Author: Peng Zhao
In the first week of November, Tokens saw a wave of widespread attention:
On the evening of November 2, Li Jiaqi's livestream listed an invisible product—a monthly membership card for Qwen Office. Priced at CNY 0.01 during the stream, nearly 100,000 units were sold within half an hour, equivalent to 2 trillion Tokens. A Token is the smallest unit for large models to read and write information; depending on the model, one Token roughly corresponds to 1.5 Chinese characters.
On November 5, Intel showcased its "Token Integrated Machine" and containerized green power energy storage compute pods at the China International Import Expo (CIIE), marking the beginning of naming computing devices after Tokens.
On November 6, three significant events occurred in the industry: First, China Mobile launched its Intelligent Network. Chairman Chen Zhongyue summarized the division of labor among the three networks at the Mobile Network Conference as: "The communication network ensures efficient byte transmission, the compute network ensures efficient token flow, and the intelligent network ensures efficient intent execution." Immediately after, China Mobile unveiled its unified IoT brand, "Mobile BiLin" (Beyond Link). On the same day, the Ministry of Industry and Information Technology (MIIT) issued the "Notice on Strengthening the Management of Smart Agent Terminal Equipment Entering the Network."
Previously, in my article titled "The 'Token Economy' of the Physical World: How IoT Can Cross the Three Steps Beyond the 'Hundred Billion Connection Trap'," I explored this concept, focusing on how electricity meter readings could be converted into settleable vouchers.
In a blink, compute Tokens have been put on shelves like mobile phone bills.
However, buying the quota is one thing; achieving results with it is another.
In July this year, George Sivulka, founder of US tech company Hebbia, wrote for a16z that just as eight out of ten employees are often busy without purpose in the workplace, eighty percent of enterprise Tokens are effectively idling. Of course, this remains a qualitative judgment that has not been strictly verified.
In contrast, survey data from software company Harness involving 700 engineering leads is more robust: According to self-reported estimates from respondent companies, approximately 26% of AI spending was wasted.
Why does compute consumption appear vigorous while actual business progress often stalls?
To unravel this mystery, let us first examine how Tokens are packaged into plans, traverse carrier networks, and ultimately land on physical world business sites.
Packaging Tokens into Plans: From One Card to One Network
For compute to be billed like phone calls, we must start with a single card.
In May this year, China Mobile launched AI-eSIM, accompanied by the "139" system (1 chip entry point; 3 major engines comprising the AIoT platform, agent services, and Byte+Token integrated operations; 9 types of scenarios).
Compared to standard eSIMs, AI-eSIM accomplishes two additional tasks:
First, it endows terminals with financial-grade hardware security identities;
Second, it realizes "Number-as-Large-Model-Account," using the device's IoT card number directly as the access account for calling cloud-based large models.
In the past, developing smart hardware required enterprises to interface separately with three systems: carriers, large model vendors, and security certification providers. Now, with built-in AI-eSIM, these three complex interfaces are merged into one. Devices connect to the network upon boot-up and can directly call large models once connected.
Beyond simplifying integration, the more critical achievement is solving the challenge of continuous billing.
Smart hardware is often a one-time sale, but the inference costs incurred when devices call large models after activation continue to accumulate.
To address this, China Mobile introduced the "Byte+Token" model, packaging traffic (Byte) and compute (Token) into the same plan for combined metering and one-stop subscription.
Six months after its launch, AI-eSIM users have exceeded 3 million, covering nine major scenarios including cultural creative toys and smart wearables.
However, this card on the edge side only solves how devices connect to large models; how massive amounts of Tokens flow and are accounted for within the network depends on supporting infrastructure on the network side.
Within this system, terminal accounts clarify who is consuming compute, while China Mobile's Model Service Platform, MoMA, addresses "which model to call and how much it costs."
MoMA acts like a backend "model switchboard," connecting over 300 large models through a unified interface. It matches the most suitable model for specific tasks, reducing unit inference costs by approximately 30%, while ensuring real-time billing and full auditability.
Above MoMA lies the newly launched "Intelligent Network." It functions like a switch and router for intelligent services, responsible for understanding user intent and precisely dispatching tasks to corresponding models, data, or agents for execution.
Looking back at the development of the communications industry, carriers' fundamental strength in each era has always been grasping the core measurement unit of the time—from minutes in the voice era, to bytes (Byte) in the data era, to tokens (Token) today.
I previously discussed paths for IoT enterprises to escape becoming mere "pipes." China Mobile has clearly chosen the hardest and most thorough route: shifting from selling connectivity products to defining industry rules. "Number-as-Account" and "Byte+Token" are the new rules it has established.
