Volcano Engine Takes 45.1% of China's MaaS Market, with Seedance Driving Over Half of Revenue

Zhidongxi reported on November 3 that today, the well-known industry research firm QuestMobile released its autumn report: in September 2026, Hongguo Free Short Dramas surged to second place with 430 million deduplicated users, overtaking Doubao and ranking behind only Douyin.

Previously, QuestMobile data showed that in July 2026, Hongguo's daily active users climbed to 168 million, a figure that has exceeded the combined daily active users of the four long-video platforms Tencent Video, iQIYI, Youku and Mango TV. Measured by total user time spent, Hongguo ranks fourth among China's mobile internet apps, behind only Douyin, WeChat and Douyin Lite.

From novel IP adaptation and AI comic-drama production to short-video distribution, the short-drama industry chain's demand for video generation models is being converted into real Token revenue. ByteDance's video generation model Seedance has become a key tool in this chain, and it has also allowed ByteDance to make a fortune in the MaaS (Model as a Service) market.

On October 29, the latest report from International Data Corporation (IDC), "China AI Software Market Semiannual Tracker, 26H1," showed that in the first half of 2026, China's MaaS market reached 10.63 billion yuan, up 889.4% year on year. Where does the money in China's MaaS market flow?

The spotlight falls on ByteDance's cloud business, Volcano Engine, which ranks first with a 45.1% market share. That means that for every 100 yuan of MaaS revenue in China, nearly 45 yuan flows to Volcano Engine.

▲China MaaS market by revenue share in the first half of 2026 (Source: IDC)

But if we look separately at the large language model (LLM) market, Volcano Engine's ranking is not so high. The IDC report shows that Alibaba Cloud ranks first with a 46.6% share, while Volcano's share is 10.2%, ranking third.

▲China MaaS market by revenue share in the first half of 2026 - large language models (Source: IDC)

So the question arises: how can a player ranked third in LLM MaaS revenue take nearly half of the revenue in the entire MaaS market?

The answer points to the video generation model Seedance. Video generation, a track once regarded by many as high-cost and difficult to monetize at scale, and a category once ruthlessly abandoned by OpenAI (Sora was shut down in April), has now become the key engine for ByteDance to top China's MaaS revenue ranking.

But the answer is not only Seedance. In a June interview with the media, Volcano Engine President Tan Dai spoke about how "model capability is the most important growth engine for the MaaS business," and also stressed that "Token revenue is the most important metric." How model capability is converted into Token revenue, and how to truly scale the Token business, is another key to understanding Volcano Engine's rise to the top this time.

▲Volcano Engine President Tan Dai

Volcano Engine officially launched in June 2020, and in 2021 it explicitly proposed becoming the "fourth cloud" in the domestic cloud market, benchmarking against Alibaba Cloud, Tencent Cloud and Huawei Cloud. Now, in this wave of AI large models, China's MaaS market has expanded to nearly nine times its size within a year, MaaS has officially entered the ten-billion-yuan track, the market shares of various players have begun to take shape, and Volcano Engine's first-place share is all the more meaningful for interpretation.

▲China MaaS market size in the first half of 2026 (Source: IDC)

Standing at this new juncture, how exactly did ByteDance capture the lion's share of the MaaS market step by step? How much of this lead comes from model capability, and how much from scenario selection and business strategy? When video generation props up half of the revenue, how long can Volcano Engine's first place last?

1. OpenAI gives up, ByteDance reaps: Seedance props up China's No. 1 MaaS player

Behind the 45.1% share is, first of all, ByteDance's strong penetration of video, a high-Token-density scenario.

First, let's look at the distribution of China's MaaS revenue sources published by IDC: in the first half of 2026, China's MaaS market revenue grew nearly ninefold year on year. Among this, large language models contributed 47.0%, video generation models contributed 32.7%, and other modality models contributed 20.2%. This means that video generation has become the second-largest revenue source in the MaaS market. For Volcano Engine, its significance is even greater.

In a June interview, Tan Dai mentioned that the video generation model Seedance is currently the largest contributor to Volcano's MaaS revenue, accounting for more than half. Previously, multiple media outlets also disclosed that Seedance 2.0's monthly revenue has exceeded 1 billion yuan, leading peers. Statistics from the industry research firm AI Pris corroborate this: in mid-to-late May this year, calculated by average daily computing power consumption, Seedance accounted for more than 80% of the domestic video generation market, Kling about 14%, and Wanxiang 2.7 about 4%.

Why can video generation support such large revenue? First, this is a scenario that naturally consumes Tokens. E-commerce advertising, short dramas and game CG are all high-frequency scenarios that can form closed loops and directly generate cash flow, and model invocation frequency and per-task consumption are far higher than ordinary text Q&A.

At the same time, ByteDance has mastered the "Seedance + content ecosystem" closed loop: on the one hand, it provides model tools for drastic cost reduction and efficiency gains—Seedance has now iterated to version 2.5, supporting 30-second video generation, multi-round continuation, high temporal consistency, and reference to 50 image-text materials, greatly compressing production costs; on the other hand, ByteDance has given Seedance an ecosystem closed loop, with Fanqie Novel providing IP, Seedance generating animated dramas, and Hongguo Short Drama and Douyin handling distribution, thereby allowing commercial demand to circulate within the ecosystem.

