The subscription model of AI essentially transforms unpredictable computing costs into a predictable revenue structure, which will be the core question the market repeatedly asks AI companies.
Author and source: Wu Duidui
Domestic AI has finally reached this point: it's no longer enough to talk about users—you must talk about profitability.
The free customer acquisition phase has ended, and AI applications are now entering the stage of "cost tiering, user tiering, and commercial validation."
Recently, the DouBao App Store page displayed paid subscription options: in addition to the free basic version, there may be a Standard Plan at 68 yuan/month, an Enhanced Plan at 200 yuan/month, and a Professional Plan at 500 yuan/month, with an annual fee of up to 5,088 yuan.
Doubao's response is that the free service will continue to be available, and the premium service plan is still under testing; related benefits have not yet been officially displayed within the product.
This matter can be viewed from several perspectives.
This is related to cost, particularly the "heavy user cost."
The most challenging aspect of a product like Doubao is that the more users enjoy using it, the higher the platform's costs become.
For an ordinary user asking a few questions occasionally, the cost may be manageable.
But if users start doing these things:
Write long-form articles, generate PPTs, perform data analysis, conduct in-depth research, generate images, create videos, enable real-time voice conversations, and execute multi-step tasks with Agents.
Then it’s not even in the same cost category.
In addition to chat, the Mac version of Doubao emphasizes capabilities such as search, image editing, writing, translation, PPT creation, and data analysis, highlighting an all-in-one workflow for image and video generation, in-depth research, meeting minutes, and document and spreadsheet processing. These functions inherently require more tokens, greater computational reasoning, and enhanced multimodal processing power than standard chat.
So DouBao's pricing is most likely not because "it can't afford to lose money on casual chats," but rather because:
High-value capabilities and power users can no longer be provided indefinitely for free.
The free version can remain available to serve as an entry point, drive daily active users, and build brand awareness; however, resource-intensive features should be tiered through memberships, usage limits, priority access, and professional plans.
The challenge in commercializing AI is "fixed revenue but variable costs."
A subscription model has an inherent contradiction:
The amount users pay each month is fixed, but the number of tokens they consume varies.
This is different from Netflix, Tencent Video, or iQIYI. For video platforms, once a show is produced, the marginal cost remains relatively limited even if users watch it multiple times. AI is different—each in-depth conversation, each video generation, and each long-context analysis requires reclaiming inference resources.
Traditional software models and the Netflix model are more like:
Develop once → Copy infinitely → Marginal cost per additional sale approaches zero
But large model services are more like:
Developing a model once → Each invocation requires computational power → The more users and the deeper their usage, the higher the inference cost.
APIs such as OpenAI and Azure OpenAI are charged per token, which essentially illustrates this point: input tokens, output tokens, long contexts, and cached inputs all have different pricing, with output tokens typically being significantly more expensive.
On OpenAI's official pricing page, GPT-5.5 short context is $2.50 per million tokens for input and $15 per million tokens for output, with cached input costing significantly less.
This is not the same economic model as selling Office, selling Photoshop, or selling an operating system.
ByteDance’s own VolcEngine also follows a similar pricing logic: the Doubao model’s pricing for developers is based on millions of tokens—for example, Doubao-Seed-2.0-pro is priced at ¥3.2 per million input tokens and ¥16 per million output tokens.
This highlights a fundamental issue:
The AI product appears to be a membership subscription, but the backend operates on a pay-per-use model. However, it is not entirely equivalent to a restaurant.
A more accurate statement would be:
AI is a hybrid of a software company, a cloud computing company, and a power-intensive manufacturing company.
If a user pays 68 yuan per month but generates PPTs, videos, and long reports excessively, the costs could consume most of their revenue.
If a user pays 500 yuan per month but primarily uses it for high-value tasks with controlled consumption, it’s a good business.
So an AI subscription essentially does one thing:
Convert uncontrollable computing costs into a predictable revenue structure.
Shift from user base competition to ARPU competition
Previously, competition among AI applications in China largely revolved around offering free services to attract users.
Why has Doubao become so successful? In addition to its product advantages, it also benefits from ByteDance’s massive traffic, strong product capabilities, and low entry barriers. Doubao is the most widely used AI chat application in China, with approximately 155 million weekly active users according to QuestMobile data, compared to DeepSeek’s 81.6 million. Meanwhile, Alibaba is also driving Qwen user growth through substantial subsidies.
But the free model has one issue:
The larger the user base, the more real the cost pressure becomes.
