Since 2026, the API call volume for Chinese models has been climbing at an unprecedented rate. As of June this year, the average daily Token call volume for the Doubao large model has exceeded 180 trillion, a growth of over 10 times in the past year. Volume is rising, and so are prices. According to Morgan Stanley, the average API price for Chinese model outputs in the second quarter has risen to 21.9 RMB per million Tokens, a year-on-year increase of over 60%.
With both volume and prices rising, all Token vendors should logically be making money, but this is not the case. There are currently three types of players in the Token market, but their gross margin trends are completely different. According to statistics from iFenxi, only major cloud providers are in the high gross margin range, while model vendors and independent vendors are struggling with losses or hovering around the break-even point.
Figure 1: Gross Margin Distribution of Three Types of Token Vendors
iFenxi believes that whether a Token vendor can make money is not determined by technology, but by five variables. In the profit formula constructed by these five variables, whoever can establish a structural advantage will be the one to make money.
Figure 2: Token Profit Formula
Let's briefly explain the above formula. The profit of Tokens first depends on the weighted unit price of Tokens, which is the unit price of Tokens weighted by input, output, and cache prices; second is the GPU utilization rate, referring to the percentage of time GPU cards are actively utilized; third is Token throughput, referring to how many Tokens can be processed per month.
Multiplying these three variables gives the Token revenue. Then, subtract two costs: one is the monthly rent for GPU servers, and the other is the model licensing fee, as open-source models have started charging large Token service providers. Using this profit formula, the fate of the three types of players can be clearly understood.
Major cloud providers have advantages in almost all five of these variables: they can independently price their self-developed models without paying licensing fees, lock in low-cost GPU monthly rental prices through long-term agreements, and have dedicated inference optimization teams to improve Token throughput. The biggest advantage of major cloud providers is the GPU utilization rate; by providing Token services both internally and externally, the GPU utilization rate is maintained at a high level year-round.
In contrast, the advantages of model vendors lie only in pricing power and model revenue sharing. They have no advantage in the other two variables, GPU utilization rate and server monthly rent, resulting in a significant drop in the gross margin range compared to major cloud providers. The most unfortunate are the independent vendors; among these five variables, they do not have independent pricing power in any of them, are fully exposed to market prices, and have no competitive advantages, which inevitably leads to structural losses.
01. The Barrier for Major Cloud Providers Is Not Technology, But That GPUs Are Never Idle
Although the high gross margin of the Token business for major cloud providers is somewhat downplayed and they are reluctant to express it too much, it is obvious that major cloud providers are the only players in the comfort zone in the Token market. Let's look at Alibaba first. During the earnings call, Alibaba explicitly stated that the Token business has a high gross margin, and with Token prices expected to continue rising over the next 1-2 years, the gross margin will continue to increase.
Explaining the reasons behind the high gross margin, Wu Yongming spoke very straightforwardly, "There is hardly an empty card in our servers." The GPUs of major cloud providers not only take external Token orders, but also have a long queue of internal businesses such as search, advertising, recommendation, and office applications waiting to use the cards.
As soon as any card becomes idle, the next internal task immediately takes it over. Therefore, for major cloud providers like Alibaba and Volcano Engine, the GPU utilization rate easily exceeds 85%. Let's look at Volcano Engine again. Although Volcano Engine denied that the gross margin of Seedance reached 70%, this move is obviously a classic case of protesting too much, and the industry generally believes that its gross margin is at least over 50%.
The high gross margin of Volcano Engine's Token business is often attributed by the market to its excellent cost control of GPU servers. However, according to iFenxi's observations, this is not the case. Volcano Engine leases GPUs from Runze Technology, and the gross margin of Runze Technology's computing power leasing business is 53%, indicating that the rent paid by Volcano Engine is not significantly lower than the market price.
Summarizing the cases of Alibaba and Volcano Engine, the root of the high gross margin for major cloud providers lies not in the procurement end, but in the GPU utilization rate. In contrast, model vendors and independent vendors need to acquire customers order by order from the market to fill their GPUs. Their GPU utilization rate has an obvious ceiling, making it difficult to exceed 70%, resulting in a large amount of idle time. In addition to the GPU utilization rate, major cloud providers have a second layer of barriers, which is the pricing power for scarce closed-source models.
The tracks for large language models and coding models are already overcrowded, but video generation models are different; there are only a few capable players. Seedance is a scarce supply among them, and its price has remained firm. According to iFenxi's research, channel agents for Seedance can only get a 5% discount, with an extremely low discount rate. At the same time, the price of Seedance including video input has risen from 28 RMB to 42 RMB per million Tokens.
With the dual advantages of GPU utilization rate and model pricing power, the Token business of major cloud providers features high gross margins and high growth, becoming the star of hope for their AI businesses.
02. Whether a Model Can Run on Domestic Chips Determines the Profitability of the Token Business
For model vendors providing Token inference services themselves, whether they can make money comes down to one key factor: whether the model can run on domestic chips. If it can run on domestic chips, there is some profit to be made. If the model parameter scale is too large and it can only run on NVIDIA chips, especially the B300, then it will basically only lose money. In other words, at the model design stage, the profit and loss space for Token inference services is already locked in.
