Summary of Key Points
AI application startups are falling into a "token dependency trap": They must spend a large amount of money to purchase tokens from top-tier AI models (such as Seedance 2.0) in order to create useful products. However, the cost of these tokens is so high that the more they sell, the greater their losses (i.e., "gross margin"). To survive, many companies either resell the tokens at a loss to generate revenue or design various "token controllers" to limit user consumption, or even rely on financing to stay afloat. Now, companies are beginning to find alternative solutions: combining hardware with applications, developing their own models, targeting enterprise customers with strong payment capabilities for services, or charging based on the final results, in order to break free from their dependence on expensive tokens.
Why Have Tokens Become the "Lifeline" for AI Applications?
For AI applications dealing with video and image processing, tokens from top-tier models are essential—without them, the products are unusable. For example, Seedance 2.0 is currently the most powerful video generation model, capable of producing videos with a cinematic quality. However, its tokens are extremely expensive: generating a 720p video for 1 second costs $0.99, which is more than 30% higher than other models. Moreover, you can't buy cheaper "gray market" tokens, as these may mix high-quality models with lower-quality versions, resulting in subpar results for users.
More importantly, buying tokens is not just about acquiring the necessary resources; it also grants access to exclusive privileges, such as high concurrency (allowing multiple users to use the model simultaneously without lag) and permissions to access large facial recognition databases. Companies like OiiOii spent $50 million on a yearly subscription to Seedance 2.0 to obtain the right to release content first and ensure high concurrency, as otherwise, users would leave if there were long waiting times. Therefore, even if it means investing 40% of their cash flow, they must buy the tokens; otherwise, they lose competitiveness and cannot survive.
Is Selling Tokens a Losing Business?
AI application companies that sell tokens act like "subletters": They purchase tokens from model manufacturers (such as Volcano Engine) at high prices and then resell them to users, but the selling price does not cover their costs. For instance, LibTV's membership allows users to generate 56-second Seedance videos for $47.4 per month, generating an income of $0.85 per second, yet the cost of using the model is $0.99 per second, resulting in a 14% loss. Even with a 10% rebate, they still lose 5%, not to mention covering other expenses (such as servers and operations).
The problem worsens when the tokens purchased from model manufacturers are non-refundable. If LibTV buys a yearly subscription for $50 million and users don't use all the tokens, the excess tokens become inventory that they have to sell at a loss (perhaps even 10% cheaper than what they bought from Volcano Engine), leading to further losses. This type of "negative gross margin" revenue is mainly used to make their annual recurring revenue (ARR) look better, which facilitates financing, but in essence, it means the more they sell, the more they lose.
Are AI Applications Becoming "Token Controllers" to Save Money?
To reduce losses, many AI applications are adopting tactics to become "token controllers":
- Limiting Usage Time: For example, they generate videos of a uniform length (e.g., 4 seconds) to avoid wasting resources on longer videos that are charged for the full duration.
- Creating Packaging Tricks: Mid-range packages do not include the best models, and high-consumption scenarios are limited by less efficient models. They also design packages that users are unlikely to exhaust, hoping users will become inactive.
- Opaque Interfaces: Companies like OiiOii do not reveal which model is used initially; only during the video editing process do users get a choice, preventing them from comparing options or optimizing their usage.
However, these measures come at the cost of a poorer user experience. Giant companies (such as Tencent WorkBuddy) do not face these issues: they have their own models and sufficient traffic, are financially robust, and can iterate on their products quickly, leaving startups with no room to compete.
How to Escape the Token Dependency?
More and more companies are seeking ways out of this dependency:
1. Combining Hardware with Applications: For example, Plaud produces AI-powered voice recorders. The hardware itself generates a 67% gross margin, and adding an AI subscription (ranging from $8 to $20 per month, with costs of only $1 to $2) raises the software's gross margin to 75%-90%. This approach avoids direct competition with model manufacturers and allows for higher pricing.
2. Developing Own Models: Companies like Mindverse have shifted from using third-party models to developing their own, which helps reduce long-term costs.
3. Targeting Wealthy Clients: By providing AI tools for industries like manufacturing or dentistry, these companies focus on generating additional revenue rather than the price per token. They are willing to pay annual fees in the tens of thousands of dollars if the tools can increase productivity and sales.
4. Charging Based on Results: Instead of charging by the number of token usage times, they charge based on the final output (e.g., a completed short video). This way, the cost of tokens no longer determines their profit.
The key to these strategies is to transform "selling tokens" into "selling value," breaking free from the dependence on top-tier model tokens. After all, tokens are merely an entry ticket, not a sustainable competitive advantage.
Conclusion
The "token dilemma" faced by AI application startups stems from their lack of core technology (their own models). To survive, they must either develop their own models or adopt different business models (using hardware, targeting businesses directly, or charging based on results). Otherwise, they will continue to sink deeper into the quagmire of high-token costs.