The AI Agent Economy Challenges the Internet’s Free Lunch Model

On April 29, 2026, Andrej Karpathy proposed during a fireside chat at Sequoia’s AI Ascent that a new economy oriented toward native agents is emerging, and that all software, documentation, and processes today designed for humans must be rewritten “for agents.” He spoke of a supply-side engineering重构, yet left unanswered the question: if everything must be rewritten, who will pay for it, and how will value be distributed?

The agent economy topic begins with this very question. The first article, "The Crossroads of AI Commercialization: Why Advertising and Subscriptions Alone Struggle to Succeed," examines how AI can generate revenue, concluding that neither advertising nor subscriptions can stand alone as viable models—shifting the focus of commercialization from "selling access" to "selling outcomes." This article digs deeper: why must the business model change? The answer lies in cost structure—the internet’s long-held assumption that "serving one more user incurs almost no additional cost" is no longer valid, and the profit-sharing frameworks built upon this assumption now require recalibration. For the past two decades, the most fundamental business belief among internet companies has been: the more users, the lower the cost, and the greater the profit.

Search engines, social media, short video platforms, content platforms, and e-commerce platforms are all built on the same fundamental premise: once the infrastructure is in place, the marginal cost of serving one additional user, displaying one more page, or distributing one more piece of content is nearly zero. The real bottleneck for internet companies is often not whether each service is profitable, but whether they can acquire sufficient user engagement time, data, advertisers, and transaction scenarios.

Under this premise, advertising has become one of the most important business models on the internet.

Because content and pages can be distributed at nearly zero cost, platforms can break down users’ search, browsing, clicking, dwell time, and purchasing paths into individual, sellable ad placements. The longer the user journey, the more commercial touchpoints the platform can integrate; the longer users compare, the higher the value of ads, recommendations, branded content, and search rankings.

But the AI agent is shaking this belief.

The AI agent is not merely distributing existing information; it is performing tasks on behalf of the user. Each conversation, each reasoning step, each file read, each tool invocation, and each result verification corresponds to real costs in GPU computing power, electricity consumption, chip depreciation, and system scheduling.

More importantly, AI has not only transformed the cost structure of internet companies but also altered the user journey on which advertising relies. In the past, users had to repeatedly search and compare across multiple pages, platforms, and brands; in the future, users may simply hand their needs over to an agent, which will filter, evaluate, rank, book, or even complete transactions on their behalf.

This means that for free users in the AI Agent era, each service provided may result in a loss, and traditional advertising will also lose part of the long-tail space it relies on for monetization.

If the keyword of the internet economy is "traffic," then the keyword of the AI Agent economy is "task." Traffic can be cheaply distributed, but tasks must be genuinely executed. In the past, advertising sold where user attention passed through; in the AI era, the more critical question becomes: who understands the task, who orchestrates it, who completes it, and which brands can enter an Agent’s candidate set and decision criteria.

Therefore, the AI Agent economy is not a natural extension of internet business models, but rather a corrective reversal of the internet growth myth and advertising logic.

The underlying logic of internet (including software) business models can be understood using the most fundamental formula in economics:

TC = FC + MC × Q (Total Cost = Fixed Cost + Marginal Cost × Quantity)

In the traditional internet era, whether developing an app, writing an article, or recording a video, the primary costs were concentrated in the initial development and production phases—resulting in high fixed costs (FC). However, once the product was launched, the additional server and bandwidth costs incurred when serving the first user versus the first million users were often negligible, with marginal costs (MC) approaching zero.

This extremely low marginal cost gives the associated business model a natural strong economies of scale. The core characteristic of digital products is their ability to be infinitely replicated at nearly zero cost. Once the software is developed and servers are deployed, the marginal cost of adding one more user is negligible—merely a few additional KB of storage space and bandwidth usage.

This feature makes it easier for platform economies to achieve economies of scale: whoever reaches the scale threshold first can convert additional revenue into profit at a higher rate, creating strong operating leverage.

On this basis, the internet evolved into three classic business models.

Subscription model: The cost of replicating software is extremely low, so growth in subscription users directly translates to revenue growth. This is why capital markets have long been willing to assign higher valuations to SaaS companies: once they surpass the break-even point, profit growth often far outpaces revenue growth.

Commission model: Build trading, payment, and fulfillment infrastructure, earning a steady commission on each additional transaction.

Advertising model: Attract a massive user base with free services, then package and sell user attention to advertisers. For example, in search advertising, once the search algorithm and ad system are deployed, the additional server cost per extra search query is relatively low, while each additional ad click generates incremental revenue.

