Why AI Entrepreneurs Are Becoming 'Gig Workers' for Large Model Vendors

Introduction: A Seemingly Booming AI Startup Wave

Since 2024, AI application-layer entrepreneurship in China has been in full swing. Countless developers and small teams have rushed into the space, rapidly building AI writing assistants, intelligent customer service tools, knowledge base Q&A products, and more using APIs from large models such as Zhipu GLM, Moonshot AI's Kimi, and Baidu's ERNIE. On the surface, the barrier to AI entrepreneurship has never been lower — no need to train your own model, just call a few APIs and you can launch a product.

However, when entrepreneurs actually start crunching the numbers, a harsh reality emerges: most of the revenue from the products they painstakingly operate ultimately flows upstream to the large model vendors. As one entrepreneur put it self-deprecatingly: "I thought I was the boss. After six months, I realized I was actually an off-the-books employee of Zhipu and Kimi."

Computing Power as a Stumbling Block: The Fatal Cost Structure Dilemma

For the vast majority of AI application-layer entrepreneurs, computing power and API call fees represent the largest cost item. A typical AI application product has roughly the following cost structure:

• Large model API call fees: 40%–70% of total costs
• Servers and infrastructure: 15%–25%
• Personnel and operations: 15%–30%

This means that for every dollar an entrepreneur earns, 50 to 70 cents may go directly or indirectly to the upstream model provider. More critically, as user volumes grow, API call costs scale almost linearly, leaving extremely limited economies of scale.

One entrepreneur building an AI document summarization tool revealed: "After our product crossed 10,000 monthly active users, the monthly API bill skyrocketed from a few thousand yuan to tens of thousands. User payment rates were already low to begin with — when you do the math, there's simply no profit. The more users we get, the more we lose."

Although vendors like Zhipu and Moonshot AI have cut prices multiple times over the past year and even offered free tiers to attract developers, this "sweet first, bitter later" strategy has left many entrepreneurs anxious — once subsidies are rolled back and prices readjust, their fragile business models could collapse overnight.

Platform Dependency: The 'Subletting Landlord' Dilemma of the AI Era

This situation is not without historical precedent. During the mobile internet era, countless merchants relied on the WeChat ecosystem, Meituan's platform, and Taobao's traffic to do business, only to discover that the platforms held pricing power, traffic allocation rights, and rule-making authority, reducing merchants to 'gig workers' within the platform ecosystem.

Today, AI application-layer entrepreneurs face an almost identical structural dilemma:

First, the technological lifeline is not in their own hands. A product's core capabilities come from third-party large models. Once a model is upgraded, its interfaces adjusted, or its strategy changed, downstream applications may be forced to adapt — or simply break. Multiple developers have already reported this year that a "silent update" to a certain large model API caused their product's output quality to plummet and user complaints to surge, and they were completely powerless to do anything about it.

Second, differentiation barriers are extremely low. When everyone is calling the same model API, the only differences between products are front-end interactions and prompt engineering. Such barriers can be replicated within weeks, and competition quickly devolves into homogeneity and price wars.

Third, there are hidden risks around data and user relationships. User interaction data is transmitted via API to the large model vendor's servers. Is this data being used for model iteration and training? Could vendors leverage API call data to identify market demand and then build competing products themselves? These questions remain unanswered, yet they hang over entrepreneurs like the Sword of Damocles.

Large Model Vendors: Playing Referee and Athlete Simultaneously

Notably, vendors like Zhipu and Moonshot AI are also actively expanding into the application layer themselves. Zhipu has launched its own AI assistant "Zhipu ChatGLM," and Moonshot AI's Kimi assistant directly targets consumer users, establishing strong brand recognition in the long-text processing space.

This creates a subtle competitive dynamic: large model vendors collect API fees from entrepreneurs on one hand, while using that revenue to fund their own application products and compete directly with downstream developers for the same users.

Industry insiders have compared this relationship to "opening a shop in your landlord's mall, while the landlord has also opened a competing shop right next door." Entrepreneurs not only have to pay rent but also face asymmetric competition from a landlord wielding information and resource advantages.

Breaking Through: What Options Do Entrepreneurs Have?

Facing this dilemma, some entrepreneurs have already begun searching for ways to break out:

1. Embrace open-source models to reduce dependence on commercial APIs. Open-source models like Llama, Qwen, and DeepSeek are continuously improving, and in some scenarios have approached the performance of commercial closed-source models. By privately deploying open-source models, entrepreneurs can drastically reduce marginal costs while maintaining technological autonomy. Of course, this requires a certain level of computing investment and technical team support, which not all small teams can afford.

2. Go deep into vertical scenarios and build data and industry moats. General-purpose AI applications are easily replaceable, but diving deep into specific industries — such as legal, healthcare, or industrial quality inspection — and accumulating proprietary data and industry know-how can create moats that large model vendors cannot easily replicate in the short term.

3. Adopt a multi-model strategy to avoid being locked into a single vendor. Design architectures that support switching between multiple models, dynamically selecting providers based on cost, performance, and stability to reduce single-point dependency risks.

4. Extend upstream by exploring model fine-tuning and training capabilities. Build domain-specific fine-tuned models on top of open-source foundations, gradually establishing proprietary model assets, and evolve from an 'API middleman' into an AI company with genuine technical moats.

Outlook: Ecosystem Win-Win or Zero-Sum Game?

From an industry development perspective, the relationship between large model vendors and application-layer entrepreneurs will largely determine the health of China's AI ecosystem.

If large model vendors continue to squeeze the survival space of the application layer by "building both infrastructure and applications," then innovation vitality at the application layer will be suppressed, ultimately harming the prosperity of the entire ecosystem. Conversely, if vendors can clearly define boundaries and provide stable, reasonably priced foundational services that allow application-layer entrepreneurs to turn a profit, a win-win dynamic similar to the AWS ecosystem in the cloud computing era could emerge.

For entrepreneurs, soberly recognizing their position in the value chain, building differentiation barriers early, and reducing dependence on any single upstream provider are the real keys to surviving this AI wave. After all, being a 'gig worker' in someone else's ecosystem isn't the scary part — what's scary is working for a long time before realizing that's what you've been doing all along.