From traffic logic to task logic, AI Agent is putting an end to the free lunch of the Internet.
[Editor's Note]
On April 29, 2026, Andrej Karpathy put forward a judgment during the fireside chat at HSG AI Ascent: a new agent-native economy is emerging, and all software, documents and processes written for human beings today will be rewritten to "be written for Agents". What he discussed is the supply-side engineering restructuring, but he left an unanswered question: if everything needs to be rewritten, who will pay for this rewriting, and how will the value be distributed.
The research topic of the agent economy starts exactly from this question. The first article At the Crossroads of AI Commercialization: Why Neither Advertising Nor Subscription Can Work Independently discusses how to monetize AI, and draws the conclusion that neither advertising nor subscription can support commercial operation alone, with the focus of commercialization shifting from "selling entry points" to "delivering results". This article digs one step further: why the business model must change. The answer lies in the cost structure. The long-held experience of the Internet that "adding one more user hardly brings extra cost" is failing, and the profit distribution system built around this experience is also due for recalculation. Over the past two decades, the core business belief of Internet companies has been: the more users, the lower the cost, and the higher the profit.
Search engines, social media, short video platforms, content platforms and e-commerce platforms are all essentially built on the same premise: after the infrastructure is completed, the marginal cost of serving one more user, displaying one more page, and distributing one more piece of content is almost zero. The real bottleneck for Internet companies is often not "whether each service loses money", but whether they can obtain sufficient user duration, data, advertisers and transaction scenarios.
It is precisely on this premise that advertising has become one of the most important business models of the Internet.
Because when content and pages can be distributed at nearly zero cost, platforms can break down users' paths of search, browsing, clicking, staying and purchasing into sellable advertising slots. The longer the user journey, the more commercial touchpoints the platform can embed; the longer users spend on comparison, the higher the value of advertisements, recommendations, brand content and search rankings.
But AI Agents are shaking this belief.
AI Agents do not simply distribute existing information, but execute tasks on behalf of users. Every conversation, every reasoning, every file reading, every tool call, and every result verification corresponds to real GPU computing power, power consumption, chip depreciation and system scheduling costs.
More importantly, AI not only changes the cost structure of Internet companies, but also reshapes the user journey on which advertising relies. In the past, users needed to search and compare repeatedly across multiple pages, platforms and brands; in the future, users may directly submit their demands to Agents, letting them filter, judge, sort, make appointments and even complete transactions.
This means that in the AI Agent era, serving each free user may result in losses, and traditional advertising will also lose part of the long-chain space on which it relies for monetization.
If the keyword of the Internet economy is "traffic", then the keyword of the AI Agent economy is "task". Traffic can be distributed at low cost, but tasks must be executed in real scenarios. In the past, advertising sold the paths that user attention passed through, while in the AI era, the more critical questions become: who understands the task, who orchestrates it, who completes it, and which brands can enter the candidate set and decision-making basis of Agents.
Therefore, the AI Agent economy is not a natural extension of the Internet business model, but a reverse correction of the Internet growth myth and advertising logic.
The Premise of Internet Advertising: Low Marginal Cost
The underlying logic of the Internet (including software) business model can be understood with the most basic formula in economics:
TC = FC + MC × Q (Total Cost = Fixed Cost + Marginal Cost × Output)
In the traditional Internet era, whether it is developing an App, writing an article or recording a video, the main cost is concentrated in the early R&D and production stage, which is the relatively high fixed cost FC. However, once the product is launched, the additional server and bandwidth costs for enterprises to serve the first user and the millionth user are often negligible, with MC close to 0.
This extremely low marginal cost endows related business models with extremely strong scale effect by nature. The core feature of digital products is that they can be copied infinitely at almost zero cost. Once the software is developed and the servers are deployed, the marginal cost of adding a new user is almost negligible, which is no more than occupying a few more KB of storage space and consuming a few more KB of bandwidth traffic.
This feature makes it easier for the platform economy to form scale advantages: the one that first reaches the critical point of scale can convert new revenue into profits at a higher proportion, thus forming a strong operating leverage.
On this basis, the Internet has evolved three classic business models.
Subscription model: The cost of software replication is extremely low, and the growth of subscription users directly corresponds to revenue growth. This is why the capital market has long been willing to give high valuations to SaaS companies: once crossing the break-even point, the profit growth rate often far exceeds the revenue growth rate.
Commission model: Build infrastructure for transactions, payment and performance, and draw a stable commission for each additional transaction.
Advertising model: Attract a large number of users with free services, and package and sell their attention to advertisers. Taking search advertising as an example, once the search algorithm and advertising system are deployed, the additional server cost for each new search request is relatively limited, and each additional ad click means incremental revenue.
