Chinese AI startup DeepSeek, founded by Wenfeng Liang, has developed a generative AI (GenAI) service that claims to surpass large language models (LLMs) in the US, such as OpenAI's, with a low-cost, open-source approach that does not rely on advanced GPUs or HBM. With a reported development cost of under US$6 million, DeepSeek developed its V3 model in just two months, positioning itself as one of the most significant disruptors in AI for 2025.
The company's R1 model has exceeded expectations with impressive depth, breadth, objectivity, and rigor at an extraordinarily low price. Many believe DeepSeek's success could signify a shift in generative AI from the cloud to edge devices, potentially opening up new opportunities for AI applications in laptops and smartphones.
However, the company's operations are constrained by internet censorship in China. As DeepSeek's data centers are located in China, content generation is subject to government regulations. Users often encounter interruptions when querying sensitive topics such as Xinjiang Uygur Autonomous Region or the Princelings, with the AI abruptly halting content generation. Furthermore, concerns have arisen regarding the legality of DeepSeek's training data, with some questioning whether it contains intellectual property used without proper consent.
In contrast, US companies like OpenAI, Meta, Google, and Microsoft have invested billions into acquiring AI hardware. A collaborative project known as the Stargate Project, spearheaded by OpenAI, SoftBank, and Oracle, is set to invest US$500 billion, mostly in AI hardware procurement. This massive initiative will also require an immense amount of electricity to power its operations.
While maintaining a lean approach, DeepSeek has managed to develop a competitive product with R1, showing remarkable performance at a fraction of the cost compared to Western counterparts. Analysts predict that the company's success could accelerate the shift of generative AI from cloud to edge, driving AI applications on consumer devices like laptops and smartphones.
However, DeepSeek's operations are heavily influenced by China's Interim Measures for the Management of Generative Artificial Intelligence Services, which came into effect on August 15, 2023. The measures require that AI services ensure their data sources are legal, avoid intellectual property infringement, and comply with stringent data protection regulations. Any illegal content must be stopped immediately and reported to authorities—explaining why DeepSeek halts content generation when it encounters sensitive subjects.
The Chinese government also mandates that any generative AI must operate using legally sourced data and models, and AI systems involving personal data must obtain consent or meet other legal criteria. Furthermore, companies must take effective measures to ensure data quality, accuracy, and diversity, while adhering to cybersecurity regulations and oversight.
In Taiwan, a report by the National Communications Commission (NCC) indicates that 52.5% of people have used AI services, primarily in the forms of smart navigation, voice input, and biometric authentication. Meanwhile, global AI investment continues to rise. A KPMG survey for the fourth quarter of 2024 found that AI companies attracted massive venture capital funding, with Databricks securing US$10 billion, OpenAI raising US$6.6 billion, xAI acquiring US$6 billion, and Anthropic pulling in US$4 billion.
Despite a second consecutive year of declines in angel and seed-stage investments, the AI industry remains a notable exception, with startups in the sector still attracting substantial venture capital across all stages of funding.
DeepSeek's journey highlights both the vast potential of generative AI and the significant challenges it faces, particularly in a geopolitically charged environment. As the company seeks to disrupt the AI landscape, its ability to navigate these obstacles while maintaining its edge will be critical to its future success.