DeepSeek Gets 82% Margins from a $7-an-Hour GPU... China's 'Survival Strategy' Against Bigger U.S. Rivals [Tech Talk]

Open-Source by Design, Under Pressure from Price Competition
AI as an Added Service on Big Tech Platforms
Extreme Efficiency Can Also Boost Margins

Editor's NoteWe make it easy to understand the technology stories behind things essential to our lives, from AI, semiconductors, telecommunications and biotechnology to technologies that may seem unfamiliar.

Global attention has turned to Anthropic, one of the leading U.S. artificial intelligence (AI) research labs, after its pre-IPO securities registration statement (S-1) became public. The company generated $16.3 billion (about 22 trillion won) in revenue through the second quarter of this year, marking 1,300% growth from a year earlier. Anthropic and OpenAI are each estimated to generate tens of trillions of won in revenue this year.

So how are Chinese AI companies competing with them doing? Surprisingly, even well-known companies such as Moonshot AI and DeepSeek, which are familiar to the general public, generate only modest revenue compared with their U.S. counterparts. Yet China's AI industry is investing capital aggressively in data centers and rushing to launch AI models. That's because Chinese AI companies make money in a different way from U.S. AI companies.

A Drop in the Ocean Compared with U.S. AI Companies

China's AI research lab DeepSeek. AP Yonhap News

Official revenue figures for Chinese AI companies have not been made public. However, Epoch AI, a global AI market research organization, compiled revenue estimates for six leading Chinese AI companies on October 6 (local time). The estimates were calculated using a range of metrics, including sales of models as a service and sales of AI tokens.

Revenue at well-known research labs such as DeepSeek and Moonshot AI is estimated at around $1 billion (about 1.4 trillion won). Revenue from AI services at major tech companies such as ByteDance and Alibaba was also estimated at $4 billion (about 5.3 trillion won) and $2.4 billion (about 3.2 trillion won), respectively. These figures are modest compared with Anthropic's estimated $65 billion (about 87 trillion won) and OpenAI's projected $40 billion (about 53 trillion won) over the same period.

China's Open-Source AI Faces a Different Business Environment from U.S. AI

Dario Amodei, CEO of Anthropic. AFP-Yonhap News Agency.

Chinese AI is already popular with people and developers around the world. DeepSeek and Alibaba's Qwen models, for example, are capable models that can be used for a wide range of tasks. Even so, they have not caught up with OpenAI and Anthropic because of their distinctive business models.

Today, OpenAI and Anthropic generate revenue by selling application programming interfaces (APIs) for their top-tier models to large customers. The companies signing API contracts with these firms are major corporations, and their employees consume enormous amounts of tokens every day. U.S. AI companies also hire "forward-deployed engineers" who work on-site at client companies to keep AI running at peak performance, providing comprehensive after-sales support.

By contrast, Chinese AI models are all open source, with core information such as model weights available to everyone. This means their designs are exposed to competitors. As soon as one company creates a new model, another soon releases a different one, prompting fierce price competition. For customers, it makes far more sense to choose the cheaper model available at any given time than to form a long-term partnership with a particular company.

AI at China's Big Tech Companies Is an Added Service on Their Platforms

ByteDance is focusing on models that enhance the platform user experience, such as AI image generation and AI filters. Screenshot of the ByteDance Seed homepage.

Why, then, do Chinese AI companies pour enormous sums into AI development despite these inherent limitations? Again, the answer lies in their business models. For one thing, major Chinese tech companies such as ByteDance and Alibaba already have huge user bases of their own. ByteDance's TikTok is a major platform that stands alongside Instagram in the United States. Chinese tech giants use AI to strengthen their services or increase the level of automation. This is similar to the strategy behind Meta's AI agent, Muse.

Epoch AI explains: "AI built into Chinese consumer applications is not general-purpose AI like U.S.-made products, but rather products specialized for specific functions. The margins on these AI products appear to be low, but the apps themselves generate revenue in various ways." In other words, AI at Chinese tech giants is closer to an added feature of their platforms than a standalone product.

Even a "Penny-Pinching" Strategy to Squeeze Every Last Drop from GPUs

DeepSeek CEO Liang Wenfeng. DeepSeek website.

Companies that make money solely from AI models, without the support of major platforms like DeepSeek, also have a strong financial foundation. Paradoxically, because Chinese AI is so cheap, it is also inexpensive to operate. Thanks to its exceptional optimization, the DeepSeek model is known to process a very large number of tokens per graphics processing unit (GPU). In other words, it can use computers, which account for most operating costs, efficiently and serve more customers with AI.

On October 20, DeepSeek published detailed information on GitHub about the efficiency of its AI inference system. According to analyses by several AI researchers, DeepSeek's gross margin is estimated at 80% to 82% per GPU costing $7 per hour (about 9,400 won). Although it earns less, its operating costs are also very low, leaving it room to generate profits and invest additional resources in research and workforce expansion.

This content was produced with the assistance of AI translation services.