Token consumption enters the trillion-billion-level computing power industry and begins to be re-queued using “delivery” and “billing”

Zhitongcaijing · 2d ago

The Zhitong Finance App learned that on November 18, the two pieces of news are worth reading together.

One from the app side: Tencent disclosed the three-quarter report. WorkBuddy's daily activity exceeded 10 million, average daily token consumption exceeded 100 trillion, and AI demand led to a 35% year-on-year increase in cloud service revenue.

Another one comes from the infrastructure side: the 2026 Computing Power Internet Conference first announced the National Computing Power Connectivity “1+M+N” industry node, and launched a regional node in Hubei.

One says “the model is actually being used,” and the other says “computing power is starting to be connected across subjects.” The two things point to the same judgment: when calls enter the actual workflow and computing power begins to be dispatched to each other, the competitive scale of the industry is no longer “how much has been built,” but “whether resources can be stably turned into usable services.”

The parameter list left, and consumption took the stage

The first half of the big model industry looked at the scale of parameters, ranking, and release pace. After entering office, programming, marketing, and enterprise services, call volume, active users, and corporate payments began to provide a harder dimension of observation.

Tencent's 100 trillion tokens are representative data of this change. A token is a unit of measurement for model processing information. Increased consumption indicates that the model is being executed, not just displayed or discussed.

However, a pot of cold water must be thrown in here: there is no simple conversion relationship between token consumption and the revenue of computing power companies. Model efficiency, service prices, task types, infrastructure costs, all affect the numbers that end up on the report. A tenfold increase in consumption does not mean a tenfold increase in income.

The enthusiasm on the demand side is not equal to the applause of the capital market. Nvidia's data center revenue for the third fiscal quarter of fiscal year 2027 was US$107.6 billion, up 110% year on year, and guidance for the next fiscal quarter continued to grow; however, in the US East trading on November 18, Nvidia's stock price fell 2.89%, and the Philadelphia Semiconductor Index only rose 0.10%.

Strong performance and market differentiation are occurring at the same time, indicating that in addition to demand, investors are also examining three things: delivery, financing, and quality of cash flow.

For computing power service providers, the stage of relying on industry sentiment to tell stories is over.

Computing power began to be interconnected, and scheduling capabilities were pushed to the front desk

Infrastructure is moving from a “single point of expansion” to a “resource connection.”

According to information released by China News Network and the Hubei Provincial Data Bureau on November 18, as of the end of September 2026, the scale of China's intelligent computing power was 2592 EFLOPS, an increase of 146% over the previous year; the national computing power Internet service platform has allocated 11.8 million computing power tags, gathered 362 EFLOPS, and built more than 70 computing power transmission channels.

Resource interconnection enhances the possibility of computing power flow, and at the same time pushes scheduling across regions, chips, and clouds to the front desk. Devices from different sources, the adaptation requirements of different models, and the task load of different customers need to be matched in the same service system.

As a result, computing power services are no longer equivalent to “equipment rental.” Platform scheduling, networking optimization, and operation and maintenance capabilities begin to directly determine actual output.

Guangdong-Hong Kong Bay Intelligent Computing (01396) is one of the few companies in this process that has the meaning of publicly disclosing the sample. The business covers integrated supply of multi-source heterogeneous computing power, “quantum power” computing power cloud services, AIDC construction and operation at home and abroad, and AI vertical application development and investment incubation. The company summarizes the delivery system as “three streams in one” for facility flow, equipment flow, and technology flow, which corresponds to the three components of intelligent computing center, server, and networking optimization.

The core of the model is not how long the business list is, but whether each link can form continuous delivery. Many participants in the industry only cover one part of it, and when customers shift from “buying resources” to “buying usable computing power,” the weight of end-to-end collaboration rises.

From “how much was built” to “how much was billed”

If you look at a computing power service provider, orders need to be viewed in two layers.

Tier 1: Delivered and billed steadily. This is the actual production capacity that has already been reported — as of June 30, Guangdong-Hong Kong Bay Intelligent Computing has delivered and entered a stable billing scale of over 20 billion yuan in computing power orders. This level is “the money is already being received” and is the hardest verification.

Level 2: In-process orders. Over the same period, the company's intended computing power technology service orders exceeded 37 billion yuan. This layer is the “capacity reserve for the next 3-5 years”. More than 95% is about 3-5 years long, and over 95% comes with advance payment.

The key reminder is that 37 billion was the caliber as of June 30, not the current scale. Since then, the company has continued to add new orders and continue to deliver online — in other words, these two numbers are a phased snapshot of rapid expansion rather than a ceiling.

The three operating indicators also came from the end of June: the computing power rate stabilized above 95%; more than 95% of new orders took about 3-5 years; and over 95% with advance payment.

Listing rates, long-term contracts, and advance payments cannot eliminate market risk, technical risk, and repayment risk. But they did an important thing: change the operating status of an asset-heavy company from “invisible” to “continuously trackable.” For a computing power service provider that has invested so much capital, this is more worth watching than any size figure.

After the market on the same day, Guangdong-Hong Kong Bay Intelligent Computing issued a voluntary announcement: the board of directors intends to repurchase shares with a total value of up to HK$200 million in the open market from time to time in accordance with existing authorization, with capital from internal resources.

The announcement is clearly written: as of the announcement date, the company has not repurchased any shares, and there is no guarantee on the time, quantity, price, and implementation of the repurchase.

Therefore, “up to HK$200 million” is the planned amount, not the completed repurchase amount. Subsequent actual implementation is still subject to company announcements and Hong Kong Stock Exchange disclosures.

From an industry perspective, capital movements should be viewed separately from management actions. The buyback program conveys the board's attitude towards long-term development, but it cannot replace order delivery, asset operation, and cash return.

Conclusions

After the computing power industry enters a new stage where “real consumption” and “interconnected scheduling” go hand in hand, the market will eventually return to the same set of basic problems:

Whether services are available, whether delivery is stable, and whether revenue and profits can continue to be generated.

Token consumption of 100 trillion dollars is the answer on the demand side. The convergence of 2592 EFLOPS and 362 EFLOPS is the answer on the supply side, while the 20 billion already billed scale, 95% listing rate, and 95% long term is the answer given by this company.

Can the three answers be correct? Delivery and billing data for the next few quarters will talk.