Exclusive: Volcano Engine fully pivots to commercialization, with MaaS achieving a gross margin above 55% in the first half of the year.

Insiders told us that in the early stages of industry development, vendors needed to increase token consumption to drive improvements in model capabilities, and also needed to verify whether the MaaS business model could be viable. But now, model capabilities have reached a high level, and the MaaS business model has been validated, so commercial revenue has become more important and better reflects the model’s true performance.

On October 28, IDC released its 2026 H1 China enterprise‑level MaaS market report, reflecting this trend. The revenue size of the China enterprise‑level MaaS market in the first half of the year has exceeded 10 billion yuan, indicating that MaaS is becoming a business with scalable revenue. In terms of revenue, Volcano Engine’s market share reached 45.1%, with Alibaba Cloud ranking second at 25.2%.

The new trend also appears in the revenue gap between different AI businesses. “LatePost” previously noted that although the consumer‑facing product Doubao App is the largest AI application by Chinese user base, its revenue in the first half of this year was still less than 100 million yuan. In contrast, Zhipu, which has a much smaller user base, achieved 954 million yuan in revenue in H1 2026, with its open platform and API business contributing 825 million yuan.

The situation overseas is similar. On October 26 (local time), Anthropic’s filing with the U.S. Securities and Exchange Commission showed that in H1 2026 its revenue reached $16.286 billion, about 14 times that of the same period last year. Of this, the business charging based on model usage contributed roughly $13.1 billion, accounting for 80% of total revenue.

It’s not just Volcano Engine; Alibaba Cloud is also beginning to focus on MaaS revenue.

Last February, Alibaba CEO Wu Yongming said that the revenue generated by the Bailei platform through API charges was relatively small, and the Tongyi Qianwen model itself might not directly bring substantial income. However, in March of this year, he publicly stated that MaaS will become the product with the largest revenue scale for Alibaba Cloud. In May, Wu further mentioned that as MaaS business revenue grows rapidly, Alibaba Cloud’s gross margin is also expected to improve.

Around 2020, when Volcano Engine entered the cloud computing market, vendors such as Alibaba Cloud, Huawei Cloud, and Tencent Cloud had already been operating for many years, boasting a massive customer base and mature product ecosystems. Cloud computing is a market where first‑mover advantage is obvious. Once a company deploys its data, systems, and business processes on a particular cloud, migration entails high costs and risks, making it difficult for newcomers to win customers.

To find its niche, Volcano Engine turned its attention to the somewhat niche area of AI. Traditional cloud computing mainly relies on CPUs for general‑purpose tasks, whereas AI workloads depend on GPUs and new compute clusters and software systems built around GPUs. However, at that time, there were not many enterprises that actually needed GPUs.

The turning point came in 2022, when large‑model startups such as MiniMax and Zhipu began using Volcano Engine’s computing power to train models; a few months later, ChatGPT sparked a global large‑model boom, and Volcano Engine finally got the opportunity for rapid growth.

At the same time, ByteDance quickly elevated Volcano Engine’s MaaS to the highest business priority.

A person from Volcano Engine told LatePost that only by applying the model to real enterprise scenarios can the R&D team understand customers’ actual needs and the model’s shortcomings. For example, Volcano Engine found in collaborations with enterprises that customers often need AI to handle long‑running, multi‑step complex tasks, including agent‑style programming. This feedback has driven ByteDance to strengthen the model’s capabilities in long‑task execution, programming, and tool invocation. As model capabilities improve, they attract more enterprise customers, generating new demands and feedback.

In the early development stage, token usage was Volcano Engine’s most critical business metric. Massive token consumption not only proves market demand for ByteDance’s models but also helps the company accumulate service experience. IDC data shows that by 2025, China’s public‑cloud large‑model token calls will reach 1.944 quadrillion tokens, with Volcano Engine capturing nearly half of the market share, ranking first.

However, entering 2026, as model capabilities continue to expand and market competition intensifies, token consumption alone can no longer fully gauge a MaaS provider’s market position and overall competitiveness; actual revenue has become a more important metric.

A senior person from a leading cloud vendor told us that as companies like Alibaba and Zhipu continue to improve programming, inference, and other capabilities, enterprises have more options. Moreover, when some model capabilities become comparable, price becomes a key factor influencing customer choice. Vendors can attract usage by lowering prices, but higher token consumption does not necessarily mean more revenue.

Models have also expanded from text Q&A to visual understanding, speech synthesis, and image/video generation, and the pricing standards for different modalities vary widely. Taking Volcano Engine’s public pricing as an example, Doubao Seed 2.1 Pro charges ¥6 per million input tokens and ¥30 for output tokens; image, speech, and other models may be billed per generated image or audio duration. Therefore, even if two vendors have the same token usage, their revenue can differ dramatically.

Another deeper shift this year is that more large models that achieve state‑of‑the‑art performance in specific domains are actually making money. For example, Zhipu leverages its leading capability in programming to attract many paying users; Kimi’s model is larger and priced higher, yet customers are still willing to pay.

This shows a clear trend in the large‑model market: as long as a model is sufficiently ahead in an important domain, users are willing to pay a premium for its superior capabilities.

