The AI Race Is Shifting from Model Capabilities to Application Advantages

Token prices for models with comparable capabilities have plunged by more than 99%, prompting a reassessment of Chinese platform companies' comparative advantages

FUZHOU, China, Nov. 4, 2026 — With inference costs for models offering performance comparable to early GPT-4 falling from US$37.50 per million tokens in March 2023 to US$0.18 in February 2025, the AI race is no longer solely about model performance. A new research report from the Research Institute of Economics and Finance at CHINA INDUSTRIAL SECURITIES CO.,LTD. (CISC) indicates with the rapid decline in token prices, AI is evolving from a costly, specialized technology into widely accessible infrastructure, potentially shifting the focus of value creation from AI models to applications and platforms. This shift has made Chinese platform companies' comparative advantages deserve a reevaluation.

1. Plunging Prices: AI Is Becoming Infrastructure

Improvements in open-source models have been raising the bar for proprietary models, noted the Research Institute at its Investment Strategy Conference in May 2026. It predicted price competition would erode returns from selling tokens alone and push AI value creation further upstream and downstream. That trend is now beginning to take shape.

According to data from Epoch AI, the lowest inference price for models with capabilities comparable to early GPT-4 fell from US$37.50 per million tokens in March 2023 to US$0.18 in February 2025, a decline of more than 99%. Meanwhile, Silicon Data's token cost index dropped 52%, from a recent peak of US$2.0651 per million tokens on May 28, 2026, to US$0.9861 on September 5.

Whether measured against a fixed performance benchmark or actual market spending, the trend is clear: Comparable-capability AI becomes more accessible at rapidly falling costs. As model capabilities become more widely available, AI is evolving from a technology with high barriers to adoption into affordable, accessible infrastructure in an accelerated manner.

2. Lessons from History: Falling Costs will Fuel Demand

Historically, once costs of a critical resource fall below a certain threshold, demand often grows at a nonlinear rate.

During the U.S. IT revolution, prices for information processing equipment fell 46.5% between 1985 and 1990, while real consumption rose to approximately 5.5 times its initial level. By 1995, prices had dropped 80.7% from 1985. Consumption growth then accelerated further, with real consumption in 2000 reaching approximately 8.7 times its 1995 level.

The rollout of 4G followed a similar pattern. Between 2011 and 2013, China's average mobile data charges fell from CNY 369 per GB to CNY 205 per GB but still at a high level. Over the same period, average monthly mobile data usage per user rose only modestly, from approximately 73 MB to 139 MB. Following the issuance of 4G licenses in late 2013 and the start of large-scale commercial deployment in 2014, mobile data prices decreased much more sharply. Between 2014 and 2018, average charges plunged 93.5% from CNY 131.3 per GB to CNY 8.5 per GB, while average monthly data usage per user increased roughly 22-fold, from approximately 0.20 GB to 4.42 GB.

These historical examples point to a common pattern: when an essential resource becomes more affordable and widely available, demand previously constrained by high prices can be rapidly unlocked. For AI, the steep decline in the cost of accessing comparable levels of intelligence suggests model usage is expanding beyond a limited set of scenarios toward broader, more frequent deployment at scale.

3. A Shift in Competitive Focus: From Building Better Models to Using AI More Effectively

Once a general-purpose technology becomes widely available as an underlying capability, competition often shifts from who possesses the technology to who can make the best use of it.

As PCs and internet access became widespread, web portals and search engines emerged as focal points of competition. Once mobile internet infrastructure matured, new forms of applications such as short videos rapidly gained traction. History has repeatedly shown that the more a technology comes to resemble essential infrastructure, the more possibilities of application innovation. Value creation also tends to move away from the underlying technology toward application integration, product design and user engagement.

Smartphones and app stores had already established the foundations for mobile devices and app distribution during the 3G era. The arrival of 4G further transformed the mobile internet by enabling more video-centric, real-time and transaction-based services. In 2013, the number of mobile video users and mobile online gaming users grew 83.8% and 54.5% year over year, respectively. In 2014, smartphones overtook PCs as the primary device for watching videos online. By 2016, both ride-hailing and online food delivery services had attracted more than 200 million users in China. In 2018, the number of short video users reached 648 million, while that of mobile online payment users totaled 583 million.

Ultimately, lower connectivity costs did more than drive higher data consumption. They gave rise to a host of high-frequency applications spanning video, gaming, payments, e-commerce and local services.

4. A Reassessment of Chinese Platform Companies' Comparative Advantages

Falling token costs are expected to increase AI model use, lifting demand for upstream hardware. At the same time, they will make it easier to integrate AI into frequently used applications and services across content, gaming, advertising, office productivity, transactions and local services.

Past waves of technological innovation show that when the cost of a key element declines, the greatest opportunities for growth often emerge at the application layer. As model prices stabilize or fall further, AI value creation is likely to extend further beyond the model layer to upstream and downstream segments of the industry.

This transition calls for a reassessment of Chinese platform companies' comparative advantages. As AI competition shifts from model performance to real-world applications, factors such as the breadth and depth of application scenarios, user base size, the strength of supporting industrial ecosystems and engineering capabilities will become increasingly important.

Against this backdrop, China's advantages in complex application scenarios, its vast user base and well-established industrial ecosystem could become increasingly valuable as AI evolves into widely accessible infrastructure. Correspondingly, companies best positioned to capture value in the next phase may no longer be limited to model providers, but could also include platform and application companies with closer ties to users, real-world applications and industry workflows.

Potential risks: Slower-than-expected progress in AI development and commercial adoption; rising uncertainty of global economy and policies.

SOURCE CHINA INDUSTRIAL SECURITIES CO.,LTD.