AI's cost structure is changing as it shifts from a tool to an autonomous executor, with both businesses and other users increasingly paying for services

image credit: Bamboo Works

Over the past two years, the business model for AI has been built on what appeared to be a solid assumption: relatively low computing costs allowed companies to scale users quickly through a "free + subscription" model. Whether in chatbots or productivity tools, AI was typically packaged as an add-on feature to enhance product appeal, rather than as a core revenue generator. Users paid fixed monthly fees for unlimited use, a model that helped accelerate widespread adoption.

But that assumption is now starting to break down.

As AI shifts from being a simple Q&A tool to agent-based products capable of completing tasks autonomously, its cost structure is changing. AI is no longer just responding to prompts. It runs continuously, repeatedly calling models and tools, and often performs multiple rounds of reasoning and refinement within a single task. This shift is transforming AI from a "passive service" into an "active executor," where the underlying consumption is no longer a one-off computation but a continuous accumulation of resources.

This shift is clearly reflected in changes in computing usage. According to media reports, daily token consumption in China's AI model market has surged from about 100 billion in early 2024 to roughly 140 trillion in 2026. This growth is driven not only by an increase in users, but also by a fundamental change in usage patterns. AI is evolving from an occasional tool into a system embedded in an enterprise's daily operations, with growth primarily driven by recurring process-based calls rather than one-off queries. As AI becomes a resource-consuming core capability, the traditional low-cost or free model becomes increasingly difficult to sustain.

This pressure is already showing up on the supply side. Between the second half of 2024 and 2026, Alibaba (NYSE:BABA) (9988.HK) and Baidu (NASDAQ:BIDU) (9888.HK) adjusted their pricing for certain cloud services, with increases ranging from 5% to 34%. As AI usage grows in both frequency and complexity, computing demand far exceeds that of traditional applications, pushing up cloud resource costs. These price adjustments reflect that rising pressure.

More importantly, the pricing model itself is evolving. In Alibaba Cloud's large model services, companies no longer pay for a product, but are charged based on token usage, with input and output priced separately. Alibaba's Ding Talk service, meanwhile, translates AI capabilities into usage counts, with different tiers offering different quotas. Whether measured in tokens or actions, the core idea is the same: AI is becoming a metered, consumable resource.

This shift is not limited to the infrastructure layer. As AI capabilities are integrated into software products, pricing structures at the application level are also rising. Kingsoft Office (3888.HK; 688111.SH) has introduced AI features such as writing, summarization and data analysis into its WPS product, similar to Microsoft Word, placing them within membership or premium tiers. Baidu has similarly bundled AI capabilities into its Wenku and cloud storage services, which now have over 40 million paying users. While these products have not universally raised their prices, they are creating new paywalls through feature tiering. For users, maintaining productivity gains often means spending more overall.

Take ByteDance's AI assistant Doubao as an example. It recently began testing three subscription tiers on top of its free version, with advanced features focused on high-complexity use cases such as generating presentations, data analysis, and content creation. While the free version remains sufficient for everyday tasks, more advanced needs — such as long-form text processing, multi-step reasoning, or structured outputs — require upgrading to a paid plan. This reflects a broader shift in AI pricing logic, where charges are no longer based solely on feature differentiation, but directly tied to underlying computing and inference costs.

In effect, this is no different from a direct price increase.

Behind these changes is a shift in cost structure. Users are paying for the move from software to computing resources. More importantly, this shift in pricing logic is reshaping how value is distributed across the industry. As AI evolves from a product into infrastructure, value begins to concentrate at the lower layers. For cloud service providers and model developers, revenue can scale alongside rising demand for computing power. In contrast, application-layer companies may benefit from efficiency gains enabled by AI, but their growing reliance on underlying resources could compress their profit margins.

AI has not eliminated costs. Instead, it has transformed them into usage-based expenses. Under this model, traditional low-cost subscription strategies will come under pressure, as costs are now directly tied to computing consumption. In their place, hybrid pricing models are emerging, combining usage volume, frequency, and even task outcomes. In this process, AI is no longer a product that can be simply priced, but is increasingly becoming a fundamental resource, much like electricity or bandwidth.

At a deeper level, this shift is rewriting the entire industry's dynamics. When computing costs become the key variable, competition is no longer defined solely by product features or user scale, but by who can secure computing power at lower cost and who can utilize model resources most efficiently. As a result, the industry's power structure is shifting, with value increasingly concentrated among providers of computing power and models.

At the same time, the logic for business expansion is also changing. In the past, scale helped dilute costs. But in the AI era, every call incurs a real expense. Under this structure, model developers and AI service providers bear the costs, agent operators convert them into revenue, and these costs are ultimately passed along to end users, creating a new layer of costs.

Benzinga Disclaimer: This article is from an unpaid external contributor. It does not represent Benzinga’s reporting and has not been edited for content or accuracy.