Doubao and Qwen Retire AI Agents: How Compute Costs Are Reshaping AI's Commercial Boundaries

A Counterintuitive Product Decision

Recently, domestic large model platforms such as Doubao and Qwen have successively retired their Agent features, sparking widespread industry discussion. On the surface, this appears to be a result of tightening regulatory oversight. But according to analysis from relevant Bilibili content creators, regulation is only one factor—the real core reason lies in the companies' internal data monitoring: the vast majority of users who use Agents daily don't pay a cent.

It's worth understanding what kind of compute consumption an "Agent" actually entails. An Agent is not an ordinary single-turn Q&A, but an autonomous AI system architecture composed of a "perception-planning-execution" loop. This architecture originates from the BDI (Belief-Desire-Intention) model in classic AI research, but modern LLM-based Agents have substantially reconstructed it. Unlike traditional rule-based automation, the planning layer of an LLM Agent relies on the emergent capabilities of large models to dynamically generate execution paths, meaning every decision step must go through a full neural network forward pass. The current mainstream implementation paradigm is the ReAct (Reasoning + Acting) framework, which interweaves "thinking" and "acting" into a single token stream—the model outputs its reasoning process before generating the next action, further increasing token consumption per call. Its core components include: a task planner (which decomposes complex goals into subtask sequences), a tool-calling layer (invoking external APIs such as search engines, code interpreters, and databases), a memory management module (maintaining short-term context and a long-term knowledge base), and a feedback-iteration mechanism (dynamically adjusting plans based on execution results). It can continuously interact with the environment—searching the web, calling APIs, executing code, managing memory—until a complex goal is achieved.

The computational overhead of this architecture scales exponentially, fundamentally because multiple cost factors compound on one another: each subtask requires independent model inference, the context window continuously accumulates and expands as tasks progress (when large language models process long contexts, the computational complexity of the attention mechanism is proportional to the square of the sequence length), and frequent tool calls introduce network latency and API costs. A single Agent task may trigger dozens of model calls, and combined with tool-call latency and context window expansion, the cost of a single task can be dozens of times that of an ordinary conversation. This is precisely why "entertainment-oriented Agent users" become the group hardest for platforms to bear—they use the service frequently, interact for long periods, yet almost never pay.

What do these users do with Agents? Mostly companionship, comedy, virtual role-play, and other entertainment scenarios. For users, this is genuinely fun; but for the companies behind them—ByteDance, Alibaba, and others—it's an unbearable compute bill. Even giants of this scale are willing to give up tens of millions of users acquired at enormous cost. This decision itself shows that the severity of the problem far exceeds what we imagine.

The Cost Structure of the AI Era Has Fundamentally Changed

To understand this decision, one must grasp the fundamental difference in cost structure between AI products and internet products.

Internet Products: Marginal Cost Approaching Zero

Take scrolling through Douyin (TikTok) as an example. If you scroll for two hours and watch a hundred videos, the cost Douyin bears for this might be just a few cents. The platform casually inserts an ad and can recoup that money—even profit handsomely. This is the classic internet logic: once content is produced, the marginal cost of distribution is extremely low; the more users and the more frequent the usage, the closer the amortized per-use cost approaches zero.

This "marginal cost approaching zero" stems from the replicability of digital content: once a video is encoded, whether distributed to one person or a hundred million, the incremental cost to the server is negligible, and CDN (content delivery network) bandwidth fees are amortized across scale. More importantly, the same video file can be cached and reused infinitely at CDN nodes—the server only needs to store one copy to respond to requests worldwide, which is the physical foundation of its scale effect. This is the physical basis on which the internet advertising business model stands, and it's why the "acquire users for free first, monetize through ads later" path has repeatedly succeeded in the internet era.

AI Products: Every Conversation Burns Real Money

The logic of AI products is exactly the opposite. Each inference must execute billions of floating-point operations on a GPU cluster, consuming real electricity and hardware depreciation. Taking a GPT-4-class model as an example, industry estimates put the inference cost per million tokens in the range of a few dollars. Each time a user has a conversation with a product like Doubao, it consumes "several cents" or even "several yuan" worth of compute cost. This means the more a user uses it and the more engaged the chat, the more heavily the platform loses money.

