The notion that AI is solely driven by "computing power" has just suffered a significant blow. On July 28, Nvidia's stock experienced a sharp decline of nearly 5%, reducing its market capitalization to $4.76 trillion, while Apple reclaimed the top position with a market cap of approximately $4.95 trillion. This marks the first reversal since Nvidia overtook Apple over a year ago.

The immediate trigger for this reversal is the ongoing market concern that returns on AI investments may not meet expectations. However, there's another significant factor to consider: Apple's strategic position.

As the leader in smart devices, Apple has reached more consumers than almost any other brand since the advent of intelligent products. Previously, due to the scarcity of "computing capacity" in the AI market, the focus was primarily on computing power and hardware companies at the production end, while the consumer end was largely overlooked.

However, since the beginning of this year, including Apple, significant changes have occurred at the consumer end, with smart devices emerging as one of the most critical sectors. This shift suggests that the market should perhaps pay more attention to the expectations of billions of consumers worldwide. At the consumption level, they are not interested in or do not possess a deep understanding of computing power, tokens, or the differences between open-source and closed-source technologies. However, if AI truly aims to revolutionize human life, focusing solely on developing a limited number of vertical software and functions for business (B-end), government (G-end), or consumer (C-end) markets is equivalent to selectively ignoring the potential of intelligent AI terminals.

Admittedly, the market could previously attribute this oversight to the underdevelopment of AI terminals. However, in 2026, the pace of strengthening the AI supply side has also begun to slow—not because model capabilities have ceased to improve, but because the marginal returns on increasing parameter scale are visibly declining. Meanwhile, the large-scale deployment of downstream applications has not materialized as anticipated. Clients purchase API services, but their usage frequency, reach, and willingness to pay fall far short of the standards expected of a so-called new industrial revolution.

Industrial progress is not constrained by a year or two, but companies must generate returns ahead of societal trends to avoid falling behind.

This has created an awkward situation: upstream computing infrastructure investors and model manufacturers are doing their utmost to expand capacity, but downstream users are slow to embrace AI-native services that they are genuinely willing to pay for continuously. The industry suddenly realizes that the computing capacity built up during the past two years of rapid expansion could turn into a massive bubble if real demand outlets are not found.

Apple's ascent to the top at this moment feels symbolic. It reminds the entire market that no matter how powerful models are, what ultimately reaches users is always a screen, a microphone, and a device. As the AI industry moves from the excitement of "what can be done" to the deeper waters of "who will use it and how," those closest to users are regaining the power to define value distribution.

1. Why AI Cannot Bypass Terminals

This is not simply a story of Apple's good fortune. The importance of terminals for AI lies in the users themselves.

Cloud-based large models can answer any question, but they cannot answer "what you need right now." For instance, a user rushing through Shanghai Pudong Airport to catch a flight has their phone aware of their location, flight information stored in their calendar, and health data showing their heart rate spiking from running. While the cloud model possesses all world knowledge, it does not know that what this specific person needs most at this moment is not a sentence but an automatically popped-up gate navigation. Real-time perception of personal contexts requires full interaction with the edge side.

Moreover, for any intelligent agent to truly act on behalf of users, it must inevitably remember their preferences, habits, social relationships, and behavioral styles over the long term.

If this data circulates in the cloud, every invocation carries a risk of privacy exposure. Edge-side processing keeps sensitive data on the device locally, with models performing local inference and only returning necessary, anonymized results to the cloud. This forms the foundational architecture for building user trust in AI.

To date, most people, even if they do not understand the technical details, find it difficult to directly enable permissions like "allow reading all emails" or "allow access to real-time location" in cloud AI assistants. However, on their phone's system-level AI, these permissions can be granted cautiously and contextually.

Given the diversity of these applications, there is a growing consensus in the industry that not all tasks require the most powerful models. Simple intent recognition, message sorting, and photo classification—these high-frequency, low-complexity inference tasks can be completed faster, more cost-effectively, and more privately using small edge-side models than by invoking large cloud models.

Lenovo Group has fully benefited from the AIPC trend this year. Its chairman, Yang Yuanqing, mentioned during an earnings call that the future goal is to process 80% of tokens on the edge or device side, with only 20% requiring cloud access. This is not just about cost but also efficiency—edge-side inference latency is millisecond-level, while cloud inference involves network requests and server queuing, with latency gaps widening dramatically in Agent scenarios requiring continuous interaction.

Qualcomm CEO Cristiano Amon has also judged that conversational AI tasks may involve around 10,000 tokens per session, complex reasoning tasks up to about 100,000 tokens, and intelligent agent tasks could surge to the million-token level. When token consumption scales up by orders of magnitude, keeping everything in the cloud poses severe challenges to both cost and efficiency for business models.

From a commercial perspective, cloud-based large models are inherently a scale-driven, winner-takes-all business, as model convergence ultimately compresses differentiation space, and customer migration costs are extremely low. However, terminals are hardware carriers, and hardware naturally has a locking effect. When a user continuously uses the same intelligent agent on their phone for half a year, the agent accumulates a deep understanding of that person's habits, preferences, and social circle, making it difficult for the user to migrate to another agent—since the new agent would have to start from scratch.

The long-term usage data accumulated on terminals constitute the unique value of that intelligent agent. This locking effect is difficult to achieve in the cloud.

