Recently, Wall Street has been playing around with the word ‘computing power’.
In early July, Meta was rumored to be preparing to sell its surplus AI computing power, causing its stock price to skyrocket by over 10%, while AI and technology stocks generally fell as a result.
But just a week later, Meta changed direction again. It has also been rumored to be preparing to continue increasing its computing power, with plans to increase it from 7 gigawatts to 14 gigawatts by 2027, and its stock price ultimately rose again.
Meta’s repeated horizontal jumps constantly touch the hearts of Wall Street elites. But this seemingly awkward move reflects a new ‘common sense’ of the AI era:
Computing power is equivalent to a circulating, real gold and silver resource, which can not only be used by oneself, but also sold.
In China, there is a company that not only focuses on buying and selling computing power, but also has raised a large amount of money through it.
On July 13th, Zhijing Technology announced the completion of its Series A financing. Within six months, it has successfully completed angel++round, Pre-A round, and A round financing, accumulating over 1 billion yuan.
This company was established at the end of 2023, positioning itself as an “AI Token Production Service Provider”.
To put it bluntly, it wants to open a factory specializing in the production and sale of tokens.
However, compared to the 1 billion yuan financing, what’s really interesting about this company is that it neither trains its own large models nor produces chips, and it doesn’t even have a large-scale GPU cluster in its hands.
Even the most basic machine for producing tokens does not belong to it, but it can raise 1 billion yuan in half a year by selling tokens.
How can Trend Micro, which doesn’t even have basic production materials, sustain a token business with a financing of 1 billion yuan?
Without its own GPU, it still runs a token factory
Speaking of Token, it can be roughly understood as the “word count limit” consumed when using AI.
The large model does not directly calculate accounts based on the number of Chinese characters or words, but first divides user input and model output into individual tokens. The more you make it read and write, the longer it usually consumes more tokens, just like making a phone call requires calculating the duration of the call.
So, Token is not a digital commodity that can be produced in advance and stored in a warehouse, but a unit of measurement that measures how much information a large model processes and generates.
And the demand for this business is rapidly growing.
At the beginning of 2024, the daily average number of token calls in China was only 100 billion; By the end of 2025, this number has become 100 trillion; By March 2026, it will exceed 140 trillion yuan, with a growth rate of over 1000 times in two years.
But to produce tokens, theoretically at least three things are needed:
Computing power equipment, pre trained large models, and an inference system responsible for deployment, caching, task scheduling, and fault handling.
For example, a typical computing company would first purchase or rent hundreds or thousands of GPUs, load models onto these machines, and then process user requests through a scheduling system. Every time the model completes input and output, it generates a batch of tokens available for billing.
The trend of not making chips and not having one’s own large model is to connect the “machine” and “model”, and then manage the intermediate production process well.
In other words, it does not possess the means of production, but it masters the methods of production.
Specifically, there are two approaches to the trend.
The first option is to borrow a machine by oneself.
It rents GPUs, CPUs or domestic NPUs from outside, and then provides deployment and reasoning services around the head models such as SmartSpectrum GLM and Kimi. Finally, it sells the reasoning ability measured by Token to model manufacturers, Internet platforms, AI application companies and large enterprises.
The second approach is to not even rent the machine yourself.
In the past few years, many smart computing centers and large enterprises have purchased chips and built computer rooms, but the utilization rate of these devices is often low – the machines are there, but they don’t know how to operate efficiently, let alone where to place orders.
The trend is to turn these idle or underutilized machines into a truly profitable production line.
It helps clients design and build inference systems, responsible for model deployment, task scheduling, resource allocation, and daily operations, and then integrates these computing power into the market, turning them into services that can be sold by token, and finally sharing the revenue.
That is to say, the trend does not produce computing power, it is just a transporter of computing power.
The equipment may be rented or belong to the customer; Trending mastery refers to the production line, process, and operational capabilities, and ultimately receiving payment based on output or revenue.
So, the lack of its own large-scale GPU cluster does not prevent it from opening a token factory.
Because in this business, what truly determines output is not just the machines themselves, but how they are organized and used.
Although the factory has started operating, there is still a more critical issue:
Everyone is using GPUs, why do customers have to pay another sum of money for Zhijing?
The answer lies in the two words’ efficiency ‘.
Same card, token production capacity increased threefold
If we treat a GPU as a machine, what Trend Micro wants to do is to make it produce more tokens in the same amount of time.
It sounds simple, but the difficulty lies in the fact that big model inference is far less direct than simply plugging in a graphics card and clicking to run.
Every time the model answers a question, it must look back at the previous conversation content. The longer the conversation, the more information needs to be processed; If these contents are recalculated every time, the GPU will keep repeating its work.
If the model size is a little larger, the video memory on one card cannot fit it, and the task needs to be split across multiple cards and servers. As long as the network is slightly slow or the task allocation is unreasonable, a valuable GPU may be waiting for other machines to submit assignments.
The technology of trendsetting mainly involves hands-on experience in these areas.
The first thing is to count a few times less.