In particular, Chinese AI products are currently engaged in a price war. DeepSeek has drastically lowered the expected cost of models, while Alibaba, ByteDance, Tencent, and Baidu are all unwilling to relinquish their entry points. As a result, consumer-facing AI is easily trapped in an awkward situation:
Users believe AI should be free; the platform knows AI cannot be indefinitely free; investors seek growth; the company seeks a sustainable business model.
DouBao's launch of a paid version means it wants to test a question:
Are Chinese users willing to pay for AI workflows?
You're not paying for "chatting"—you're paying for someone who saves you time, creates PPTs, writes reports, conducts research, processes data, and generates videos.
This difference is crucial.
Users find it hard to pay 500 yuan per month for "you keep me company."
But if it can truly save a content creator, salesperson, teacher, student, operator, or consultant 1–2 hours each day, then the willingness to pay 68 yuan, 200 yuan, or 500 yuan becomes entirely different.
This also indicates that the free version of AI will remain available, but it will become increasingly limited.
In the future, domestic AI-native applications will likely not use a one-size-fits-all pricing model, but rather a four-tier structure:
Layer 1: Free Plan
Used for customer acquisition, building usage habits, and maintaining market share. Basic chat, simple Q&A, and lightweight search will remain free.
Layer 2: Low-Cost Membership
For regular high-frequency users, such as higher limits, faster speeds, less waiting, and better models.
Layer 3: Professional Edition
Selling PPTs, data analysis, in-depth research, document processing, code, and long-context capabilities to content creators, professionals, students, programmers, and researchers.
Layer 4: Enterprise/API/Agent Services
Pay-as-you-go, or a plan plus overage charges. This is where the real business model comes into play.
The three tiers of 68, 200, and 500 currently being offered by DouBao are essentially testing this tiered structure.
The free version addresses user scale; the standard version addresses light monetization; the enhanced and professional versions address cost recovery for heavy users.
ChatGPT, Claude, Gemini, Kimi, Tongyi, Zhipu, and DouBao are all moving toward similar structures. The only differences lie in which offers the strongest free version, which provides the most noticeable paid benefits, and which has the best cost control.
Why are AI subscriptions harder than traditional SaaS?
Each additional user, conversation, long-form summary, or agent task execution consumes more GPU inference, power, VRAM, bandwidth, storage, and engineering operations.
So the most critical issue for AI application companies is not:
Are there any users?
Instead:
The more users, the more profitable—or the more costly—does it become?
This is very different from traditional SaaS. In traditional SaaS, once the system is set up, the gross profit from new customers is typically high; however, with AI products, if users engage heavily, it can lead to significantly higher inference costs. The current market concern over Big Tech’s return on AI investment essentially stems from this issue. This year, giants like Alphabet, Microsoft, Meta, and Amazon have made massive investments in AI, and investors are now paying closer attention to when these AI expenditures will generate sufficient returns.
However, an AI subscription cannot be simply compared to a restaurant, as it is difficult for a restaurant to reduce the cost of "a bowl of noodles" by 80% annually.
But AI can.
Because model inference costs are continuously reduced by several factors:
First, chips become more powerful. Second, models become smaller, with improved distillation, quantization, and more precise MoE routing. Third, caching, batching, and context reuse reduce the cost of redundant computations. Fourth, many tasks do not require the most powerful models and can be accomplished with smaller ones. Fifth, enterprises will shift from “blindly stacking tokens” to “consuming fewer tokens per business outcome.”
So the marginal cost of AI is not zero, but neither is it a fixed ingredient cost.
It’s similar to early cloud computing: initially expensive, but costs continue to decline through scaling, hardware, and software optimizations.
This is also why OpenAI’s pricing treats “cached inputs” as significantly cheaper than regular inputs. The very existence of a caching mechanism shows that AI providers are striving to transform repetitive computations into lower-cost, software-like processes.
This means AI companies must simultaneously answer three questions:
First, how much are users willing to pay? This is the revenue side.
Second, how many tokens does a user consume each month? This is the cost side.
Third, can the model's cost decline faster than usage growth? This relates to profitability.
If the answer is:
The user is willing to pay 200 yuan, but the monthly cost is 150 yuan, so this business is average.
If the answer is: the user is willing to pay 200 yuan; the cost is only 20 yuan, and it can be reduced further to 10 yuan with model optimization.
AI applications are once again approaching a good software business.
So the true core metric of an AI business model is not DAU, nor is it downloads, but:
Revenue per paying user / inference cost per paying user.
That is, the unit economics model powered by AI.
