MaaS Has Arrived: Anthropic Turns an Operating Profit for the Quarter, China's Public Cloud MaaS Revenue Reaches 10.63 Billion Yuan in Half a Year

Can models actually make money? Recently, there have finally been a few clear bills.

In the prospectus made public on October 26, Anthropic disclosed that its revenue in the second quarter of this year was $11.555 billion, and it achieved an operating profit of $974 million for the first time. Moreover, this profit has already deducted the operating expenses recognized in the current period, including R&D, personnel, sales, and compute.

What has supported this company to get here is the business of selling intelligence by usage. In the first half of this year, about 80% of Anthropic's revenue came from usage-based billing, with customers purchasing model capabilities through the API and Claude's own products. It calls this model "paying for intelligence."

A few days ago, when DeepSeek open-sourced its inference framework DeepInfer, it also provided a very interesting set of numbers: a single cluster serves a throughput of 155 trillion tokens per day; calculated at API off-peak prices, across the entire cluster per card, the revenue is $38 per hour. Previously, The Information also reported that DeepSeek's API business gross margin reached 82.9% in the first seven months of this year.

For a long time, people have been accustomed to treating models as a technology that is still burning money and waiting to find a business model. Now, some have obtained sustained revenue by selling model capabilities, some have achieved gross profit in inference services, and some have for the first time used these revenues to cover current operating costs.

The commercialization of foundation models has finally found a path that has been proven first. These models, called by different names, have a unified name: MaaS.

MaaS Is Fierce

MaaS means "Model as a Service." Service providers handle model deployment, inference optimization, and compute scheduling. Enterprises connect models into their own products and businesses through interfaces and supporting tools, and then pay based on call volume and other methods. Users can start from a specific link, verify the results, and then expand usage as the business grows.

This purchasing method, which originally still needed explanation, has already formed a tens-of-billions market in China.

The latest report released by IDC shows that in the first half of 2026, revenue in China's public cloud MaaS market reached 10.63 billion yuan, a year-on-year increase of 889.4%. In the first half of last year, this market was only about 1.08 billion yuan, and in the second half about 2 billion yuan. In half a year this year, it has already exceeded three times the previous full year.

In China's public cloud PaaS market, MaaS's revenue share also rose from 5.2% in the second half of last year to 21.0% in the first half of this year.

More notably, those crowding into this business are already far more than cloud vendors.

In IDC's overall revenue ranking, Zhipu holds a 5.0% share and enters the top three. Model development companies have begun to directly participate in the distribution of the enterprise services market. The chips they hold are the models they have developed themselves and that customers are willing to pay to use.

Zhipu's semi-annual report presents this change more concretely. In the first half of this year, the company's total revenue was about 954 million yuan, a year-on-year increase of 399.7%; among it, revenue from the open platform and API reached 825 million yuan, a year-on-year increase of 2735.7%, accounting for 86.5% of total revenue. In the same period a year earlier, this part of the business accounted for only 15.2%.

The company's revenue focus has shifted from localized deployment to continuously providing model services.

This is directly related to the capability evolution of GLM over the past period. Zhipu introduced in its semi-annual report that the model has advanced from programming recommendations to code file generation and engineering project execution. Longer tasks allow users to make continuous calls and also give a single use the opportunity to correspond to higher work value. At the same time, the company disclosed that the inference cost per token has dropped by 80% compared with the beginning of the year, giving larger-scale services a cost basis.

Improved capabilities open up tasks that people are willing to pay for; lower inference costs allow these tasks to be handed to the model more frequently. The two things have begun to meet in revenue.

DeepSeek presents another path. Low prices and low inference costs make it suitable for a large number of programming, tool-calling, and Agent tasks, and the permissive MIT license allows more service providers to deploy and sell its model. Users can purchase directly from DeepSeek or from other platforms that provide inference services.

These examples show that MaaS is accommodating different model positioning. Lower prices can bring broad usage, stronger capabilities can enter more complex tasks, and speed and supply stability can also become reasons for customers to pay.

And the same model can simultaneously support the business of the original vendor and third parties. The technological progress of model companies can be monetized through their own APIs and also enter the market through deployment on other platforms.

Therefore, the model being used the most and the model's original vendor obtaining the most revenue are not necessarily the same thing. Price, supply capability, and which platform the calls occur on will all affect where revenue ultimately lands.

In this IDC report, the highest revenue share still belongs to Volcano Engine, reaching 45.1%. Estimated by market size, that is about 4.8 billion yuan. Against the backdrop of such fierce competition, changes in the landscape, and new players entering, Volcano has held onto its first position.

Volcano has ByteDance's own models, and its infrastructure capabilities can also be used to provide third-party model services. In addition, it keeps launching end-to-end model products. It is a sample of this model being executed quite thoroughly.

From these changes, it is clearly felt that MaaS has finally reached a moment that brings a turning point to the entire cloud computing market.

The logic of cloud has reached a time for change

In the past, the relationship between models and cloud was often explained as: models need a lot of computing power, so they will drive cloud growth.

Now, this explanation is no longer sufficient.

Enterprises have begun to directly purchase the work that models can accomplish. When choosing a service provider, it first needs to judge whether the model is usable, how high the cost of handling tasks is, and whether the service can be delivered stably. As for how many cards are needed behind it and how scheduling is done, these are increasingly becoming problems that suppliers need to solve.

The object of procurement has changed, and so has the unit of competition.

DeepSeek's disclosed revenue per card per hour provides a very concrete observation angle. For the same card, how many effective tokens it can produce, how many concurrent tasks it can serve, and how much cache can be reused will all affect the revenue it ultimately creates. Inference optimization and cluster scheduling thus directly participate in determining the gross margin of a business.