For MaaS vendors, the commercialization opportunity is even greater. Enterprises need different models for various scenarios such as programming, office work, image generation, and video generation. MaaS vendors can offer both in‑house and third‑party models, allowing customers to choose based on actual needs and pay according to usage. As more models demonstrate leading capabilities across different domains, the demand that MaaS platforms can handle will also grow. By continuously providing competitive models and controlling inference costs, there is a chance to generate revenue from more companies and more scenarios.

Since the rise of large models at the end of 2022, they have been around for almost four years, and commercialization has become increasingly important. In the early stage, call volume could prove that a model had users and market demand; now, as models enter actual enterprise production workflows, vendors also need to demonstrate that customers are willing to pay continuously. High R&D and inference costs make commercial returns even more urgent.

“For Volcano Engine, MaaS business revenue is the top priority and the only important goal,” said an insider.

MaaS has become the most certain business in the AI industry

In recent years, figuring out how to keep making money has been a persistent challenge for many AI companies.

Consumer‑facing AI applications have amassed large user bases, but scenarios that encourage long‑term paid usage remain limited. In China’s enterprise services market, companies have historically shown low willingness to pay for efficiency tools, and traditional SaaS software, which relies on selling accounts and subscription fees, struggles to achieve a large revenue scale.

ByteDance’s internal AI apps such as Doubao Work, Dream, and Doubao follow a similar model: users can pay a subscription fee and also purchase credits to get more model usage quota. Sources told LatePost that even if AI tool subscription revenue exceeds that of traditional SaaS, it still represents only “a drop in the bucket” within the overall business.

In contrast, MaaS’s business model is simple: enterprises call the model and pay per usage. Whether it’s AI programming, data analysis, or generating images and videos, as long as the model is continuously used, cloud service providers can earn revenue. As AI becomes part of more enterprises’ daily work, model calls will shift from occasional needs to ongoing consumption, making MaaS a sustainable and scalable business.

Even as model prices keep dropping, MaaS can still generate gross profit. Currently, all inference services offered by Volcano Engine across various models have positive gross margins, including the self-developed Doubao large language model, the Seedance video generation model, and third‑party DeepSeek models.

Take DeepSeek, which maintains low prices over the long term, as an example. It reduces inference costs by optimizing model architecture, caching mechanisms, and other techniques. Volcano Engine can further lower costs by increasing GPU utilization and optimizing compute scheduling. Since it does not bear DeepSeek’s upfront R&D and training expenses, Volcano Engine only needs to handle model deployment and inference services to earn gross profit.

However, for MaaS to keep growing, continual improvements in model capabilities are still essential. Each new capability a model acquires opens up new enterprise use cases, driving additional call demand.

Since the beginning of this year, Seed 2.1 Pro has further improved performance on programming and complex agent tasks, pushing the model into office agents, automated phone and computer operations, and embodied intelligent data annotation scenarios. The video generation model Seedance 2.0 has increased the usability of video generation, also driving more applications of the model in film, advertising, e‑commerce and other enterprise production processes. In addition, Zhipu’s rapid progress in the programming field has also boosted Volcano Engine’s revenue growth.

It is reported that the office scenario is expected to become the next major revenue growth point for Volcano Engine’s MaaS business.

On July 30 this year, ByteDance reorganized the structures of the Doubao, Feishu, and Volcano Engine businesses. The Feishu product team merged with the Doubao product team, while Feishu’s sales, marketing, and customer service teams integrated with Volcano Engine, further linking models, office products, and enterprise sales channels.

According to LatePost, Doubao’s services will be opened to enterprise customers, allowing integration of many third‑party models beyond ByteDance’s own developments.

The person mentioned said that Doubao Work is essentially a form of MaaS product, using a combination of subscription fees and pay‑as‑you‑go pricing. In addition to purchasing subscription plans, users can buy fuel packs to obtain additional model usage quotas. As users handle more tasks and the complexity increases, model invocation volume and corresponding revenue have the opportunity to keep growing.

Doubao Work also expands Volcano Engine’s coverage of enterprise customers. Within the same company, developers can directly call model APIs, technical teams can use the hosted Agent development tool to build their own applications, and ordinary employees can directly use Doubao Work to complete tasks. By offering products at different tiers, Volcano Engine can meet the needs of more people inside an organization, allowing employees who would not normally use APIs to directly leverage large models, thereby generating more model calls.

Compared with simple chat Q&A, enterprise employees also use Doubao Work to handle more complex, long‑chain tasks. More real‑world usage feedback can help R&D teams identify model shortcomings and improve capabilities; stronger models then have the chance to take on more complex tasks, driving additional call demand.

A senior executive from a leading cloud provider told us that when MaaS first emerged, many cloud vendors saw it as a good business, but not the most important one. They are more accustomed to using MaaS as a way to attract customers, thereby driving sales of traditional cloud services such as compute and storage. “It’s kind of like using an old map to find a new continent.”

After several years of development, the commercial value of MaaS itself has gradually become an industry consensus. “MaaS is the AI business with the highest certainty right now,” the aforementioned senior cloud executive said.