More critically, the marginal cost of AI inference does not approach zero, fundamentally because its computational demand is directly tied to the physical limits of hardware. Understanding this at the chip-architecture level is especially important: the inference process of large language models is essentially a "memory bandwidth bound" task rather than a "compute bound" one. When generating each token, the GPU must read the complete model weights from HBM (high-bandwidth memory) into SRAM for computation—unlike during training, where the memory-read cost can be amortized through large-batch parallelism. Take the NVIDIA H100 as an example: its HBM3 memory bandwidth is approximately 3.35 TB/s, but for a GPT-4-class model with 175 billion parameters, even storing at fp8 precision requires about 175GB of memory, and the bandwidth utilization for a single token generation is extremely low—this physical constraint is called the "Inference Wall." Every inference requires loading the complete model weights into GPU memory, and regardless of user scale, the compute consumed per request is nearly fixed. GPU memory bandwidth is the physical bottleneck of inference speed—every token generated requires the model to fully read all weights from memory, and this physical process cannot be amortized by increasing the number of users. By contrast, internet content distribution relies on CDN caching, where the same content can be reused infinitely; but each AI-generated output is unique and cannot be cached or reused. This fundamental difference makes the "scale effect" far less pronounced on the inference side than on the training side. AI platforms cannot replicate the internet's "bigger equals cheaper" law—instead, they face a linear loss curve of "more users, more losses."

Even worse is the monetization dilemma: if a platform wants to recoup this compute cost through advertising, how many ads must it stuff in front of users? Stuffing until users uninstall in unbearable frustration still may not cover the cost. Entertainment-oriented, companion-chat AI applications are essentially a business model where "the more active the user, the more the company loses"—which is unsustainable in the long run.

A worthwhile reference point: the early cloud computing industry likewise went through a phase of "burn money to scale first, find paying scenarios later." After AWS launched in 2006, it continued to lose money for several years until the enterprise market matured and profitability was achieved. But the key difference between AI and cloud computing lies in the slope of the cost curve: cloud computing's marginal cost drops rapidly as scale expands (raising server utilization from 20% to 80% brings a significant drop in unit cost), whereas AI inference is constrained by the "Inference Wall," and its cost curve declines far more slowly than cloud computing's. This means the time window for AI companies waiting to "trade scale for profit" is longer and burns more capital. Against this backdrop, big companies proactively scaling back entertainment-oriented Agent features is better understood not as a strategic retreat, but as "proactively stopping the bleeding"—concentrating limited compute resources on professional users with higher commercial conversion rates.

Shifting from Mass Entertainment to Professional Markets

Having recognized this cost structure, the strategic adjustments of the big companies follow logically. The next directions for Doubao, Qwen, and others will be very clear:

• Moving toward professional knowledge assistants, focusing on scenarios that can generate real value;
• Abandoning virtual companion chat and emotional services, because such users won't pay for compute;
• Targeting professional markets willing to pay, letting those who truly need AI to create value share the cost.

The logic behind this is simple: paying will only occur naturally when AI creates value for users that far exceeds the compute cost. Companion chat brings emotional value, but emotional value is hard to quantify into willingness to pay; whereas professional scenarios—writing code, doing analysis, running workflows—bring measurable efficiency gains. This is the proper path for AI commercialization.

This strategic shift has a clear corresponding logic in economics: companies are proactively screening for "high-value users" and abandoning "high-cost, low-return users." In the traditional internet era, such screening was almost unnecessary, because the marginal cost of serving each user approached zero; but in the AI era, the cost of serving each user is real and significant, and the importance of user quality (i.e., ability and willingness to pay) is greatly elevated.

AI Coding: The First Paying Scenario to Prove Itself

Among all professional scenarios, AI Coding is currently one of the most commercially mature directions with the most intuitive value. But its threshold is not low either.

The development of AI coding tools has gone through three stages: early code completion (represented by GitHub Copilot, supporting line-level and function-level completion); mid-stage conversational programming (represented by ChatGPT Code Interpreter, supporting natural-language descriptions to generate complete code blocks); and the current autonomous coding Agent stage (represented by Cursor, Codex CLI, and Devin, capable of understanding an entire code repository, autonomously planning modification paths, executing tests, and iterating). Each capability leap is accompanied by a multiplicative increase in compute consumption—from millisecond-level inference for a single completion, to multi-round planning across an entire codebase, to closed-loop systems that can autonomously run tests, read error messages, and iterate on fixes. Compute demand grows by orders of magnitude.