But returning to the initial question, do AI-eSIM and MoMA alone suffice to end "Token idling"?
Idling has two root causes: unclear task definition and lack of result verification. Billing and scheduling solve the former—determining who receives the task, which model to use, and the cost. The latter can only be completed at the business site.
Beyond Connectivity: When Compute Tokens Meet Physical Tokens
Completing result verification at the business site requires support from the IoT foundation.
On the same day, China Mobile launched its IoT brand "Mobile BiLin" (Beyond Link). Its product line extends from chips and modules to AIoT platforms and solutions. Supporting this are 1.5 billion IoT card connections, over 300 million devices connected to the AIoT platform, and more than 1,000 standard object models accumulated over time.
The carrier named its IoT brand "Beyond Link," implying value derived from, based on, and exceeding connectivity. This clearly indicates a shift in value focus to above the connection layer.
This value above connectivity corresponds precisely to the "Physical Token" concept I proposed earlier.
A Physical Token refers to a voucher of physical output measured by accepted methods at the business site, verified by third parties, and directly usable for settlement—for example, the actual regulated electricity recorded in a virtual power plant after being sealed by measuring devices.
Although both Physical Tokens and Compute Tokens are called "Tokens," they account for two different worlds.
As shown in the table, "Result Content" (the proportion of compute Token consumption convertible into verifiable results) and "Effective Control Rate" (the proportion of connected capacity actually mobilizable according to instructions) measure idling degrees in the digital and physical worlds, respectively.
In fact, idling phenomena also exist in IoT, though they are more easily exposed against actual meter readings.
Taking virtual power plants as an example, aligning the compute and physical accounts requires bridging three links, all of which already have deployments:
First is Trusted Identity. On-site data serving as settlement vouchers must never be falsified. The hardware security identity provided by AI-eSIM can serve as the "trusted root" for on-site metering terminals.
Second is Precise Metering. Smart meter reading is the foundational base of billions of IoT connections, and AIoT platforms have already implemented projects in the virtual power plant domain.
Third is Compute-Power Coordination. China Mobile's Guangzhou and Zhanjiang data centers, along with Unicom's Shaoguan data center, participate in electricity spot trading via the Yunengtou Virtual Power Plant Operation Platform. They increase compute tasks during low electricity price periods and decrease them during peaks. This marks the first time large data centers have participated in spot trading as virtual power plants.
From Consuming Compute to Delivering Results: The IoT Watershed
Truly linking the compute account and the physical account depends on IoT enterprises rooted in business sites.
Between these two accounts lies an "exchange rate": How much verifiable physical result is exchanged for each unit of compute Token consumed?
The core value of IoT enterprises lies in continuously raising this exchange rate.
This requires overcoming three hurdles on-site:
First, providing industry context. Converting on-site conditions, equipment mechanisms, and sensor data into language directly understandable by large models. The "standard object models" mentioned earlier serve as foundational drafts.
Second, encapsulating execution actions. Tasks dispatched from the cloud must be executed via physical interfaces upon reaching the site. This involves encapsulating operations such as opening valves, adjusting pressure, and scheduling air compressors into callable standardized skills with safety guardrails.
Third, closing the loop with result verification. Directly linking each model call to actual metrics such as energy savings, yield rates, or downtime duration, using real meter readings to validate the effectiveness of compute consumption.
Industry regulatory trends are evolving simultaneously.
The MIIT notice dated November 6 requires strengthening management of smart agents regarding permission invocation and task execution, standardizing live network upgrades, and conducting adaptation verification and risk assessment during the network entry phase.
It is evident that regulatory focus is extending from cloud-based "models that talk" to on-site "terminals that act."
Any work performed in the physical world must leave a trace and undergo verification.
At this point, the division of labor in the supply chain is clear: Carriers' intelligent networks deliver task intents to the doorstep of the site; however, how to perform the work inside and how to settle accounts falls to on-site enterprises holding equipment dispatch rights, deeply familiar with industry conditions, and capable of passing meter verification.
Compute Tokens are priced by consumption, with unit prices trending downward as inference costs drop; Physical Tokens are settled by outcomes, with their premium determined directly by energy saved or yield improved.
Conclusion
From AI-eSIM to the Intelligent Network, the original intention of this entire infrastructure is to enable efficient intent execution. However, whether for execution or verification, the last mile must return to the business site.
Backend systems can accurately record how many Compute Tokens were used; but proving what was actually accomplished on-site relies on real meter readings and business outcomes.