On March 1, 2025, DeepSeek, a domestic AI innovator, unveiled the remarkable theoretical cost-profit ratio of its inference system for the first time—a staggering 545%. This revelation not only resets industry benchmarks but also underscores the potential for AI large models to bridge the gap from technological marvels to profitable commercial ventures.

Behind this milestone lies not just a showcase of technical prowess but a rigorous test of the monetization models for large AI models. This article delves into the fundamental logic and future trajectories of these commercialization paths, drawing insights from DeepSeek's pioneering practices.

01. API Sales: Balancing Scale and Cost

As understanding deepens, markets and customers are increasingly pragmatic about AI large models. Many startups are transitioning from foundational models to applications and toolchains. For investors, commercial viability remains paramount.

Currently, the primary commercial path for large models involves selling APIs, with pricing based on tokens.

Recently, DeepSeek unveiled five core technologies over five days, spanning computational optimization, communication acceleration, and storage architecture. This essentially exposed its entire AI Infra technology stack, significantly boosting hardware efficiency for large model training and inference. In its article titled "Deep Dive into DeepSeek-V3/R1 Inference System," DeepSeek detailed how these technologies are integrated, forming a cohesive system. The company employs large-scale cross-node expert parallelism (EP) and a range of technical strategies to optimize the large model inference system, achieving impressive performance and efficiency.

Concluding the article, DeepSeek presented theoretical cost and profit calculations, capping off a week focused on cost reduction through open sourcing: Assuming a GPU rental cost of $2 per hour, the daily cost totals $87,072; with all tokens priced as per DeepSeek R1, the theoretical daily revenue is $562,027, yielding a cost-profit ratio of 545%.

This model's advantage lies in asset-light operations, focusing on technical optimization rather than direct end-user interaction. Market penetration is swift, with strategies like night-time API price reductions (to 25%) accelerating adoption by SMEs and fostering demand stickiness. However, risks include potential price wars and profit margin erosion due to low-price competition from domestic cloud providers.

Essentially, this model charges based on the number of calls, relying on high throughput and low costs to achieve economies of scale. The core of the API economy is a "revolution in computing power efficiency," with DeepSeek maximizing computational output per dollar through EP strategies and dynamic load balancing.

02. Advertising Revenue: The Rise of Super Traffic Gateways

The second mainstream path involves monetizing through advertising. As AI large model technology becomes a tech industry trend, downstream enterprises rush to deploy related technologies. Upstream chip manufacturers, cloud service providers, and other "water sellers" providing computing power infrastructure are in high demand, generating revenue by providing computing power to downstream clients, achieving early commercialization.

While many downstream To C enterprises have integrated AI large models into their core products, they struggle to expand revenue channels, leading to unprofitable operations due to significant operational costs.

In response, embedding ads in AI searches has emerged as a commercial strategy for downstream enterprises. During the PC internet era, traditional search engines displayed related websites, guiding users to relevant pages. With the mobile internet era, super apps like WeChat, Douyin, Taobao, and Gaode aggregated content, extending user engagement. Entering the intelligent internet era, ChatGPT further enhances the appeal of super apps by providing complete answers during interactions and integrating upstream applications' content and functionalities through open plugins, enabling users to complete complex tasks without switching apps, embodying the All-in-one concept.

Essentially, large models will become new super traffic gateways. The underlying logic involves a three-step leap: "improvement in interaction efficiency → scenario integration → economies of scale," ultimately forming an irreversible entrance monopoly. Future enterprises with advanced models and traffic matrices will dominate the distribution rights of the next-generation digital ecosystem. However, unlike traditional search engines displaying extensive content, AI search integrates information, displaying highly precise results. Excessive commercialization may clash with user interests.

03. Subscription Services: Unlocking C-end Market Value

The third mainstream path involves profiting through subscriptions, exemplified by OpenAI's ChatGPT Plus service.

This model offers stable cash flow, providing predictable revenue and alleviating financing pressure. It also allows for brand premium, with high-paying users often being tech enthusiasts or professionals, driving word-of-mouth promotion.

The challenge lies in continuously providing differentiated value to retain users. Compliance risks, such as adhering to privacy regulations across countries, increase operational complexity. While European and American users have strong payment habits, making subscription models viable, the Chinese market, with weaker payment habits, presents uncertainties regarding the model's commercial viability.

Yang Zhilin, founder of Darkside of the Moon, noted: "User-based charging fails to create greater commercial value as products evolve. Subscriptions won't be the ultimate business model."

Essentially, this model charges individual users a monthly fee for advanced features (like unlimited access and priority response). The core of the C-end subscription model is "monetizing user experience." Practically, such models tend to be closed-source, relying on "technological monopoly + paid subscriptions." However, DeepSeek's open-source strategy disrupts this model. According to Zhu Xiaohu from GSR Ventures, if open-source models' performance nears that of closed-source models, the latter will lose value.

Notably, DeepSeek's API pricing is one-tenth of the industry average, supporting privatized deployment, enabling enterprises to customize AI services independently. Outperforming numerous closed-source models, open-source DeepSeek is recognized as a victory for open-source models, stirring up the global AI large model community.

As the model gap widens, high marketing expenses transform from a competitive advantage to a financial burden. Consequently, many DeepSeek competitors have ceased massive channel investments. Media reports suggest Darkside of the Moon recently decided to significantly reduce its product launch budget, including suspending multiple Android channel launches and third-party advertising platform collaborations.

Darkside of the Moon has publicly stated that the company has been affected by "external factors and internal strategic adjustments."

Whether selling APIs, advertising, or subscription services, DeepSeek's emergence underscores the trend towards greater inclusivity in AI large model development.

04. Conclusion: DeepSeek Resolves the Large Model Commercialization 'Three-Body Problem'

In summary, large model commercialization is no easy feat. DeepSeek's case reveals the core contradiction—balancing technology, ecosystem, and capital. Its success stems from:

Technological Extremism: Achieving a qualitative leap in computing power efficiency through innovations like EP parallelism and dynamic load balancing;

Ecosystem Openness Strategy: Attracting long-tail innovation through open sourcing and screening high-value customers through tiered services;

Capital Patience: Financial support from QuantOptics enabling it to avoid short-term profitability pressures and focus on long-term technological investments.

Future large model competition will enter "deep waters": General models will pursue ultimate efficiency, vertical models will delve into industry Know-How, and the infrastructure layer will evolve into a collaborative battleground for computing power and algorithms. Only by simultaneously addressing the "three-body problem" of technological usability, commercial sustainability, and ecological richness can one lead in this AI revolution.

- End -