The advertising model has sustained itself not just because of the large number of users, but because the user journey is sufficiently long. A user’s path from identifying a need to completing a purchase typically involves multiple stages—searching, browsing, comparing, saving, consulting, navigating, and placing an order. Platforms can place ads on search results pages, promote products within content feeds, enhance conversions on product detail pages, and continue influencing decisions through retargeting. The value of advertising fundamentally stems from layered engagement throughout the user journey.

The platform does not bear the full cost of service for every user decision. Search engines direct users to specific webpages, social platforms direct users to particular brand accounts, short-video platforms direct users to specific live streams, and e-commerce platforms direct users to specific product pages. The platform primarily facilitates the matching of information with attention.

The internet distributes existing information, while AI generates information in real time. This qualitative shift from “copying” to “generating” results in entirely different cost structures and business models.

First, AI inference costs are explicit, and the marginal cost of using AI is significantly greater than zero (MC > 0).

For every token input or output by a large model, the GPU operates at high speed, consuming real electricity and chip depreciation costs. The larger the model parameters, the longer the context, and the more complex the task, the higher the cost of a single inference. The cost of a simple Q&A, a Deep Research task, code development, or cross-software operation can differ by tens or even hundreds of times. The AI Agent economy not only has non-zero marginal costs, but these marginal costs are also highly heterogeneous.

Although the price per token is rapidly declining, this does not mean that marginal costs will continue to approach zero: as agent capabilities improve, the number of tokens consumed per task is quickly rising from hundreds to hundreds of thousands—planning, calling tools, reading files, self-validation, and retrying failures all consume substantial computational resources. Tokens are deflating, but tasks are inflating; the widening gap between these two trends forms the core of AI economics: the cheaper tokens become, the more confidently people task AI with heavier workloads, resulting in larger overall bills—this is precisely the “Jevons Paradox” in the AI domain.

A single token's price decline is not one-directional. According to Morgan Stanley data, the average API output price for domestic large models in Q2 2026 increased by approximately 80% compared to Q1 2025, as the industry shifts from price competition to value-based pricing. Prices themselves are also stratifying: commoditized models continue to decline, while frontier models are rising in price. Frontier intelligence is priced based on value, and its cost does not decrease in tandem with commoditized models.

Although unit costs can be reduced through methods such as batch processing to improve GPU utilization, KV caching, building in-house inference clusters, and developing proprietary chips, the rate of marginal cost reduction is far lower than in internet-based economies and is offset by increased usage depth. This is the core reason why ChatGPT rapidly introduced a paid subscription model after experiencing a surge in user numbers.

At the same time, fixed costs have shifted from “one-time” to “ongoing.” Training costs are no longer a one-time capital expenditure but rather a continuous capital outlay required to iterate to a new model generation every six to twelve months. Large models have become “consumable assets” with a “shelf life” of 6–12 months; failing to retrain and keep up leads to rapid obsolescence, with single training runs costing tens of millions or even hundreds of millions of dollars. For example, OpenAI reported $13.07 billion in revenue in 2025 with a gross margin of approximately 43% (this metric only deducts inference compute costs; the most expensive training compute is classified separately as R&D expense). However, its R&D expenditure reached $19.18 billion, of which approximately $10.6 billion was paid to Microsoft for training compute—this being the primary reason for its losses.

Therefore, the AI Agent economy is characterized by the following: reasoning costs (variable costs) increase linearly with usage and are highly differentiated; infrastructure and talent costs (semi-fixed costs) rise in “step-like” increments as user scale grows—when user volume exceeds a certain threshold, a new batch of GPU servers must be added; meanwhile, R&D costs (i.e., model training costs) are substantial. These costs stack on top of one another, causing AI companies to remain on a path where all profits are continuously reinvested, with no “sweet spot” akin to the internet industry’s phenomenon of exponential profit growth after crossing the break-even point.

This cost logic has also reshaped the industry’s moat: while the core barrier in the internet era was network effects—where the product becomes more valuable with more users—the barrier for Agents has shifted to data flywheels and switching costs. The new moat stems from accumulated context: users’ memories, preferences, files, workflows, and authorized data. The longer a user engages with an Agent, the better it adapts to them, and the higher the cost of switching providers. The internet locked in users’ relationship networks; Agents lock in users’ work contexts. This is why industry players are fiercely competing to secure entry points—it’s not just about traffic, but about the foundational accumulation of context.

Clearly, we need to adopt a different logic when viewing the AI Agent economy: internet platforms compete for "user access points," while AI Agent platforms compete for "task access 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 merely indicates that a user has entered a page, whereas a task signifies that a user has entrusted an AI to achieve a specific goal—requiring the Agent to understand the objective, break down the task, invoke tools, read files, operate software, repeatedly verify results, and, when necessary, request user authorization. This makes the cost structure of AI Agents more akin to that of a "digital employee" than a "digital medium."

