The Shock of China's AI "DeepSeek": 3 Years Old, Zero Overtime, and Taking on ChatGPT
機械翻訳 / Machine-translated
A
AI Friends
3日前・10分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
What you'll learn from this article
Chinese AI company DeepSeek, founded in 2023, has grown to a $60 billion (approximately ¥9 trillion) scale in just three years. Since its public release in January 2025, the app has surpassed 173 million downloads. It has become a presence that stands shoulder to shoulder with products from giants like OpenAI's ChatGPT and Google's Gemini.
Behind this rapid growth lies a unique organizational approach that drives efficient development with a small team, as well as technological innovations that have dramatically reduced costs.
DeepSeek was founded by Liang Wenfeng, born in 1985. In 2015, he established the quantitative hedge fund "High-Flyer," which by 2021 had grown into one of China's largest funds, managing $14 billion (approximately ¥2 trillion) in assets.
A quantitative hedge fund is a system that generates returns from stock investments using AI and mathematics. The large-scale computing infrastructure built at High-Flyer was directly leveraged for DeepSeek's AI development. The secret to its rapid growth lies in having started on a well-prepared foundation.
Among Chinese tech companies, a work style known as "996" has spread — working from 9 a.m. to 9 p.m., six days a week, an grueling schedule. However, DeepSeek takes a completely opposite approach.
In a dialogue with staff in August 2026, founder Liang explicitly stated, "We basically don't do overtime." Even more surprisingly, the company doesn't set KPIs (Key Performance Indicators) for its employees.
Liang has said, "Research requires a relaxed environment," and the company pursues a strategy of concentrating a small, elite team on carefully selected themes. It is believed that this work-style reform has drawn out employees' creativity and enabled rapid technological innovation in a short period of time.
"DeepSeek-V4," announced by DeepSeek in April 2026, delivered a major shock to the AI industry. Despite costing approximately 85% less than OpenAI's GPT-5.5, this model has produced results that match or exceed it in performance.
The price for DeepSeek-V4-Flash is $0.14 per one million tokens (approximately 750,000 characters) of input. Compared to services from OpenAI and Anthropic, this works out to be up to 28 times cheaper. Furthermore, the model is released under the MIT license, allowing companies to freely modify and use it.
This combination of "affordable, high-performance, and open" has caused companies around the world to take notice of DeepSeek.
In August 2026, DeepSeek released an even more powerful model specialized for coding. This model is characterized by its low price, capable of processing enormous volumes of code for just a few pennies (a few yen).
Conventionally, AI-powered code generation was dominated by high-priced services. However, DeepSeek's emergence is upending that conventional wisdom. It is now possible to access functionality equivalent to paid services like GitHub Copilot and Claude Code at a far lower cost.
Major players such as OpenAI and Google are now being forced to choose between lowering their prices or releasing even higher-performance models.
Interest in DeepSeek is also growing in Japan. In January 2025, a Japanese-language specialized model was released, and CyberAgent published a fine-tuned version.
However, there are also points of caution when adopting it. Due to China's data regulations, there is a possibility that the Chinese government could access data entered by companies, so careful judgment is required when using it for tasks involving confidential information.
On the other hand, DeepSeek's success offers important lessons for Japanese companies: that even a small team can challenge large corporations through efficient organizational management; that work-style reform has the potential to enhance creativity; and that an open-source strategy can be used to capture market share all at once.
This article is a cross-post from AI Friends.
