On the evening of August 26, Zhīpǔ officially announced its possession of the "Niú Lái" model.
On the evening of August 31, Zhipu released its first semi-annual report since listing in January this year.
The next day, the key words that flooded my social media feed were mainly two: revenue soared 400% in the first half of the year, and MaaS's August ARR was $1.6 billion. To be honest, both numbers are quite impressive.
The following views are problematic:
A 400% surge in revenue is actually not that important, it's a highlight, but it shouldn't be the focus.
ARR stands at $1.6 billion—though the metric is something of an industry-wide vanity figure, extrapolated from August numbers to project the next 12 months.
The arrival of a turning point in profitability has not actually occurred, as both operating losses and adjusted net losses have not shown a true turning point.
The business model has not yet been proven, it's too early to say that. It hasn't reached that stage yet.
The narrowing loss is actually an illusion created by the numbers on the balance sheet.
Disclaimer: This article is based on the author's personal experience of regularly reviewing internet companies' financial reports and does not constitute any investment advice.
In the past, when discussing Chinese model startups, it all boiled down to competing for a high score, ranking on a leaderboard, and seeing if they could make it into the top tier - otherwise, it would be embarrassing to even join the conversation.
Getting a seat at the table used to be a make-or-break proposition. If the model wasn't up to snuff, that was the end of the line.
But by now, Zhī Pǔ has gone public, and GLM has undergone six iterations in just eleven months. The latest GLM-5.3-Flash has even anonymously topped the call volume chart on OpenRouter - it's worth noting that this is the call volume chart, not the benchmark chart. Continuing to focus solely on whether the model can perform is somewhat narrow-minded, as the ability to fight has transformed from the ultimate question to a mere entry qualification.
The real new issue is: how a company makes money after a model is already good enough.
For now, this isn't a problem for Alibaba, Tencent, or ByteDance. Alibaba builds models—if the API doesn't make money, it can sell cloud. Tencent builds models, and both advertising and gaming stand to benefit. As for ByteDance, I've always believed it has a business flywheel. Just plug the models directly into its vast content product ecosystem, and the flywheel spins on its own.
Zhichu is different, it has no e-commerce, no advertising, no games, and no large public cloud business to underpin its models.
Thus, companies like Zhīpǔ and MiniMax have become ideal examples for researching whether developing large models is a viable business in China.
First, it's necessary to understand how Zhimu used to make money.
Turnkey.
Previously, the main source of income for ZhiPu was from private deployment, customized development, and system integration for enterprise and government clients, which involved completing a project, conducting acceptance tests, and then receiving payment, a model known as a "turnkey project".
Turnkey projects are also real gold and silver, and local deployment is not fake revenue - but it's still necessary to clarify that local deployment is not the same as a local model.
However, the turnkey project has its own problems. Firstly, growth may be limited because each client's needs are not consistent. Secondly, a more acute issue arises: are clients paying for the GLM model itself, or primarily for a suite of AI solutions?
The company's scalability is in question.
In 2025, localized deployment-based revenue accounted for 73.7% of ZhiPu's total revenue. More than 70% of so-called AI revenue is negotiated through human efforts in conference rooms.
The biggest highlight of this semi-annual report is that the proportion mentioned above has completely reversed.
In the first half of the year, Zhi Pu's total revenue was 954 million yuan, up 399.7% year-over-year, even surpassing last year's full-year revenue - but as I mentioned at the beginning, it's not important. What's important is that the revenue from the MaaS open platform and API business was 825 million yuan (this is not ARR, but confirmed revenue), up 2736% year-over-year, with the proportion increasing from 15.2% last year to 86.5%.
(There should be a graph illustrating the change in revenue structure here. I'm not good at math, but the jump from 15.2% to 86.5% is obvious even without a graph.)
Secretary of the Board Xiao Lei's statement at the performance meeting is very important, crucial, and carries significant weight: compared to revenue growth rate, the shift in revenue composition is more worthy of attention.
A 400% increase isn't that impressive. Any percentage looks good when the base is low. Zhicheng's estimated annual revenue for 2023 is just over 100 million yuan, 312 million yuan for 2024, 724 million yuan for 2025, and 954 million yuan for the first half of 2026 - from this perspective, a 400% increase is to be expected and not worth making a big fuss about.
The essence of the revenue-structure shift is a shortened business chain, which is critical to profitability.
The previous chain was: model –> sales –> project –> customization –> deployment –> acceptance –> revenue –> accounts receivable.
The current chain is: model –> API call –> token –> revenue.
This is very healthy, very standard, and easy to replicate on a large scale.
By the way, around 85% of Company A's revenue also comes from APIs, so this path is not unique to ZhiPu, but ZhiPu seems to be on the right track for now.
Regarding the call volume of GLM, there's one thing we should understand as observers: in the AI industry, there are many ways to manufacture impressive call volumes, such as lowering prices, providing subsidies, offering free quotas, and burning GPU resources to attract users, making the numbers look very appealing.
So, the growth of Token should be viewed in conjunction with three dimensions of data, is it desperately trying to make its point clear?
The first metric—volume—is truly striking: MaaS platform Token call volume has grown more than 40-fold since the start of the year, paid daily active users are up 603%, and daily call volume among the top ten customers has grown 98-fold.
The second digit "price" - the average selling price of APIs did not get caught up in a price war, but instead increased by about 101%. The financial report figures show that revenue growth was not achieved through price cuts.