The reason why the advertising model can survive for a long time is not only because of the large number of users, but also because the user journey is long enough. A user often goes through multiple links such as searching, browsing, comparing, collecting, consulting, jumping, and placing an order from generating demand to completing purchase. Platforms can sell advertisements on search result pages, promote products in information feeds, strengthen conversion on detail pages, and continue to influence user decisions in remarketing links. The value of advertising essentially comes from the layered intervention in the user journey.
Platforms do not need to bear the full service cost for every user decision. Search engines direct users to certain web pages, social platforms direct users to certain brand accounts, short video platforms direct users to certain live streaming rooms, and e-commerce platforms direct users to certain product detail pages. Platforms are more about "matching information and attention".
The Dilemma of the AI Agent Economy: Marginal Cost + Training Cost
The Internet distributes existing information, while AI generates information in real time. This qualitative change from "copying" to "generating" brings a completely different cost structure and business logic.
First of all, AI reasoning cost is explicit, and the marginal cost of AI usage is significantly greater than zero (MC > 0).
Every time a token is input or output by the large model, the GPU is running at high speed, consuming real power and chip depreciation costs. The larger the model parameters, the longer the context, and the more complex the task, the higher the single reasoning cost. The cost of a simple Q&A, a Deep Research, a code development, and a cross-software operation may differ by dozens or even hundreds of times. The AI Agent economy not only has non-zero marginal cost, but the marginal cost itself is highly heterogeneous.
Although the price per token is dropping rapidly, it does not mean that the marginal cost will continue to approach zero: with the upgrading of Agent capabilities, the token consumption per task is rising rapidly from hundreds to hundreds of thousands — planning, tool calling, file reading, self-verification, and failure retry all consume a large amount of computing power. The unit token is deflating, while the single task is inflating, and the scissors gap between the two curves constitutes the core of AI economics: the cheaper the token, the bolder people are to let AI handle heavier work, and the larger the total bill, which is exactly the "Jevons Paradox" in the AI field.
The price reduction of per token is not a one-way downward trend. Statistics from Morgan Stanley show that the average API output price of domestic large models in the second quarter of 2026 increased by about 80% compared with the first quarter of 2025, and the industry is shifting from price war to value-based pricing. The price itself is stratified: the commercialized models continue to decline, while the cutting-edge models are adjusted upward instead. Cutting-edge intelligence is priced by value, and its cost does not decline synchronously with commercialization.
Although unit cost can be reduced by improving GPU utilization through batch processing, KV caching, self-built reasoning clusters, self-developed chips and other methods, the decline range of marginal cost is far less than that of the Internet economy, and will be offset by the growth of usage depth. This is also the core reason why ChatGPT quickly launched a paid subscription version after the explosion of user volume.
At the same time, the fixed cost has changed from "one-time" to "perpetual". The training cost is no longer a one-time fixed input, but a perpetual capital expenditure for iterating new generation models every six months to one year. Large models have become "consumable assets" with a "shelf life" of 6-12 months. They will be eliminated quickly if they are not retrained and updated, and the cost of a single training can reach tens of millions or even hundreds of millions of dollars. Taking OpenAI as an example, its revenue in 2025 was 13.07 billion US dollars, with a gross profit margin of about 43% (this caliber only deducts reasoning computing power, and the most expensive training computing power is included in R&D expenditure separately), but its R&D expenditure reached as high as 19.18 billion US dollars, of which about 10.6 billion US dollars was used to pay Microsoft for training computing power, which is also the main reason for its losses.
Therefore, the AI Agent economy presents the following characteristics: reasoning cost (variable cost) grows linearly with usage and is highly differentiated; infrastructure and talent cost (semi-fixed cost) rises in a stepwise manner with the expansion of user scale — when the number of users breaks through a certain threshold, a new batch of GPU servers must be added; meanwhile, R&D cost (i.e. model training cost) is huge. These costs are superimposed layer by layer, making AI companies always stay on the track of full reinvestment of profits, and there is no such sweet spot in the Internet where "profits grow exponentially after crossing the break-even point".
This cost logic also reshapes the source of the industry's moat: the core barrier of the Internet is network effect, the more users the product is easier to use; the type of barrier for Agents has switched from the network effect of relationship chain to data flywheel and conversion cost, and the new barrier comes from the precipitated context — users' memories, preferences, files, workflows and authorization data. The longer it is used, the higher the adaptation degree of the Agent to users, and the greater the migration cost of replacing the service provider. The Internet locks users' relationship chains, while Agents lock users' working context, which is also the core reason why the industry is competing to lay out entry points: what they are fighting for is not just traffic, but the original accumulation of context.
The Underlying Logic of the AI Agent Economy: From Attention Monetization to Value Delivery
Obviously, we need to adopt a new logic to view the AI Agent economy: Internet platforms compete for "user entry points", while AI Agent platforms compete for "task entry points".
