Stronger LLMs, Harder Business: What the Kimi Usage-Cap Controversy Reveals About AI Monetization

A Quota Incident That Reflects a Larger Industry Problem

Less than 12 hours after a new model launch, controversy erupted from two directions. On one side, promotional materials were reported to have asked partners to hide paid-collaboration labels. On the other, many users complained that their tasks hit usage caps after only a short period of operation. Two days later, the company behind the model issued a public apology, restored 25% of the monthly quota for subscribers as a one-time measure, and promised to optimize unnecessary resource consumption.

This was already the product’s third public apology for similar issues within half a year. Yet the quota incident has not truly been left behind. It is better understood as a symptom of the broader commercialization dilemma facing the large language model industry: technical capability is advancing rapidly, while the gap between user experience and commercial sustainability continues to widen.

Capability Gains and Quota Consumption: A Double-Edged Sword

Among domestic large model providers, this product has built a loyal user base through its long-context processing and content generation capabilities, and it has performed competitively against frontier models globally. But complaints about how quickly subscription quotas are consumed have persisted throughout this year.

“Those without money can’t afford it. Those with money get better cost efficiency from overseas models,” one heavy AI user commented. At the same price, the product offers far less usable quota than mainstream overseas alternatives. Another user compared the situation bluntly: “Paying 199 yuan for a certain domestic model lasts me a month. Paying the same amount for this product lasts about a day.”

For these users, a large model is not an occasional chatbot. It is production infrastructure deeply embedded in their daily workflows. As a result, they are highly sensitive to quota consumption. The launch of the new model intensified this tension. The model showed clear improvements in code generation, long-horizon task execution, and multimodal modeling. Market reaction was positive, with some observers suggesting that the gap between Chinese and American AI capabilities had narrowed substantially.

Looking at API pricing alone, the product still tries to maintain a cost advantage. The new model reduced input prices by about one-third compared with the previous generation, and output prices also dropped significantly. But for users on subscription plans, stronger model capability does not mean lower real-world cost. On the contrary, the new model consumes more tokens to complete a single task, so the total cost ends up roughly the same as before.

The Logic Behind “Stronger Models, Higher Consumption”

To understand this paradox, it helps to break down the cost structure of large models into three key dimensions.

First, model scale and inference cost are positively correlated. A stronger model usually means more parameters, longer context windows, and more complex reasoning chains. Although per-token pricing is falling, the number of tokens required to complete a task of equivalent complexity often grows several times over. In long-context processing, for example, when a model can handle a context of one million tokens, users’ actual usage scale increases accordingly. A single task can jump from thousands of tokens to tens of thousands or even hundreds of thousands.

Second, agentic behavior amplifies hidden consumption. New-generation models commonly support task planning, tool calling, and self-correction. A seemingly simple user request may trigger multiple rounds of internal reasoning and multiple external tool invocations, each consuming quota. From the user’s perspective, they submitted one task. Behind the scenes, the model may have been invoked dozens of times. This experience of quota disappearing faster than expected naturally causes frustration.

Third, subscription pricing is misaligned with usage expectations. Subscription plans usually market a fixed monthly quota. But when model capability improves dramatically, per-task consumption rises and the number of tasks that can actually be completed under a fixed quota declines. In user perception, this is equivalent to a hidden price increase.

The Dual Business Model Problem: To B and To C

Large model providers serve enterprise and consumer customers simultaneously. These two groups have different needs and different willingness-to-pay logic, yet they compete within the same pool of computing resources.

In the enterprise market, customers care most about stability, security, and predictable costs. Usage-based API pricing is relatively transparent, and enterprises can budget based on historical consumption. But the increased token volatility of new models makes cost forecasting harder. Some enterprise customers report that after switching to a new model, monthly costs did not rise dramatically, but project scheduling and resource estimation became noticeably more complex.

In the consumer market, users are extremely price-sensitive. Subscription plans are intended to provide an “all-you-can-eat” experience. When that becomes “only a few bites,” trust collapses quickly. Even more problematic, consumer users do not care much about technical details. They only perceive that they paid money and the work did not get done. Once this negative experience forms, it spreads quickly on social media and becomes a brand reputation risk.

Four Strategies for Breaking the Impasse

To escape the “stronger models, harder business” trap, large model providers need to make systematic adjustments across four dimensions: product design, business model, technical optimization, and user communication.

First, make quota measurement transparent and granular. Provide real-time token consumption details in the user interface, including input and output token counts for each task and the number of internal model invocations. Let users see clearly where their quota went instead of facing a vague “X% used” progress bar. In addition, support consumption alerts based on task type. For example, before starting a long-document processing or complex reasoning task, show an estimated consumption warning so users can make informed decisions.

Second, adopt scenario-based differentiated pricing. Resource consumption varies widely across task types. For high-frequency, low-consumption scenarios such as lightweight Q&A and simple text rewriting, offer cheaper lightweight plans. For low-frequency, high-consumption scenarios such as deep reasoning and long-horizon task execution, provide per-task billing or dedicated compute packs. Avoid a one-size-fits-all subscription model and allow users to combine options according to their actual usage patterns.

Third, optimize models for efficiency. While maintaining core capabilities, optimize attention mechanisms and inference paths to reduce unnecessary internal loops and redundant computation. For example, simple questions should converge quickly. Complex tasks can be executed in stages with intermediate results exposed, so users can detect deviations early and stop the task before the model burns through more tokens in the wrong direction.

Fourth, establish a communication mechanism that aligns capability upgrades with cost expectations. When a model upgrade significantly improves capability, publish a cost comparison guide at the same time. Explain in plain language why the new model consumes more tokens per task but delivers lower cost per unit of useful output. At the same time, offer existing users transition-period quota compensation or buffering to avoid the perception that an upgrade made the product suddenly more expensive.

The Core Commercial Question: Value, Not Price

The three apologies mentioned at the beginning of this article were not caused by immature technology. The root cause is that the business model has not kept pace with the speed of technical iteration. The large model industry is moving from a competition over parameters to a competition over experience, cost, and trust. Users never cared only about how powerful a model is. They care about whether it can solve their problems at a predictable cost.

A healthy large model business needs to satisfy three parties at once: users gain perceivable value, providers earn sustainable profit, and capability upgrades generate positive market feedback. This requires providers to be not just technical leaders but also careful planners of commercial viability.

“Stronger models, harder business” is not an unavoidable fate. When technical capability translates into user value, cost structure supports reasonable pricing, and the business model matches user expectations, large model commercialization can enter a virtuous cycle. Every public apology should be a chance to rethink the underlying business logic, not just a way to cover deeper structural problems with more compensation.