Previously, users would open a search engine and ask, "Which home renovation company is good?"—and the search engine would return web pages, ads, and map results. Now, users might directly tell an agent: "Help me shortlist three reliable home renovation companies, compare their quotes, reputations, and construction guarantees, and book a consultation for me." The commercial value of the former came from exposure and clicks; the latter derives value from task completion and decision-making delegation. Ads are no longer merely information displayed on a page—they become candidates, data sources, trusted references, or transaction interfaces within the agent’s decision-making process, helping users ultimately complete their tasks.

Model companies are not just providing foundational capabilities—they are becoming the next-generation task operating systems. Both AI-native companies and large conglomerates are evolving from "large models" to "agent platforms," rolling out products such as Claude Cowork, Codex, and Workbuddy. Whoever controls the user’s task entry point will determine the order of tool invocation, the priority of data sources, the logic of commercial recommendations, and the transaction闭环.

"Even knowing that unit economics are negative, companies are burning cash to secure entry points"—this is the core strategic gamble driving the industry today and the source of valuation disputes. Because behind the entry point lies not just user volume or conversation counts, but control over the future distribution of tasks, commercial recommendations, and transaction闭环.

This shift in logic will also inevitably lead to another trend: layering.

Internet products strive for extreme standardization, using a single codebase to serve all users in order to maximize the advantage of zero marginal cost. AI agents, by contrast, are naturally suited to tiered delivery—tasks of varying complexity differ greatly in cost, with simple Q&A potentially costing dozens or even hundreds of times less than complex reasoning. As a result, AI companies offer products at different tiers: a free version using smaller models with usage limits; a basic paid version using medium-sized models; an advanced version using the most powerful models; and an enterprise customized version providing dedicated deployment and optimization.

This tiered pricing model is fundamentally cost-based grading with value-based pricing, where value correlates directly with cost. The greater the capability users gain, the higher the cost they incur. As a result, the AI Agent ecosystem will evolve into a “dual-speed economy”: lagging-edge intelligence will be extremely cheap (open-source plus big tech subsidies, becoming infrastructure like water, electricity, and gas), while cutting-edge intelligence will always remain expensive (the most powerful models will always reside at the highest end of the cost curve). The free intelligence is yesterday’s intelligence; the paid intelligence is today’s.

The deepest impact of the AI agent economy on traditional economics is not the emergence of a new category of AI products in the market, but rather the fact that user entry points previously scattered across websites, apps, and software are being re争夺 by a new layer of task agents. Smartphone manufacturers, super apps, and others are each developing their own agents, resulting in even greater fragmentation rather than unified entry points. In the past, internet profits were centered around “where traffic flowed”; in the future, they are more likely to be reallocated around “who orchestrates tasks and who completes them.”

First rewritten are the entry and distribution rights.

In the internet era, search engines, app stores, content platforms, and super apps control access points, and companies compete for rankings, featured placements, and user engagement time. But with the emergence of agents, users may no longer need to open apps one by one—booking flights, finding suppliers, analyzing data, and replying to emails can all be initiated from a single task entry point. Websites and apps won’t disappear, but they may increasingly become data sources, tools, and fulfillment interfaces operating behind the scenes for agents.

This means that the new platform's power is no longer just about directing users, but about determining how tasks are completed: which model, software, merchant, or dataset to invoke. Part of the “distribution tax” previously controlled by search engines, app stores, and traffic platforms may gradually shift to Agent platforms.

There is a subtle but critical impact: AI agents are significantly shortening the customer journey, thereby reducing the space available for advertising.

In the past, users had to search ten times, compare multiple websites and platforms, before reaching the purchase stage; under AI scenarios, users may complete information filtering, option comparison, and purchase decisions with just three to five consecutive queries. Industries characterized by high intent, strong decision-making, and intensive comparison are most likely to have their distribution logic rewritten by AI. Users no longer need to constantly navigate between search results, content pages, and merchant homepages—they can delegate the decision-making process directly to AI. As a result, the number of touchpoints where advertising can intervene decreases, and the influence of traditional advertising is diminished.

The real challenge is that advertisements in AI responses must be highly transparent. In the past, ads could be embedded throughout the user journey in various contexts such as search rankings, native feeds, influencer recommendations, and branded content. But in AI responses, any subtle, hidden, or unmarked commercial recommendation directly undermines users’ trust in the platform’s objectivity. The more the AI behaves like an agent making decisions on behalf of the user, the less tolerant users will be of its being influenced by commercial interests without transparency.

The second-layer impact will soon occur in the platform and intermediary economy.