If the basic unit of the Internet economy is "impression/click", then the basic unit of the AI Agent economy is "task". A click only means that the user enters a certain page, while a task means that the user entrusts AI to complete a certain goal — Agents need to understand the goal, split the task, call tools, read files, operate software, verify results repeatedly, and request authorization from users when necessary. This makes the cost structure of AI Agents closer to that of "digital employees" rather than "digital media".
In the past, users opened search engines and asked "which decoration company is good", and search engines returned web pages, advertisements and map results; now users may directly say to Agents: "Help me filter three reliable decoration companies, compare their quotations, reputations and construction guarantees, and help me make an appointment for communication". The commercial value of the former comes from impressions and clicks, while the commercial value of the latter comes from task completion and decision-making agency. Advertising is no longer just information displayed on pages, but becomes candidates, data sources, trusted information sources or transaction interfaces in the Agent decision-making process, helping users complete tasks eventually.
Model companies are not only providing underlying capabilities, but will become a new generation of task operating systems. Whether it is AI-native manufacturers or comprehensive large-scale factories, they are all evolving from "large model" to "Agent platform", and have successively launched a series of products such as Claude Cowork, Codex and Workbuddy — the one that can grasp the user's task entry point can determine the tool calling sequence, data source priority, commercial recommendation logic and transaction closed loop.
"Knowing that the unit economy is negative, we still need to burn money to seize entry points" is exactly the core strategic gamble in the current industry, and also the source of valuation disputes. Because behind the entry point is not just the number of users, nor the number of conversations, but the control right of task distribution, commercial recommendation and transaction closed loop in the future.
This change of logic will surely bring another trend: stratification.
Internet products pursue the ultimate standardization, with one set of code serving all users, so as to maximize the advantage of zero marginal cost. AI Agents are naturally suitable for stratified supply — tasks of different complexity have huge cost differences, and the cost of simple Q&A and complex reasoning may differ by dozens or hundreds of times. Therefore, AI companies will launch products of different levels: the free version uses small models with limited usage; the basic paid version uses medium models; the advanced version uses the strongest models; the enterprise customized version provides exclusive deployment and optimization.
This stratified pricing is essentially grading by cost and charging by value, and value is positively correlated with cost. The stronger the capabilities users obtain, the higher the cost they pay. Therefore, the AI Agent ecosystem will evolve a "two-speed economy": the intelligence that is half a generation behind is extremely low-priced (open source + subsidies from large manufacturers, becoming infrastructure like water, electricity and coal), while the cutting-edge intelligence is always expensive (the strongest models are always at the most expensive end of the cost curve). The free one is the intelligence of "yesterday", and the paid one is the intelligence of "today".
The "Impact Effect" of the AI Agent Economy
The deepest impact of the AI Agent economy on the traditional economy is not that a new type of AI products appear in the market, but that the user entry points originally scattered in websites, Apps and software are being re-competed by a new batch of task agent layers. Mobile phone manufacturers, super Apps and other parties are developing their own Agents, and the entry points are not unified, but more fragmented instead. In the past, the Internet profit pool was formed around "where traffic passes through", and in the future, it will more likely be redistributed around "who orchestrates and completes the task".
The first to be rewritten is the entry point and distribution right.
In the Internet era, search engines, app stores, content platforms and super Apps control the entry points, and enterprises compete for rankings, recommendation slots and user duration. However, after the emergence of Agents, users may not need to open applications one by one: booking air tickets, finding suppliers, analyzing data, and replying to emails can all be initiated from the same task entry point. Websites and Apps will not disappear, but will more and more become data sources, tools and performance interfaces behind Agents.
This means that the new platform power is no longer just "where to direct users", but to decide how the task is completed: which model, which software, which merchant, and which set of data to call. Part of the "distribution tax" that used to be controlled by search engines, app stores and traffic platforms may gradually transfer to the hands of Agent platforms.
There is a subtle but crucial impact here: AI Agents are greatly shortening the customer journey, thus compressing the space for advertising business.
In the past, users needed to search ten times and compare multiple web pages and platforms before entering the transaction stage; in the AI scenario, users may only need to ask 3-5 consecutive questions to complete information screening, solution comparison and purchase decision. The industries with high intent, strong decision-making and heavy comparison are more likely to have their distribution logic rewritten by AI. Users no longer need to shuttle repeatedly between search results, content pages and merchant homepages, but directly entrust the AI to complete the decision-making task. At this time, the nodes where advertising can intervene will decrease, and the influence of traditional advertising will also be weakened.
The real difficulty is that the advertisements in AI responses must be highly explicit. In the past, advertisements could be embedded in the user journey in multiple scenarios such as search rankings, native information feeds, internet celebrity recommendations, and brand content, while in AI responses, any soft, hidden or unmarked commercial recommendation will directly impact users' trust in the objectivity of the platform. The more AI acts like an agent making judgments for users, the more intolerant users are that it is influenced by commercial