Author: Peng Zhao (Founder of Zhicifang, Co-Founding Partner of Yunhe Capital) Original
On July 14, "Miaoshi," an emotional AI product under NetEase Cloud Music, officially ceased operations.
On July 15, the Interim Measures for the Management of Anthropomorphic Interactive Services of Artificial Intelligence (hereinafter referred to as the "Measures") will officially come into effect, and the agent functions of Doubao and Qwen will also go offline on the same day.
Just a few days later, on July 17, the 2026 World Artificial Intelligence Conference will open in Shanghai. According to the advance announcement by Ni Fei, Senior Vice President of ZTE Corporation: the world's first AI agent smartphone comes from Nubia. Positioned as a mass-produced flagship, it will make its public debut at the conference.
Within a short week, agents have collectively bowed out of apps, only to make a high-profile debut on smart hardware. This is not merely a shift in carriers, but a deep binding of a "physical plug-in": transforming from AI agents in a purely digital world into AIoT agents equipped with physical perception and execution capabilities.
On one hand, leading software platforms are cutting off their arms to divest the agent ecosystem; on the other hand, terminal hardware manufacturers are taking it as a core selling point and betting heavily on it. This contrast is no coincidence, but an inevitable choice under the interweaving of business logic and compliance pressure.
Behind this contrast lie two questions that strike at the soul of the industry: Why must agents leave general-purpose apps? Once moved into hardware and evolved into AIoT agents, does the "tightening spell" of anthropomorphic regulation become invalid?
This article will answer these two questions in turn.
Cutting Off the Arm to Survive: Why Are Apps Collectively Shutting Down "Open Personas"?
First, let's correct a misinterpretation spreading in the industry: this round of offline transitions is not a comprehensive retreat of agents.
A careful reading of the offline notices for Doubao and Qwen on July 4 clearly states the target: "user-built agents," which refers to the function allowing any user to customize the personality, rules, and interaction logic of AI characters. What the platforms have actually shut down is just one thing: the open production of personas.
Why did the platforms shut it down? Because there are three accounts that simply can no longer be balanced.
The first account is the computing power cost.
Volcano Engine disclosed in late June that the daily average Token usage of the Doubao large model exceeded 180 trillion, a 1,500-fold increase compared to May 2024; at the end of June, Doubao launched paid subscriptions, with tiers ranging from 68 to 500 RMB. According to LatePost, in the first half of 2026, Doubao's total daily revenue was less than 1 million RMB, while the daily average computing power cost reached tens of millions of RMB. Additionally, industry estimates indicate that the Token consumption for an agent to execute a complex task can be 5 to 10 times that of a normal conversation. The traditional rule for internet platforms is "the more users, the lower the marginal cost"; however, large model products have reversed this rule—the more active the users, the heavier the bill.
The second account is the content moderation cost.
The open production of personas is endless. Internet analyst Zhang Shule once described this type of user-generated content as being able to "spread like weeds." In the first phase of Shanghai's "Qinglang" (Clear and Bright) special campaign, a cumulative total of over 14,000 non-compliant agents were removed. After the implementation of the Measures, every openly produced persona corresponds to a solid obligation for content moderation, algorithm filing, and minor protection. This means that the platform's compliance obligations expand linearly with the number of personas, but revenue does not grow accordingly.
The third account is the permission cost.
Agents parasitic in the app form cannot obtain system-level permissions; they can only operate other applications through "screen reading + simulated clicks." This path is both fragile and dangerous, highly susceptible to system interception or failure due to UI updates.
Combining these three accounts, the conclusion emerges: Apps are essentially traffic containers, and traffic containers cannot sustain agents.
This is a brand-new species that burns money with every operation and accumulates compliance obligations with every persona; it urgently needs a completely new habitat.
Compliance Diversion: The "Billing Container" Where Emotion Goes Left and Tasks Go Right
This great migration ultimately led to two completely different diversion paths, and the basis for this diversion is written in Article 2 of the Measures.
This regulation focuses on supervising "continuous emotional interactive services that simulate the personality characteristics, thinking patterns, and communication styles of natural persons," but it also provides explicit exemptions: "Those providing services such as intelligent customer service, knowledge Q&A, work assistants, learning and education, and scientific research, which do not involve continuous emotional interaction, are not subject to these Measures."
