#夏日创作营 DeepSeek exposes a highly profitable business model: annualized 120% return crushes real estate— the era of sloppy money-making is finally over!

Recently, DeepSeek founder Liang Wenfeng publicly disclosed the core business logic: the top-tier profit business of API calls was fully laid out for outsiders. By bulk purchasing GPUs to build computing clusters, and relying on open-source large models to provide interface services externally, the entire hardware investment can be recouped in just 10 months. Within the 3 to 5 year GPU usage cycle, total profit can reach 6 times the initial cost, with annualized returns as high as 120%. This profitability not only far exceeds the domestic real-estate golden cycle, but also remains out of reach for traditional high-profit industries, tearing open the new wealth rules of the AI tech era.

Looking back at the past two decades, real estate has long been widely recognized as the common people’s wealth-making track. In China’s real-estate sector, the golden period’s compound annualized return rate is about 9.4%. Even with leverage effects layered on, the annualized returns of leading players are still hard to break through 30%. Overseas, the long-term annualized appreciation of real assets is only 4%–6%; REITs products that include rental income maintain average returns around 10%. DeepSeek’s API business with a 120% annualized return is more than ten times the traditional real-estate business, and its level of “making big profits” far exceeds the past rough, resource- and relationship-based trades. What’s even more interesting is that Liang Wenfeng publicly reveals the business model without reservation, yet he doesn’t seem to care about competitors rushing in. The core confidence lies in extreme technical cost barriers.

Industry data shows that the gap in inference and training costs among mainstream AI large-model vendors is stark. Overseas giants like OpenAI and Anthropic have extremely high overall operating costs. Second-tier model companies in China, such as Zhipu and MiniMax, also have their overall compute costs at 20 times or more than DeepSeek. Even if ordinary capital buys GPUs with huge sums and copies open-source models, it still cannot replicate DeepSeek’s accumulation at the core technical level—such as its MoE architecture, KV cache compression, and software-hardware co-optimization. In the end, they will only fall into a dilemma of high costs and low profits. This also explains why traditional entrepreneurs with massive capital, even knowing AI API is highly profitable, have still been unable to enter and get a share.

The money-making logic in the real-estate era is very simple: if you have money, connections, and can obtain land, you can basically “lie back and earn.” The barriers are concentrated in resources, with almost no technical barriers. But in the AI computing power services track, funding is only the basic threshold. The hard-core technical barriers that determine life or death—such as model R&D, inference optimization, cluster scheduling, and engineering deployment—are the real crux. A large amount of long-established capital that’s accustomed to rough expansion has heavy funds but nowhere to invest, leaving them helpless in this high-end, technology-intensive race track—this is precisely the clearest snapshot of how the times are iterating.

UBS research data confirms this trend: in 2026, more than 60% of global AI capital expenditures will flow to inference businesses. Competition in the industry is shifting from “competing on compute scale” to “competing on cost efficiency.” DeepSeek chooses to restrain profits, setting the pricing floor at payback in 10 months, giving up short-term super profits. Instead, it builds a solid moat. The open-source model not only fails to backfire on its own API business, but also forces competitors into a position where they cannot seize market share by undercutting prices.

The old era of running blind and arbitraging resources has already ended. The golden age of tech—winning through hard-core technology, long-term R&D, and fine-grained efficiency—has officially arrived. In the future, wealth allocation will continue to tilt toward teams that deeply develop technology. Opportunities to simply profit off capital and connections will become increasingly scarce. For entrepreneurs and investors, understanding DeepSeek’s profit logic is, in essence, understanding the survival rule of the next technology cycle: only by deeply developing core capabilities can you stand firm amid changes in the times.

Risk warning: This article only organizes and analyzes industry logic and public information, and does not constitute investment or entrepreneurship advice. The AI industry’s technology iterates quickly, and business models face multiple uncertainties such as policies and technology iteration.