2. Operating Systems Are Becoming the New Battleground

The logic of terminal competition is clear, but who can define the AI experience on terminals remains an open question.

Looking at a series of events that occurred intensively in July, we believe operating systems are the current key battleground:

The Ministry of Industry and Information Technology (MIIT) released test results for the national standard "Grading of Intelligence for Artificial Intelligence Terminals" for the first time, with 11 smartphone models from Huawei, Xiaomi, OPPO, vivo, Honor, and others achieving L3 assistance level. The Cyberspace Administration of China (CAC) announced filing information for seven smartphone-side generative AI services, including Apple, Huawei, Samsung, Honor, and Nubia.

At WAIC, StepOn AI released Step AOS, an agent-native operating system. Honor announced its long-term partnership with Alibaba to jointly develop the next-generation terminal operating system Agentic OS. Nubia, in collaboration with ByteDance's Doubao, showcased the second-generation AI agent smartphone NaviX Ultra.

These signals collectively reveal factions in the AI terminal competition, with the core question for major enterprises being "who can define the operating system for the agent era."

Why operating systems? Because in the traditional smartphone era, the value of operating systems lay in defining application distribution rules. iOS and Android packaged apps for download via app stores, with apps isolated from each other and the system layer handling resource scheduling and security. This architecture worked well for over a decade, but it was built on a fundamental assumption—that the phone had only one active agent: the human user.

When AI agents become the second active agent, they need system permissions to call data and services across apps, remain resident to understand user behavior and preferences, and run continuously in the background to proactively provide services at the right moments. Traditional operating systems were not designed with a "second agent" in mind regarding permission systems and scheduling logic. This has become the fundamental motivation for major manufacturers to intensively reconstruct operating systems this year:

StepOn AI's Step AOS elevates agents to a system-level identity close to users, uniformly scheduling CPU, GPU, and NPU, establishing cross-app semantic memory, and decomposing system capabilities into atomic services callable by agents. Honor's Agentic OS emphasizes a human-centric approach, enabling the system to proactively understand user intent and coordinate across devices. Huawei's HarmonyOS Agent-based reconstruction transforms the underlying architecture across communication, interaction, data, and AI dimensions.

Their paths differ, but all aim to redefine interaction and information distribution in the AI era. And it's not just hardware manufacturers sensing this trend.

Late last year, Doubao's first-generation smartphone faced restrictions from WeChat, Taobao, and some banking apps, with platforms instinctively trying to block agents.

However, by July this year, Alipay announced a system-level collaboration with StepOn AI at WAIC, while WeChat was reported to be advancing A2A protocols with multiple smartphone manufacturers. Platforms are realizing that if they continue to cling to closed ecosystems and refuse to support system-level agent protocols, their gateway status could truly be replaced in the agent era. Rather than passively defending, they are choosing to actively participate in rule-setting.

3. Potential Issues During the Concentrated Evolution Phase

From the current stage, while AI smartphones are the most promising form of intelligent terminals for widespread adoption, they are still constrained by the sluggish consumer electronics market.

IDC data shows that global smartphone shipments reached 277.5 million units in Q2 2026, down 6.7% year-on-year. Rising memory chip prices have pressured flagship phone costs, while consumer upgrade willingness remains low. The industry desperately needs a narrative that can truly drive upgrade demand. Currently, while smartphones incorporate AI agents, they are not yet true AI terminals.

One issue is how to handle the numerous non-standard scenarios and unexpected bugs in daily use. In stable demo environments, agents can smoothly cross apps to book flights, plan itineraries, and compare prices. However, in real-world scenarios, popup ads, network fluctuations, App version updates altering button positions, or users changing their minds—any minor variable can disrupt the agent's task chain.

When task processes become complex enough, will users accept a failure rate of one in every two tasks? The answer is likely no. Users have far lower tolerance for errors on smartphones than on chatbots.

The second issue remains security-related to payments and privacy. Nubia President Ni Fei once summarized AI phone capabilities as "understanding, acting, remembering, and being secure." But what does "being secure" mean?

The new filing system requires edge-side models to complete filing before entering systems and going public commercially. However, filing is just the starting point of the regulatory framework. When agents perform operations involving financial security, such as payment processing, the boundaries of user confirmation, audit integrity, and reliability of withdrawal mechanisms all require finer technical standards and industry consensus.

Finally, there's the persistent issue of business models. If agents reduce the number of Apps users open, App ad revenue from splash screens and recommendations will decline. If agents prioritize Ctrip over Meituan for hotel bookings, platform bidding logic must be redesigned. These interests are deeply entangled, and clearly cannot be resolved through a few AI hardware launches.

The industry eagerly awaits an "iPhone moment." However, it took two to three years from the iPhone's stunning debut to truly changing daily habits worldwide. Moreover, the iPhone later defined an entire Apple ecosystem. No player in today's AI field wields such influence, and we are far from the integration phase.

Going forward, the market will focus more on the speed and depth of AI penetration into real life. The endgame for the AI industry may never be a binary choice between "the cloud devours everything" or "terminals dominate everything," but rather a deeper synergy. The endpoint of technological revolution is always human life. A technology truly fulfills its mission not when it must be actively "used," but when it naturally fades into the background of daily life. For AI, this process begins by slipping into the screen in your pocket and quietly becoming part of your life.

Source: Songguo Finance