Consequently, emotional agents moved towards licensed vertical containers. For example, ByteDance migrated related capabilities to its Maoxiang app, where a strict moderation and anti-addiction system was built independently. Task-oriented agents, on the other hand, moved towards hardware, officially kicking off the evolution path of AIoT agents.
This directly answers the first question at the beginning: Why do agent smartphones dare to make a high-profile debut? Because all their selling points—booking tickets, comparing prices, placing orders, and cross-app operations—fall perfectly within the exemption list. It does not meet the anthropomorphic regulatory requirements, but exits this regulatory category by relying on "service types."
More importantly, as a natural physical plug-in for AIoT agents, hardware provides task agents with five things that apps cannot:
1. Billing. At the end of 2025, ByteDance Doubao and ZTE jointly released the engineering prototype Nubia M153, priced at 3,499 RMB. The first batch of about 30,000 units sold out immediately upon release, with second-hand platforms even seeing a premium of hundreds of RMB. Users might be unwilling to pay monthly for virtual conversations, but they are extremely willing to pay a one-time premium for a piece of hardware that "can do work for them." The 2025 AI Toy Industry Report by LeadLeo Research Institute also confirms this: LOVOT has a monthly fee of 880 RMB, and BubblePal has a repurchase rate of 40%. The "hardware + subscription" billing structure has long been proven in other categories.
2. Permissions. Permission delegation is completed at once at the system level. The M153 adopts hierarchical permission management, mandating that sensitive operations be handed over to the user for manual confirmation. Authorization changes from tedious bargaining in every conversation to a one-time contract upon purchasing the device.
3. Identity. Regulation needs an address to knock on. Millions of user-built personas are unaccountable, but a hardware product has a name, a filing entity, and a recall channel.
4. Memory. Edge-side storage allows long-term memory to compliantly remain on local devices without needing to upload everything back to the cloud, significantly reducing privacy risks.
5. Physical Interaction (The Soul of AIoT). Pure software agents can only live inside screens, whereas AIoT agents perceive the world through sensors and change the physical environment through hardware control. Hardware is no longer just a shell, but the senses and limbs of the agent extending into the real world.
The "billing container" catches two and a half of the three accounts that the traffic container cannot resolve. The remaining half account tears open the irreconcilable contradiction between AI hardware and internet giants.
At the end of 2025, leading apps such as WeChat, Taobao, and Meituan successively imposed restrictions on the Doubao AI smartphone. Although hardware controls the system, it cannot truly control the ecosystem. Because in the eyes of the giants, Doubao's unconventional approach based on GUI agents is not innovation at all, but a sneak attack without a declaration of war. It is like a kind of "digital parasitism," bypassing high walls and relying on simulated clicks to freeride services. This kind of "traffic hijacking," which attempts to cross interface barriers and brutally infringe upon the platform's data sovereignty, is destined to face the most resolute resistance from super apps.
Therefore, what the industry should truly advocate is by no means this GUI-based "digital parasitism," but rather "ecosystem symbiosis" based on underlying protocols and API authorization. AI agents should not act as barbarians breaking down doors, but should establish positive interface calls with super apps under the premise of respecting the platform's data sovereignty. Only by transforming "zero-sum traffic hijacking" into "incremental value co-creation" can this remaining half account be truly settled.
Macro data also stands on the side of hardware. An official from the National Development and Reform Commission revealed at a press conference on July 7 that last year, the shipment volume of smart terminals such as AI smartphones and AI PCs exceeded 100 million units, and this year it is expected to surpass non-AI products for the first time.
The Embodiment Trap: Hardware is Absolutely Not a "Safe Haven" for Anthropomorphic Regulation
The story seems to have a perfect ending here: emotional agents go to licensed apps, and task agents move into hardware to embody AIoT agents, each finding its proper place. But the second question remains unresolved: emotional agents can also be moved into hardware; AI dolls, companion robots, and speakers with personas are all ready-made physical plug-ins. Once moved into hardware, can the anthropomorphic regulatory requirements be bypassed?
The answer is exactly the opposite. The Measures regulate "services," and the virtual or physical nature of the carrier does not change the nature of the service; on the contrary, the deep integration of the "embodied form" with AIoT will exponentially amplify compliance issues.