The Price Barrier of Top-Tier Foreign Tools

Mainstream foreign AI coding tools like Claude and Codex start at $20 per month—and that's just the entry tier, which may only allow half an hour of use per day. To truly max out the compute and support a full day of high-intensity coding, you often need to pay up to $200 per month.

The reason for the steep pricing is that they consume the most top-tier reasoning model compute, layered with substantial Context Engineering—which is the core technical challenge of the current autonomous coding Agent stage, referring to how to efficiently organize project information within a limited token window. The complexity of this challenge far exceeds what's apparent on the surface: large code repositories may contain millions of lines of code, and mainstream models also suffer from the "Lost in the Middle" phenomenon—the model's attention to the middle portion of the context is significantly weaker than to the beginning and end, making the effectively usable context capacity far smaller than the nominal window size. Coding Agents need to combine multiple retrieval strategies: vector-embedding-based semantic search to find functionally similar code snippets, BM25-based term-frequency retrieval for precise matching of function names and variable names, as well as code retrieval augmentation (RAG), abstract syntax tree analysis (AST), dependency graph construction, and other techniques, intelligently filtering out the code snippets most relevant to the current task to feed the model, while maintaining cross-session project memory.

Among these, abstract syntax tree (AST) analysis is a key means of understanding the semantic structure of code: AST parses source code into a tree data structure, enabling the AI to understand function call relationships, variable scopes, and class inheritance hierarchies, rather than merely treating code as a text string. The dependency graph records reference relationships between modules, helping the Agent judge which downstream components might be affected by modifying a certain file, thereby avoiding unexpected side effects when planning modification paths. This complex engineering system is precisely the core competitiveness of top-tier tools like Cursor and Devin, and the underlying support for their high pricing.

For the few who aren't short of money, a few hundred or thousand yuan is nothing; but for cost-sensitive small and medium enterprises and individual developers, this is a very real threshold. Many of them have yet to experience the efficiency revolution brought by AI coding, largely because they've been shut out by cost.

The "Pinduoduo-Version Codex" Approach

Addressing this pain point, some entrepreneurs have launched AI coding platforms positioned as a "Pinduoduo-version Codex"—offering a coding experience close to top-tier tools at a lower price. By doing extensive Agent optimization on top of base models, they aim to achieve coding results not inferior to Codex while pushing prices below foreign products.

It should be noted that "lowering prices" is not just a business strategy—it also requires real breakthroughs at the technical level. Model distillation is one of the key paths: using the output of large teacher models (such as GPT-4, Claude 3.5) as training signals to guide small student models in learning complex reasoning capabilities, thereby retaining most of the performance while shrinking the parameter count by several times. The core principle of distillation is that the probability distribution output by the teacher model (soft labels) contains far richer semantic information than manual annotations (hard labels)—for example, the model considers "this code is a Python function" while also assigning a 30% probability that it might be a class method; this fine-grained uncertainty signal is extremely valuable for training the student model. In AI coding scenarios, distillation strategies typically perform domain-specific optimization for particular task types (such as Python code generation, SQL queries, front-end component writing), making small models perform close to general large models in vertical scenarios. In addition, KV cache reuse technology can reuse already-computed intermediate results in similar coding tasks, further lowering inference costs. The combination of these two technologies makes affordable AI coding services theoretically feasible—but without such real technical breakthroughs, low prices are merely subsidies rather than a sustainable business model. Such performance claims come from a single, promotional source, and readers should evaluate them rationally. But the market pain point they reveal is real: the cost-effectiveness of AI coding tools is becoming a core consideration in the tool selection of small and medium developers.

AI Coding Is Reshaping Workflows

Setting aside product promotion, the practical case studies of front-line developers are more worth noting. A developer who has written code for over a decade says that in recent months he has "not written a single line of code" and relies entirely on AI Coding:

• Sales: Having AI develop workflows to automatically mine high-intent users, improving outreach efficiency;
• Operations: Having AI develop tools for automatic editing and automatic video production, replacing inefficient manual work.

The key is that these transformations "require no technical work and can be completed entirely through natural-language conversation." Behind this is a key breakthrough of large models in the field of code generation, corresponding to the emerging paradigm of "Vibe Coding"—a concept proposed by OpenAI co-founder Andrej Karpathy in early 2025, describing a programming approach driven by natural-language intent with code generated entirely by AI throughout.

