Abstract
The Internet economy has long relied on near‑zero marginal cost as its core economic foundation, supporting dominant business models including subscription, commission‑based revenue and advertising. The rise of agent‑native AI overturns this underlying assumption. Every task executed by AI Agents consumes computing resources, driving non‑trivial marginal costs and continuous training overhead. This paper dissects the cost‑structure shift triggered by AI Agents, compares legacy‑Internet business logic against agent‑native economics, and analyses cascading impacts across user entry points, advertising distribution, intermediary platforms and software‑service boundaries. When enterprises orchestrate multi‑model Agent workflows in production, an API gateway can unify access control for heterogeneous large‑model endpoints. 4sapi offers consolidated routing capability to streamline traffic across diverse model backends for Agent deployments.
1. Introduction: The Onset of Agent‑Native Economic Paradigm
On April 29, 2026, Andrej Karpathy raised a core viewpoint during his talk at Red‑Hat AI Ascend: an agent‑native economy is emerging. Existing software, documents and workflows will need to be rebuilt for intelligent agents. Nevertheless, he did not elaborate on critical practical topics, including billing mechanisms and value‑distribution rules within this new economic order.
This article builds upon discussions around AI commercialisation. Prior analysis pointed out that advertising and subscription models face bottlenecks when applied independently to AI products; monetisation must shift from selling user access toward selling task‑level outcomes. This paper further explains why fundamental business‑model reconstruction is unavoidable, rooted in radical changes to cost structures brought by AI Agent systems.
Traditional Internet economics follows a well‑known cost formula:
> Total Cost (TC) = Fixed Cost (FC) + Marginal Cost (MC) × Output Volume (Q)
In classic Internet business logic, heavy one‑time R&D constitutes high fixed cost FC. Once products launch, incremental costs for additional users remain minimal. Server and bandwidth overhead barely rises as user volume expands, pushing marginal cost MC close to zero. Near‑zero marginal cost creates powerful scale effects for digital goods. Duplicating digital content incurs almost no extra expense. New users consume negligible extra storage and bandwidth resources.
Three mature business models grew directly out of this low‑marginal‑cost property:
- Subscription model: Digital copies carry minimal extra cost. Growing subscriber numbers translate directly into revenue expansion. This mechanism underpins high market valuations for SaaS vendors. Once companies cross the break‑even threshold, profit grows far faster than operational costs.
- Commission model: Platforms build underlying infrastructure for transactions, payments and contract enforcement. They collect stable commission fees for each completed transaction.
- Advertising model: Free services attract massive user traffic. Platforms resell user attention to advertisers. For search‑engine services, after search‑algorithm and advertising infrastructure are deployed, handling extra search requests adds limited marginal overhead; revenue comes primarily from advertisement clicks.
Advertising‑driven business models are sustainable not purely because of large user bases, but because platforms capture value at multiple steps within user decision journeys. Internet platforms act as information‑and‑attention matchmakers. They avoid bearing the full service cost for every individual user decision.
AI Agent systems break these established rules. Unlike traditional Internet products which reuse pre‑existing static content, Agents generate outputs dynamically in real‑time. This difference rewrites core cost structures and commercial logic.
2. Economic Dilemma for AI Agents: Non‑Zero Marginal Cost plus Recurring Training Overhead
Two critical cost characteristics distinguish Agent‑native systems from conventional Internet services.
First, marginal costs become materially positive (MC > 0). Every token processed by large‑model systems consumes GPU power and hardware amortisation expenses. Marginal expense varies with model parameters, context window size and task complexity. Although per‑token prices have declined historically, per‑task token consumption surges dramatically, creating what industry observers call the “Jevons paradox” for AI.
Unit‑token pricing is no longer trending uniformly downward. Industry statistics indicate average domestic large‑model API prices in Q2 2026 rose roughly 80% compared with Q1 2025. The sector has transitioned toward value‑based pricing. While engineering optimisations can suppress per‑unit cost growth, overall marginal‑cost reduction is far less pronounced than within Internet services, and cost decreases get offset by rising usage depth. This factor explains why services such as ChatGPT introduced paid subscription tiers.
Second, fixed costs transform into recurring capital expenditure. Training costs are no longer one‑time investments. Model iterations happen every six to twelve months. Large‑model weights become depreciable assets with an effective service life of approximately 6‑12 months. Obsolete model versions get phased out, and retraining carries substantial expense. Take OpenAI as an example: its 2025 R&D spending reached 19.18 billion US dollars, of which nearly 10.6 billion US dollars went toward compute resources for model training. This represents a major source of operating losses.
AI Agent economics deliver non‑linear scaling. Cost increases show high divergence against user‑base expansion. Infrastructure and talent expenses rise in step with user‑volume growth. R&D spending remains persistently high. Enterprises operating Agent services must keep investing capital continuously. There exists no “cross‑over point” after which profit grows exponentially as seen in traditional Internet businesses.
Cost‑structure shifts also reshape competitive moats. Legacy Internet platforms rely on network‑effect moats. For Agent‑native services, competitive advantages shift toward data flywheels and conversion costs. Accumulated conversational context becomes a valuable asset. The longer an Agent serves one user, the better it adapts to individual requirements. Switching Agent providers imposes high migration costs. Market competition is no longer purely about traffic acquisition; companies compete to capture persistent context generated by end users.