This will, in turn, affect AI market trends.
In relation to the stock market, this matter is actually very important.
The market is currently trading AI; the first phase looks at:
Will demand for computing power surge?
So NVIDIA, TSMC, Broadcom, storage, power equipment, and data centers rose.
In the second stage, the market will ask:
Do AI applications have users?
Therefore, the user bases of ChatGPT, Doubao, Kimi, Qwen, Copilot, and Gemini will be monitored.
The third stage, which is the most critical next phase, the market will ask:
Can these users pay? Can they make money after paying?
The announcement of Baicai's pricing marks the beginning of phase three.
If the future sees these signals, the AI market will be healthier:
The paid conversion rate is strong; users have not massively churned due to pricing; there is demand for the premium high-end plan; enterprise customers are beginning to scale their purchases; inference costs continue to decline; AI features are enabling real price increases.
But if you see the opposite signal:
Users are only willing to use the free version; the paid version has poor reviews; the platform continuously lowers prices with promotions; high-frequency users have driven costs through the roof; AI application revenue is growing rapidly but with low gross margins;
Then the market will begin to wonder:
Is the AI application layer a good business?
This will further ripple upstream, because if the application layer isn’t generating profits, cloud providers and model vendors will be asked: Why are you continuing to increase your capex?
Different AI companies have completely different economic models.
Another issue is that you can't treat all AI companies as the same.
1. NVIDIA, TSMC, storage, power equipment
These are the ones selling shovels. The more others use AI, the more they profit.
They do not directly bear the end-user token costs; instead, they absorb the capital expenditures driven by the expansion of AI inference and training.
2. Cloud providers: Microsoft, Google, Amazon
They are in between.
On one hand, AI drives growth in cloud revenue; on the other, it incurs massive capital expenditures, depreciation, electricity, and data center costs. Reuters Breakingviews notes that major companies' AI spending is expanding significantly, but the market is increasingly concerned whether these investments will generate clear returns.
So the issue with cloud providers is:
Can the growth in AI cloud revenue cover the costs of data centers, GPUs, depreciation, and electricity?
3. AI Application Companies: Copilot, ChatGPT, Various Agents
The more users use the service, the higher the cost. If it’s a fixed subscription model—such as charging a flat monthly fee—but users consume heavily, the gross margin will be eroded.
Therefore, the ideal state for AI applications is not "users chatting endlessly," but:
Users are willing to pay a high fee, but actual token consumption remains controllable.
For example, if a company is willing to pay $30, $50, or $100 per month for an AI sales assistant, AI coding assistant, or AI legal assistant, while the underlying inference cost is only a few dollars, that’s a great business.
4. Traditional software companies plus AI
For companies like Microsoft, Adobe, and Salesforce, if they can integrate AI features into their existing software to increase ARPU without letting costs spiral out of control, they can turn AI into a pricing tool.
For AI, it’s not a new startup—it’s the existing software distribution channel plus an AI premium package.
So, the largest valuation discrepancy for AI lies here.
There's no need to argue whether AI is useful or has a future—AI is undoubtedly the future.
A deeper question is: Is AI fundamentally a high-margin software business or a capital-intensive industry?
Optimists believe:
AI costs will drop rapidly, applications will surge, ARPU will rise, and in the end, it will still be a software business with high margins.
Pessimists believe:
AI will become an arms race, with everyone buying GPUs, building data centers, and paying electricity bills, but users may not be willing to pay high enough prices for each token, ultimately eroding profits through infrastructure costs.
I think the truth lies in the middle:
Foundation models and cloud infrastructure will increasingly resemble capital-intensive industries; only AI applications with true distribution, real-world scenarios, and pricing power have the chance to become software businesses again.
This also explains why AI market trends may diverge.
In the first phase, the market is buying:
Anyone associated with AI sees gains.
In the second phase, the market will ask:
Who can turn AI into revenue?
In the third stage, the market will continue to ask:
Who can turn AI revenue into profit and free cash flow?
Unlike traditional software, where selling an additional unit costs almost nothing, AI consumes computational power with each service, giving it inherent cost characteristics similar to restaurants, cloud computing, and manufacturing businesses.
But AI isn't as linear as a restaurant, because model optimization, caching, chip advancements, batching, and small model routing continuously reduce the unit cost.
So what really matters in an AI business model isn't "whether there's revenue," but:
How much GPU, electricity, and token cost is required for every dollar of AI revenue.
This will be the core question that the market repeatedly asks AI companies going forward.
What is the future profit margin of AI?