Kimi and Zhipu AI, two model vendors, happen to demonstrate completely different results. Kimi's K3 has 2.8 trillion parameters, and the weight file after MXFP4 quantization is as high as 1.4TB. The total VRAM capacity of 8 H200 cards is only 1.1TB, which is simply not enough, so it can only use the B300. Based on Wafer's benchmark tests, iFenxi estimated the monthly profit and loss of K3 on B300 servers. Under the premise of 100% GPU utilization rate and 90% cache hit rate, Kimi loses 65,000 RMB per month, with a gross margin of -34%. If the GPU utilization rate is estimated at the industry standard of around 70%, the losses will be even more severe.
Figure 3: Monthly Token Loss of K3
Looking at Zhipu AI's GLM-5.2 in contrast. Although the output price of GLM-5.2 is significantly lower than that of K3, it can run on Ascend servers. Running fully for a month, the profit margin can reach 20%.
The parameter count of GLM-5.2 is 744 billion, only 1/4 of K3, and the weight file after MXFP4 quantization is only over 200GB, which can completely run on Ascend servers. The monthly rent for Ascend servers is between 70,000 and 80,000 RMB, which is just a fraction of the B300's cost. Calculating for both, the result is exactly what was said at the beginning: whoever can fit their model into cheaper domestic chips can make money.
Therefore, there is no gap in technical strength between GLM-5.2 and K3, but rather different choices in technical routes. K3's choice of B300 does not mean it cannot make any profit at all. When opening model weights, K3 charges model licensing fees to partners with MaaS revenue exceeding 20 million USD. This asset-light revenue-sharing model will be the main way for K3 to make a profit.
Under this model, Kimi does not need to invest in computing power costs, and all revenue is pure profit, which is also a unique structural advantage for model vendors. iFenxi expects that in the future, Zhipu AI and DeepSeek will also consider taking a cut from large MaaS service providers while open-sourcing their models.
03. The Future of Independent Vendors Is Uncertain, and No Amount of Optimization Will Bring Profits
Independent vendors are the most difficult type among the three types of players. When selling Tokens, they have no model pricing power because they run open-source models, the market is fully competitive, and they have no price advantage. They pay the full market price for server monthly rent and lack the capital strength to sign long-term contracts to lock in prices. They do their best to achieve a 70% GPU utilization rate, which is already an excellent level in the industry, but there is a fixed gap of 10-20 percentage points compared to major cloud providers. In addition, they also have to pay licensing fees to model vendors. For these variables, independent vendors cannot call the shots on any of them; all prices are set by others.
The only card independent vendors have in their hands is inference acceleration technology. Xia Lixue, CEO of Infinigence-AI, proposed the idea of "optimization equals profit" in July 2026. The team's actual tests showed that through architecture optimization, the cost of a single Token could be reduced by 37.5%. This number sounds significant, but when put into the bill of K3 losing 65,000 RMB per month on B300, the situation changes.
Figure 4: Gross Margin Improvement Brought by Inference Acceleration
With all these optimizations fully applied, the best result is just changing the monthly loss from 65,000 RMB to a loss of 10,000 to 20,000 RMB. The direction hasn't changed; it's still a loss. The true value of optimization is not to make independent vendors profitable, but to help them lose a little less and survive a bit longer. iFenxi judges that the larger the scale independent vendors build, the more they lose; this is a business where the numbers simply don't add up.
Moreover, the survival space for independent vendors is narrowing from both directions. On the one hand, the parameter scale of open-source models continues to expand, and fewer and fewer models can fit into domestic chips. On the other hand, the monthly rent for NVIDIA GPUs is rising rapidly, with the B300 increasing from 120,000 RMB at the beginning of the year to over 190,000 RMB in July.
If independent vendors insist on only doing inference acceleration optimization, the only profitable path is to use idle computing power in the market. Holders of idle computing power, represented by local government and state-owned enterprise computing centers, will be important partners for independent vendors. For these government and enterprise computing centers, GPU procurement and deployment have already been completed, making GPU servers sunk costs, where every cent of Token revenue is important. Due to a lack of usage scenarios, the computing power remains idle all the time, and the cost for independent vendors to access this computing power is extremely low.
For the same B300 servers, independent vendors pay a monthly rent of 190,000 RMB at market prices, while the cost of accessing idle computing power from government and enterprises is even cheaper than Ascend servers. Running the same K3 model, it is entirely possible to achieve a positive gross margin. However, the window period for this path is limited, and the amortization cycle for idle GPUs in government and enterprises is finite. In another two years, these GPUs will be eliminated by the market and cannot be used for Token inference services.
04. Telecom Operators Are About to Enter the Market, Making Competition Even Fiercer
The current Token market is dominated by three types of players, but a fourth role is in the making. Telecom operators will join the market competition with a wholesale-to-retail distribution model, centrally procuring Tokens from model vendors and then retailing them to government and enterprise customers. Among them, China Telecom has launched a 10-billion-level Token factory centralized procurement, and China Mobile has launched general Token services. For telecom operators, this business is absolutely in their comfort zone. Holding two core sales channels for consumers and government/enterprise clients, turning Tokens into a product on par with voice and data packages is just a matter of time.
Moreover, resource resale businesses like wholesale-to-retail are inherently what telecom operators are best at. Therefore, the entry of telecom operators will squeeze the space of some market players. Players lacking market bargaining power and competitive advantages, especially independent vendors, will face greater tests of survival pressure.
Ultimately, every sustainably profitable model for Token vendors is essentially tied to a cost structure that others cannot replicate. Major cloud providers are tied to the deterministic GPU utilization rate brought by their internal business ecosystems, model vendors are tied to their proprietary models and matching chip choices, and telecom operators are tied to their distributor roles. Only independent vendors have yet to find their own exclusive advantages.