products.
Chinese leader Xi Jinping, along with Liang, has signaled support for DeepSeek, a sentiment echoed by experts who observe widespread endorsement among domestic stakeholders. However, some governments—from South Korea to Italy—have removed DeepSeek’s applications from their national app stores over privacy concerns. Stephen Wu, an AI expert and founder of Carthage Capital, warned that if DeepSeek becomes the primary AI model for Chinese state entities, Western regulators might leverage this to justify further restrictions on AI chip exports and software collaborations.
Liang acknowledged that while funding has never been a constraint, the U.S. export ban on advanced AI chips remains a significant challenge for the company.
Overall, DeepSeek’s efforts to accelerate the launch of its R2 model could mark a pivotal moment in the AI industry, potentially reshaping competitive dynamics and prompting global firms to reevaluate their own strategies.