3. Underlying Logic of Agent‑Native Economy: Monetisation Shifts From Attention Capture to Task‑Outcome Delivery
People need revised mental models to understand AI Agent economics. Traditional Internet platforms compete to seize user entry points. Agent‑native platforms compete to seize task entry points.
The basic transaction unit for Internet economics is impressions or clicks. For AI Agent economics, the fundamental unit is completed tasks. A click merely represents page access. A task stands for a concrete objective delegated to AI for resolution. Agent‑system cost profiles resemble digital employees more than digital media assets.
In the past, users submitted queries to search engines. Value came from exposure and click events. When users delegate full tasks to Agents, commercial value originates from task completion and decision‑making delegation. Advertising evolves into one component inside Agent decision workflows, supporting Agents to finish end‑to‑end user assignments.
Large‑model vendors are evolving toward Agent‑platform operators, regardless of whether they originate as base‑model developers or integrated application vendors. Platform operators gain control over task‑execution workflows: they manage tool‑call sequences, data‑source priorities, business‑logic rules and closed‑loop transaction flows.
The industry is currently in a “burn‑capital‑to‑grab‑entry‑point” phase. Control over task distribution, commercial recommendation logic and transaction closure sits behind these entry points, creating intense valuation debates.
This paradigm shift drives dual‑tier economic trends. Internet products pursue standardisation to exploit near‑zero‑marginal‑cost advantages. AI Agents suit tiered supply frameworks. Task complexity directly drives cost variance. AI vendors roll out stratified product lines with tiered pricing aligned with task value. Lower‑complexity intelligent tasks carry low price tags, while cutting‑edge high‑capability Agents command premium rates.
4. Shock Effects Brought by AI‑Agent‑Driven Economic Transformation
4.1 Restructuring User Entry and Distribution Power
During the Internet era, portals, app stores and search engines dominated user entry channels. Enterprises competed fiercely for ranking positions, exposure durations and user time allocation. After Agents emerge, users can initiate diverse tasks via unified Agent‑oriented interfaces. Existing websites and applications may degrade into background data sources and tool‑call endpoints for Agents. New platform power emerges around task orchestration. Part of the “distribution tax” historically collected by traditional portals may transfer to Agent platforms.
Meanwhile, Agents drastically compress user decision journeys and shrink available room for advertising placement. Previously users needed multiple rounds of searching and comparison. Agents can produce actionable decisions within very few interaction rounds. Industries built around high‑intent comparison shopping suffer the most direct impact. Advertising insertion points shrink. Any advertisements surfaced by Agents must maintain high transparency; otherwise users lose trust in platform objectivity.
4.2 Impact on Intermediaries and Marketplace Platforms
Marketplace platforms historically delivered value by searching, filtering and matching options for end‑users. Many of these intermediate functions become automatable by Agents. Pure information‑listing intermediaries face compression. Platform competitive moats will migrate toward real‑time supply‑and‑demand coordination capabilities that Agents cannot readily replace.
For large‑scale incumbent platforms, Agents create short‑term disruption for existing advertising and transaction‑based revenue streams, yet deliver long‑term potential as business amplifiers. Platforms can reorganise commercial value flows. Instead of charging excessive advertising fees, a reasonable approach involves collecting technical service fees ranging from 1% to 10% for each Agent‑completed transaction.
4.3 Blurred Boundaries Between Software Economy and Service Economy
Conventional SaaS business models centre on seat‑based licensing. Enterprises purchase access seats, and revenue grows in line with rising employee seat counts. AI Agents break this assumption. Organisations buy capacity to finish concrete tasks. One Agent instance can run multiple workflows simultaneously. This challenges the classic “SaaS revenue scales with seat numbers” logic.
SaaS systems may split into two layers: the lower layer stores core proprietary data; the upper layer delivers task‑execution capabilities. These two layers can become decoupled. Software vendors controlling core data retain strong competitive barriers. Products whose primary value comes from graphical user‑interface wrapping risk being invoked merely as backend interfaces for Agent systems.
Software markets start eroding traditional service‑industry territory. Corporate budgets previously allocated toward hiring human labour gradually shift toward purchasing machine‑executed task capacity. Software market boundaries keep expanding into “human‑labour budget” territory.
5. Conclusion
Over the past two decades, Internet economics followed a proven growth formula: scale user volume first, then explore monetisation paths. Marginal‑cost increments for each new user stayed modest. AI Agents end this linear‑growth pattern. Real resource consumption accompanies every Agent‑completed task. More users do not guarantee higher profit margins; complexity growth of delegated tasks may push costs upward faster than revenue.
Enterprises must calculate three dimensions comprehensively: task‑completion costs, user willingness‑to‑pay and residual value retention. The new scarce resource becomes capability to fulfil assigned tasks. AI Agents redefine transaction units for digital economies. Business models, platform power structures and profit‑sharing arrangements built atop traffic metrics will undergo comprehensive reshuffling. For organisations building multi‑agent production systems, 4sapi helps harmonise traffic routing across heterogeneous large‑model backends.
International access: https://4sapi.com
Domestic access: https://4sapi.cn