Shopping, travel, food delivery, and local life transaction platforms fundamentally help users search, compare, and facilitate transactions—areas where agents are most easily automated. Intermediate layers that merely provide “information lists” will be compressed, while truly indispensable elements will become exclusive supply, payments, credit, logistics, after-sales service, and offline fulfillment.

The moat for trading platforms will shift from "how much traffic I have" to "how much real supply and execution capability I control that agents cannot replace."

For large companies with a closed-loop transaction system, AI may indeed disrupt existing advertising and commerce operations in the short term, but in the long run, the involvement of AI agents will become a new business amplifier. By simultaneously understanding user demand and controlling merchant supply, transactions, payments, fulfillment, and after-sales service, the platform can reorganize commercial value in a holistic, integrated way: merchants advertise on e-commerce platforms → drive higher GMV → AI prioritizes displaying brands with high GMV, high ratings, and reliable fulfillment → brands achieve a closed-loop transaction within AI-driven scenarios → the platform simultaneously enhances the value of its advertising, commerce, and AI distribution capabilities.

The key boundary here is that AI cannot charge excessively high advertising fees like traditional ad platforms. If AI charges too much, it undermines its most core assets: objectivity and user trust. A more reasonable approach is to charge a service fee at a rate aligned with industry standards, typically between 1% and 10%, primarily to cover computing power, usage, and service costs.

The third layer, and a more profound change, is the blurring of the boundary between the software economy and the service economy.

Traditional SaaS (including software) is fundamentally still a tool. Salesforce provides a CRM system, but it is salespeople who maintain customer relationships; Adobe offers design software, but it is designers who complete the designs; accounting software records data, but analysis and decision-making are still performed by humans. As a result, the SaaS business model has long revolved around the “seat”: companies purchase licenses based on the number of employees using the software. In recent years, usage-based pricing has emerged, and hybrid pricing has become mainstream, but the seat-based model remains prevalent in collaboration software.

The AI agent changes this premise.

When software begins to autonomously read data, make decisions about the next steps, and execute tasks, businesses will no longer be purchasing merely “software for an employee to use,” but rather a direct capability to accomplish tasks. An AI agent can simultaneously operate CRM, ERP, BI, email, and customer service systems, compressing processes that once required multiple people switching between different platforms into a single task invocation. At this point, the traditional growth model—“more employees, more seats, higher SaaS revenue”—will be challenged.

More importantly, SaaS itself can be layered: at the bottom is the "System of Record" that stores core data such as customers, orders, finances, and inventory; above it emerges "Systems of Action" that understand requirements, invoke the system, and complete tasks.

In the past, both were bundled within the same software interface, but in the future, they may separate. Software that controls core data, permissions, compliance, and business rules will still maintain strong barriers; however, software whose primary value comes from interface, process, and functional encapsulation faces significant risk—it could degrade from being software users open daily into mere backend interfaces called by agents.

Conversely, this will mark the first large-scale incursion of software into traditional service industries. In the past, software companies competed for enterprise IT budgets; in the future, AI agent companies will target the much larger budgets for labor, outsourcing, and professional services. Functions such as customer service, basic research, sales operations, recruitment screening, financial processing, and even certain consulting tasks may shift from “purchasing human time” to “purchasing tasks completed by machines.” As a result, the market boundary of software is expanding beyond its historical domain of “software spend” and beginning to encroach upon “labor spend.”

The internet economy of the past two decades: more users meant more traffic, more traffic meant more advertising, transaction, and subscription revenue, while the cost of acquiring new users did not increase significantly. “Grow the user base first, then find ways to monetize” became the most classic growth formula of the internet era.

The AI agent is ending this simple linear relationship.

Because AI agents now do more than provide information—they actively consume computing resources, invoke external tools, and complete tasks. More users don’t necessarily mean higher profits; more complex tasks may even lead to faster cost growth. What businesses truly need to calculate is no longer just DAU, retention rate, and user session duration, but rather: how much it costs to complete a task, how much users are willing to pay for the outcome, and how much value the platform can retain.

What is truly scarce in the AI era is changing.

In the internet era, traffic was scarce, so companies competed for users' attention; in the AI Agent era, what may truly be scarce are high-value tasks, along with the data, permissions, tools, and execution capabilities needed to accomplish them.

The AI agent is not the next product form of the internet, but rather it is transforming the most fundamental unit of digital economy transactions. What agents sell is the ability to complete tasks.

When the fundamental unit of the economy shifts from “clicks” to “tasks,” and from “using tools” to “delivering results,” the business models, platform power, and profit distributions built around traffic over the past two decades will inevitably undergo a major realignment.

This article is from the WeChat public account "Tencent Research Institute" (ID: cyberlawrc), authored by the AI Economy Research Team.