It calls this system ‘mooncake’.
Simply put, it is to save the processed context of the model through KV Cache. When users continue to ask questions, the model can directly read the previous results, reducing the number of calculations from scratch.
This is a bit like allowing you to flip through the just calculated draft during an exam, without having to redo every question.
The second thing is not to let the GPU do everything.
The “six in one” heterogeneous inference system of the trend will split the computation to CPU, GPU, and domestic NPU according to the characteristics of the task.
Expensive GPUs focus on processing their strongest parts, while other tasks are delegated to lower cost or more compatible devices.
The purpose of doing this is only one:
Reduce GPU waiting, idling, and repetitive tasks.
The reason why Zhijing was able to come up with such technology is related to the research direction of the company’s founding team in the past.
The company was born in the Institute of High Performance Computing of Tsinghua University. Its CEO, Ai Zhiyuan, is a doctor of computer science of Tsinghua University. Zheng Weimin, an academician of the CAE Member, serves as the chief scientific adviser, and Wu Yongwei, a professor of Tsinghua University, serves as the chief scientist.
The KTransformers open-source project developed by Trend Micro and the Tsinghua team focuses on how to utilize different devices such as CPUs and GPUs to enable large models to run at lower hardware costs.
The number of stars for this project on GitHub has exceeded 17000.
The core members of the trend group have also participated in large-scale inference projects such as Mooncake in the past, with a research focus on caching and distributed inference.
These technology names may seem complex and ultimately all end up on the same outcome:
How many tokens can a card produce.
According to the data disclosed by Trend Micro, the cache hit rate of “Mooncake” can reach up to 90%; Liuhe “can reduce the operating costs of a 10000 card intelligent computing cluster by more than 20%.
Since the Spring Festival of 2026, the average token production efficiency per unit of computing power in the company has increased by more than three times.
With the expansion of computing power and business volume, its total token output has increased by more than 30 times; One of the trillion parameter level models has already surpassed the trillion level in daily token production.
We need to distinguish between two numbers here.
The increase of 30 times is the total token production, which also includes the growth of computing power scale and order volume; The production efficiency of a single computing power is increased by more than three times.
However, these efficiency data currently mainly come from trend disclosure, and the complete testing environment and control baseline have not been made public.
But for the trend, the logic of this business is not complicated.
Renting a batch of computing power requires paying relatively fixed rent, electricity bills, and maintenance costs. If the same card can handle more tokens after optimization, the cost evenly distributed among each token will decrease.
What is saved is the computational cost of real gold and silver.
And this price difference is precisely the space where Trend Micro, as a software and operator service provider, can collect money from it.
Just saving space on the books, has the trend actually made money?
Can selling tokens make money
According to the data disclosed by the company, this business has already started generating revenue.
By June 2026, the monthly revenue of Trend Micro has exceeded that of the entire year of 2025; Some mature businesses have also crossed the cost line.
This number sounds very strong, but it leaves a lot of blank space.
Trend Micro has not disclosed specific revenue for 2025, nor has it disclosed the current gross profit margin.
The outside world still doesn’t know how much it will cost to rent computing power, which customers will continue to place orders, and whether the monthly revenue exceeds last year’s total. Is this year’s growth fast enough or last year’s revenue base still very low.
More importantly, improving the efficiency of a card only solves the first problem of making money.
The biggest fear for token factories is that the machines have no work to do.
There is a similarity between big model inference and traditional factories: once the equipment is turned on, rental, electricity, and maintenance costs begin to arise. No matter how fast you spit out a card, if there are no orders for half a day, the account still cannot be calculated.
The judgment of the trend is that the vast majority of token demand in the market is already concentrated on a few top models. Instead of adapting dozens or hundreds of models at once and spreading R&D resources thinly, it is better to focus on serving a few top models and high-value enterprise scenarios.
It summarizes this route as’ fewer models, deep optimization ‘.
A smaller number of models allows for deeper caching, segmentation, scheduling, and fault recovery for each model, making it easier to control research and adaptation costs.
At the same time, providing services around head models with larger demand also provides more opportunities to undertake stable scale orders, such as intelligent computing centers in multiple regions.
However, behind the seemingly lucrative business, more and more difficulties are beginning to entangle the trend.
Meta’s intention to sell surplus computing power is a signal.
Alibaba Cloud, telecom operators, and big model vendors are all continuously reducing inference costs and have the ability to directly connect models, computing power, and customers together. The constantly maturing open-source inference framework is also lowering the threshold for customers to complete deployment and optimization on their own.
The trend needs to prove that it’s not just about being able to produce more tokens with one card today.
It also needs to prove that this efficiency advantage can be maintained in the long run; Compared to customers doing it themselves or directly purchasing cloud services from large companies, its solution is still more cost-effective.
At the end of the day, there is not much suspense about the continued growth of Token demand; The question that still needs to be answered is whether every token sold by oneself can ultimately leave a profit.
After all, obtaining 1 billion yuan in financing and earning 1 billion yuan are two completely different businesses.
相关文章