First, embodiment amplifies dependency.
The physical touch, voice, and sense of presence make the emotional connection between the user and the device far stronger than a cold chat window. This point exactly hits the legislative core of the regulation: preventing dependency, preventing addiction, and preventing emotional manipulation. Moving into hardware not only does not constitute evasion, but actually constitutes an increase in risk.
Second, there is a severe mismatch in the compliance context.
Most of the execution mechanisms of the Measures are designed on the premise of "personal account + screen interface": prominent identification of AI identity, identification and mode switching of suspected minors, remote control on the guardian's end, and reminders for continuous use exceeding 2 hours. However, the usage scenarios for AIoT hardware are often "family-shared devices + screenless environment interactions." How can an AI doll in the living room verify the age of every user? How can a screenless plush toy prominently identify itself as AI in some way? Where is the guardian's end installed?
Third, lifecycle responsibilities are heavier.
When Miaoshi ceases operations, users just lose an app; but if an AI toy manufacturer shuts down cloud services, the dolls in children's hands will collectively lose their voices. Beyond the emotional break, there is the superimposition of tangible property rights losses. Additionally, there is an unanswered black hole of rights and responsibilities: the model comes from a large company, and the hardware comes from another startup. Who is the "service provider" defined by the Measures? Who files the record, who conducts the security assessment, and who takes responsibility when emotional dependency damage occurs?
Thus, it can be seen that compliance requirements such as filing, security assessment, annual audit, and minor protection systems are all fixed costs that do not scale with the selling price. A 3,499 RMB AI smartphone might be able to afford it, but a 199 RMB AI doll is absolutely unable to bear it. If the first wave of regulation on July 15 mainly hit apps, then the second wave will inevitably precisely hit the "warmth hardware" focusing on emotional companionship.
Faced with high compliance costs, the future smart hardware industry will inevitably diverge into two paths:
First is "tool-based degradation," proactively cutting off anthropomorphic personas, long-term memory, and proactive care, and retreating to the safe regulatory exemption zone;
Second is "licensed upgrading," transforming compliance into a moat and migrating towards a high-ticket subscription model.
However, the realistic challenge lies in the fact that there is no clear-cut wall between a "caring task assistant" and "continuous emotional interaction," only a red line that constantly moves with the scale of regulation. To cope with this ambiguity, it is foreseeable that a brand-new product design logic will become the standard for future terminals, namely the "physical separation of tool mode and companionship mode."
Under this design, the tool mode stays in the exemption zone, enabled by default, free to use, and stripped of emotional memory; the companionship mode, as a value-added service, is filed separately, has a separate switch, undergoes separate age verification, and adopts subscription billing. This "dual-mode separation" will evolve from a mere passive compliance action into a new variable that reshapes the definition of smart hardware products.
Following this compliance boundary, we can precisely re-examine a classic proposition in the IoT (Internet of Things) industry.
Over the past decade, one of the selling points of IoT products has been "warmth," with companies dedicated to making cold devices understand and care for users. But when IoT fully evolves into AIoT agents, the rules of the game have changed completely: "warmth" is no longer just a glib marketing term, but has officially become a business scope that requires "working with a license."
To sell "warmth," one must bear the matching regulatory costs. This is not only the evolution of AI technology but also a reconstruction of the underlying logic of the entire smart hardware industry.
Final Thoughts
Returning to the timeline at the beginning, there may be three layers to the upcoming transformation:
1. Diversion Judgment: Emotional agents fully enter the licensed era, and AIoT agents (task-oriented) enter the competitive era of physical plug-ins. The variable determining the life and death of AI applications has shifted from mere "model capability" to "billing containers" and "physical execution power."
2. Second Wave Prediction: Before the end of 2026, the first filing announcement or regulatory interview targeting anthropomorphic services in hardware form may appear, and the AI companion toy category may experience the first round of industry consolidation.
3. Boundary Prediction: Within the next 12 months, the market may see a batch of mass-produced AI terminals with separated designs for "tool mode" and "companionship mode."
As the tide recedes, the reefs of compliance have been revealed. The long journey of AIoT agents has only just taken its first step.