For third-party service providers, open model weights only solve the problem of obtaining models. Running models stably and delivering them to customers at an appropriate price is still a set of work that requires engineering capability.

According to Volcano, programming demand on Ark has already made extensive use of third-party models such as DeepSeek. The platform adapts and optimizes around KV Cache, inference services, and underlying operators to provide call services to customers. Such deployments are already able to generate positive gross margin. Technological progress in open-source models can therefore also become a source of revenue for deployers.

Its own models give it another type of supply.

Seedance is the easiest to see. But the large number of enterprise uses of video models actually occurs at different stages of the production pipeline: creative and storyboard previews, shot and asset generation, post-production modifications, as well as e-commerce displays and multi-version marketing production. What the model outputs is video; what enterprises purchase is production capability.

According to IDC statistics, video generation contributed 32.7% of MaaS market revenue in the first half of the year, about 3.5 billion yuan. It is already a productivity market with real budgets.

On the language model side, according to Volcano, the Seed series is used for visual understanding, embodied intelligence data annotation, and long-chain tasks such as office agents and mobile and computer operation. Beyond video, some capabilities that are not so easily seen directly in consumer products are also forming enterprise demand.

One thing Volcano has done right here is to operate both models and inference services at the same time: its own models open up new uses, third-party models supplement choices customers have already validated, and engineering optimization then turns both into products that can be delivered continuously.

What it operates is enterprises' procurement of intelligence.

This also helps explain why ByteDance's positioning of this business has changed. The outside world used to find it easier to understand ByteDance's goals in AI from the growth of consumer products such as Doubao. But in June this year, a remark by Liang Rubo at the Volcano Engine FORCE conference seemed not to have attracted much attention.

He described it as, "Volcano's MaaS business is becoming ByteDance's foundational business."

He also regarded serving MaaS customers well as an important manifestation of truly doing large models well. And exactly how good model capabilities are has begun to need verification in external operations.

This change is also happening inside model companies. Zhipu's open platform and API business reached a gross margin of 24.6% in the first half of the year. But in the same period, the company's R&D expenses were still 2.131 billion yuan, and overall operations still need a larger revenue scale to support them.

These changes all show that providing model services can already generate gross margin, while continuing to develop the next generation of models requires this business to continue expanding.

Original vendors need to cover training, R&D, and service costs with revenue over the model life cycle; service providers deploying third-party models mainly need to cover inference, licensing, and operating costs. They all need real demand to absorb supply.

The significance of Anthropic achieving quarterly operating profitability for the first time lies in the fact that it has already shown: when usage scale, model capability, and service efficiency reach a certain level, these costs can be covered by revenue in the current period.

This turns MaaS from a technology monetization idea awaiting verification into a business that can continue to be operated and expanded.

MaaS decides the winner

Back to Anthropic, what it needs to do next is actually already clear.

Continue to improve model capabilities so that Claude can take on more complex tasks; then, through products such as Claude Code and Cowork, put these capabilities into the hands of more people. Developers can use the API directly, and other employees can also assign work through the products.

The same enterprise can therefore continuously purchase the same kind of model capability through different entry points.

This is also an important growth space for MaaS going forward. It will extend from a single model call to a task that can be continuously executed, checked, and adjusted. Work that customers originally needed to organize models, tools, and processes themselves is beginning to be undertaken by suppliers.

Volcano's Managed Agents provides a corresponding example. Seedance MA connects understanding requirements, retrieving references, organizing materials, generating content, and checking revisions, and enterprises submit creative tasks through the Agent API. In the official e-commerce detail image example, the Agent needs to understand the delivery requirements for a vertical long image, arrange product angles, details, and selling points, and then complete generation and layout.

The capabilities users purchase thus include how to organize this production work.

Within ByteDance, the MaaS business goal is taken very seriously, and resources have already been concentrated around that goal. The organizational adjustment announced in July brought this direction into the sales and customer service system: the relevant Feishu teams were integrated with Volcano Engine to uniformly undertake enterprises' MaaS service needs.

The integration of Doubao Work and Feishu continues to add entry points for model capabilities to enter enterprises. Enterprises can purchase through the API, use subscription and incremental-payment work products, or build their own Managed Agent on Ark and then distribute it to employees for use. Feishu had previously already supported connecting third-party models, and enterprises' existing model choices can also be combined with the platform's knowledge, tools, and collaboration capabilities.

From this perspective, end-to-end products are also a distribution method for model services. They allow more people who do not directly call APIs to continuously use the models behind them.

Alibaba is also adjusting in a similar direction. In the latest disclosed quarterly results for June, Alibaba Cloud's external commercialization revenue grew 45% year over year, and AI-related product revenue achieved triple-digit growth for the twelfth consecutive quarter. This business scope is broader than MaaS, but Qwen Office, which Alibaba also emphasizes, has already connected models, Agents, and Alibaba Cloud and DingTalk's enterprise entry points.

Another kind of change is taking place in the commercial licensing of models.

DeepSeek V4.1 Flash continues to use the MIT license; Kimi K3's public license stipulates that users operating MaaS businesses and reaching the agreed revenue threshold need to separately sign an agreement with the original manufacturer. Model weights can be open, while the rights for large-scale commercial use are beginning to be arranged separately. Some revenue-sharing ratios have already reached as high as 30%.

Of course, MaaS revenue will still need to withstand the test of customer returns, renewals, and R&D investment. But compared with repeatedly discussing whether foundation models have a business model, today we can already discuss more real questions and see more genuine competition. Intelligent progress can finally, along an operable path, turn into revenue and then support its own next round of progress.

This article comes from the WeChat public account "Silicon Star Pro", author: Wang Zhaoyang, published by 36Kr with authorization.