The third digit is "cost". The cost per token for inference has decreased by 80% compared to the beginning of the year, and the revenue corresponding to the unit computing power investment has increased by 14 times.
Thus, the gross margin of Zhiku's open platform and API business increased from -0.4% in the same period last year to 24.6%.
This is even more important than its 954 million yuan in revenue, as its model invocation has started to generate gross profit - of course, DeepSeek is even more impressive, with The Information revealing that the gross margin for its API business has exceeded 80%.
So 24.6% is much more important than 400%.
By the way, Zhimu's overall gross margin has declined, dropping from 50% to 24.6% now. The reason is simple: the revenue structure has changed. Previously, the gross margin for turnkey projects was high, but now it has shifted to selling services.
The above is the good news, now let's move on to the not-so-good news (but not bad news yet).
Revenue was 950 million yuan. Gross profit was 250 million yuan. Research and development expenses were 1.33 billion yuan.
That is, the profit margin multiplied by eight is barely enough for research and development. Once this ratio is laid out, the not-so-good news is clear. It's equivalent to a student earning a meager 2,500 yuan from a part-time job, but facing a tuition bill of 21,000 yuan for the next academic year.
In fact, Zhimu did not spend heavily on administration, sales, and marketing, with sales expenses decreasing by 15% year-over-year and administrative expenses dropping by 44%, which implies a certain degree of cost reduction and efficiency enhancement. Meanwhile, research and development expenses grew by 33.6%.
Revenue increased fivefold, with operating leverage becoming apparent.
What is truly weighing on the company, however, is the economic structure of the foundational model industry itself: in order to maintain the competitiveness of GLM, it is necessary to continue training the next-generation model, and the bill for the next-generation model is denominated in tens of billions.
Can the money earned from selling tokens support the development of the next-generation GLM?
This question was not answered by the financial report.
More bad news, or so-called narrowing losses.
The surface numbers are: a net loss of 207.2 million yuan for the period, narrowing by 12.1% year-over-year, which looks pretty good.
However, after adjusting, the net loss was 1.964 billion yuan, compared to 1.752 billion yuan in the same period last year, indicating that ZhiPu's operating losses are actually still expanding.
The net loss narrowed mainly because the previous year's financial statement included around 600 million yuan in non-cash items related to financial instruments prior to listing, whereas this period the difference was only just over 100 million yuan.
To be frank, the narrowing of losses is largely due to accounting items, rather than actual business performance.
Zhi Pu has not yet reached a profitability turning point.
Finally, regarding that $1.6 billion.
Management disclosed that as of the end of August, the annualized revenue run rate (ARR) was approximately $1.6 billion on a monthly basis. If calculated on a weekly basis, it could exceed $2 billion. The aggressiveness of this metric can be assessed by yourselves.
It has an annualized value of $1.6 billion and a scale of over 10 billion yuan, but its actual confirmed revenue in the first half of the year was less than 1 billion yuan, a difference of one order of magnitude.
I have always been skeptical of the ARR in the AI industry, which has become overly inflated. In the US, what was once just a reference point, TAM (total addressable market), is now being touted as a key metric. Company A recently announced a TAM of $3 trillion, claiming it as their potential revenue.
In the SaaS industry, ARR refers to the already signed, highly recurring subscription revenue, where once a customer signs a contract, the revenue for the next year is essentially locked in.
API business revenue is highly dependent on whether customers will continue to call the API the following month and how much they will use it. The volume of API calls can change drastically overnight if the contract changes or the model is replaced, even if it's replaced with a model from another company.
So the $1.6 billion refers to the current monthly revenue run rate multiplied by 12, and not the already locked-in contract revenue for the next year.
These two things have completely different values. If we look at the ARR of APIs using the traditional SaaS ARR model, to put it politely, it would be a loss.
To put it more broadly.
Alibaba develops models, but API offerings are insufficient, so it can sell cloud services. Tencent develops models, which can benefit its advertising, gaming, and WeChat ecosystem. ByteDance embeds models into its product system. On the Google side, its search, advertising, and cloud businesses have a strong foundation.
Zhipu is different. Its business could be described charitably as focused, or less charitably as thin: can a model itself directly become a business big enough to stand on its own?
GLM's benchmark scores are set to climb a few more points next month, and the Niu Lai model is genuinely impressive—certainly worth watching. But what matters more is how many links in that economic chain can actually hold together.
Model capabilities –> Usage –> Revenue –> Gross profit –> R&D expenses coverage –> Free cash flow.
It is currently at the stage of transitioning from revenue to gross profit, the most difficult part, where gross profit covers the development of cutting-edge models, and still has a long way to go.
This financial report has, for the first time, provided an answer to the question of whether independent foundation model companies can be profitable, and that answer can be taken as a step forward.
But it is only one step.
The First Nitrogen-Fixing Bacteria Strain
Additionally, there's another variable I haven't elaborated on: computing power. Zhixu discussed the story of large-scale reasoning with 100,000 domestic chips in its financial report, which is directly related to the 80% reduction in token reasoning cost per unit. The credibility of this story is as important as the financial report itself, but due to insufficient information, I won't make any unfounded comments.
As for MiniMax, Zhi Pu originally had a completely different story, but after seeing its revenue structure undergo a drastic change in this financial report, there are now some similarities. However, MiniMax still has its own unique story.