Its technical foundation lies in modern code generation models being trained not only on massive code corpora from platforms like GitHub (GitHub Copilot's training data exceeds 540 billion tokens), but also optimized through reinforcement learning (RLHF/RLAIF) to map natural-language requirements into executable code. Specifically, RLHF (reinforcement learning from human feedback) has the model generate multiple candidate code solutions, ranked and scored by human engineers, then uses this signal to train a reward model, ultimately internalizing the human judgment of "the code runs and meets the requirements" as the model's optimization objective—reducing the translation cost between business intent and implementation detail to nearly zero. In traditional software development, business staff need to convert requirements into product documents, product managers then convert them into technical specifications, and developers finally write the code—each translation carries information loss and communication cost. When AI can directly understand natural-language descriptions and generate runnable code, this chain is drastically compressed.

The deeper significance of Vibe Coding lies in how it changes the "relations of production in software production." Traditional software development has strict knowledge barriers: even the simplest business automation requirement requires understanding computer science concepts like data structures, call stacks, and exception handling, long dividing "those who can describe requirements" and "those who can implement requirements" into two distinct professional groups. When AI dissolves this barrier, business experts gain the ability, for the first time, to directly materialize ideas into software tools—this is economically analogous to "disintermediation." Just as e-commerce platforms compressed the circulation layers between producers and consumers, AI coding tools are compressing the translation layers between business requirements and software implementation.

It's worth noting that Vibe Coding still has clear boundaries: in scenarios requiring deep system design, performance tuning, or complex debugging, AI-generated code often contains hidden architectural problems that require experienced engineers to review. Common issues include: generated code lacking consideration for concurrency safety, performance bottlenecks under large data volumes, or the introduction of hard-to-trace implicit dependencies. Therefore, it is currently better suited for prototyping, script automation, business process transformation, and other scenarios with relatively relaxed code-quality requirements. But even so, its disruptiveness lies not in making programmers faster, but in enabling non-programmers to participate in software production, fundamentally expanding the potential user market boundary of AI coding tools.

This represents an important trend in AI coding—it is no longer just a tool exclusive to programmers, but a universal means for ordinary professionals and business owners to transform their business processes.

Compute Cost Determines AI's Commercial Boundaries

Doubao and Qwen retiring their Agents taught the entire industry a lesson: the business logic of AI products is ultimately constrained by the hard limit of compute cost. Entertainment-oriented, free mass applications may seem lively but are difficult to sustain; only professional scenarios that create quantifiable value and make users willing to pay can truly close the commercial loop.

From a more macro perspective, this adjustment also signals that the AI industry is about to enter a phase of "intensive cultivation." The previous logic of relying on financing subsidies to burn compute and trading user scale for valuation is failing. Whether a company can find real paying scenarios and build a sustainable revenue model will become the core variable distinguishing life and death for AI companies. The declining cost curve of GPU compute (referencing the historical pattern of roughly 2-3x cost-performance improvement per Nvidia GPU generation) will give the industry room to breathe, but before compute costs drop low enough, a clear-eyed understanding of the business model matters more than technical leadership.

For entrepreneurs and enterprises, rather than agonizing over "whether AI can chat as a companion," it's better to think about "how much labor AI can save me and how much revenue it can create." For the entire industry, how to find a sustainable business model along the curve of declining compute costs remains the most critical proposition for the coming years.

Key Takeaways

• Agent compute consumption scales exponentially: independent inference per subtask, continuously expanding context, and compounded tool-call latency make the cost of a single Agent task dozens of times that of an ordinary conversation;
• AI inference cost has an insurmountable physical boundary: GPU memory bandwidth is a hard constraint (the "Inference Wall"), and the uniqueness of each generated output means it cannot be cached and reused like internet content—the scale effect is far weaker on the inference side than on the training side;
• Entertainment-oriented free users are the group AI platforms can least afford: high-frequency, long-duration use combined with near-zero payment forms a death spiral of "the more active the user, the more the company loses";
• AI coding is currently the clearest commercial professional scenario: value is quantifiable, willingness to pay is strong, and with the rise of Vibe Coding, the potential user base is expanding from professional programmers to ordinary professionals;
• The feasibility of affordable AI coding tools depends on real technical breakthroughs: technologies like model distillation and KV cache reuse are the underlying support for "low price ≠ low quality"; low prices without a technical moat are merely unsustainable subsidies.