▲Hongguo Short Drama App page (Image source: Zhejiang Business Magazine)

It is worth noting that IDC counts public cloud large-model service invocations that cloud vendors provide to external customers, and does not include invocations by ByteDance's own businesses such as Doubao, Jimeng, and Douyin. So the 45.1% first place was not built up by internal traffic.

In this high-density scenario, ByteDance pocketed the revenue before competitors had formed effective supply.

II. Models Are Hard to Monopolize, So Bundles Stake Out Territory: Agent Plan Seizes the Agent Entry Point

Seedance solved the bulk of Volcano Engine's MaaS revenue, and Agent Plan is now making it earn even more.

On May 11, 2026, Volcano Engine released Agent Plan. According to ByteDance, this is the industry's first subscription-based Agent Plan package aimed at Agent scenarios, bundling multi-model + Harness (agent orchestration framework) toolchains. Unlike the Token Plans on the market that only sell model Token quotas, Volcano Engine's Agent Plan supports users in running complete agent tasks out of the box.

After developers integrate, they can invoke these models and capabilities in tools such as Claude Code, OpenCode, TRAE, and OpenClaw. For example, when having an Agent build a website, it can first invoke GLM or Kimi to handle the task, then invoke Seedream to generate images and Seedance to generate videos, and finally complete the code build.

Being the first in the industry to lay out Agent Plan reflects ByteDance's judgment on the next stage of competition: the competitive logic of the model market is changing, and developers often choose multiple models according to the task. ByteDance acknowledges that the model layer cannot retain everyone, that relying only on the single-item capabilities of self-developed models makes it hard to cover all needs, and that it is also hard to ensure users always stay on the same platform, so it chooses to retreat to the position of collecting tolls.

Volcano Engine's Agent Plan has four subscription tiers: 40 yuan, 200 yuan, 500 yuan, and 1,000 yuan. It is the first to introduce AFP (Agent Fuel Points) as a unified resource measurement unit, uniformly measuring model and Harness usage, with Auto mode scheduling models by task, and web search and embedding also bundled in. What it sells is a runtime package for developers that can run Agents.

Compared with other Token Plan team-edition packages on the market, where a standard seat costs about 200 yuan/seat/month and what is sold is centralized management of enterprise AI budgets, seat allocation, usage analysis, SSO, and multi-tenant isolation, Volcano's Agent Plan has a minimum threshold of only one-fifth of that and focuses more on providing a runtime environment.

It can be seen that ByteDance may judge that it is difficult for it to establish an absolute advantage at the large language model layer, so it tries to use Agent packages to create switching friction.

Whether this model can form genuine customer stickiness still needs time to verify, but ByteDance has already begun competing for entry points beyond models.

III. Concentrated Firepower + Feishu Synergy: Volcano MaaS Achieves Lightning-Fast Revenue Growth

Tan Dai joined Baidu in 2008 and was once responsible for search, cloud storage, and infrastructure; he joined ByteDance in 2020 and subsequently took charge of Volcano Engine. He is familiar with search engines and also with the engineering efficiency of large-scale systems, and this is the organizational leadership backdrop behind Volcano Engine's rapid surprise attacks.

He once used one sentence to explain Volcano's strategic choice of going all in on AI: "If you went back to 2012, the year ByteDance was founded, would you go big into PC search?" The answer is obviously no. We have seen that when ByteDance entered search back then, it avoided Baidu's traditional web search main battlefield and relied on short-video content differentiation and engineering-efficiency tactics to capture incremental traffic at low cost.

Therefore, when entering the public cloud battlefield, Volcano Engine chose to avoid the stock competition on the traditional cloud frontal battlefield, using MaaS as a wedge to enter the market and using model services to drive cloud consumption. At the same time, Volcano Engine did not spread resources evenly across ten general scenarios, but concentrated on video generation, a point with high Token density and clear willingness to pay.

Its pricing logic also bears traces of the search war. Tan Dai does not work backward from gross margin to set prices, but looks at how much value a Token can create. The benefit of customer migration must be two to three times higher than the migration cost before people are willing to use it. Following this logic, in 2024 Volcano Engine pushed large-model prices to the floor, because at that time models could only do Chatbots with limited value. Now Seedance can enter the production processes of advertising, short dramas, and games, so prices can move upward.

Organizationally, in July 2026 the Feishu GTM team was integrated with Volcano Engine to establish the Creativity Service Platform, with Tan Dai responsible for ByteDance's MaaS and SaaS sales system. The model business and the office workflow business have become more closely tied, and Feishu scenarios have become the second layer of switching cost for MaaS.

Together, these "blitzkrieg"-style moves also drove the rapid revenue growth of Volcano Engine's MaaS business.

IV. Is the Top Spot Stable? Three Hurdles to Pass: Lopsidedness, Stickiness, and Cash Burn

To sit firmly on the throne of China's No. 1 MaaS player, ByteDance still faces several tests.

The first question is that the revenue structure relies too heavily on Seedance. ByteDance's advantage is built on the creation, distribution, and consumption closed loop linked by Hongguo, Jimeng, and Douyin, but outside that closed loop, overseas customers and professional film and television customers may well have cheaper alternatives.

The video generation track is far from settled. Kuaishou Kling's Q2 2026 revenue exceeded 850 million yuan, up more than 200% year on year; Kunlun Wanwei's AI short drama business is also growing rapidly; MiniMax is entering with low prices and open weights. Google Veo, Alibaba HappyHorse, and others continue to step up... Seedance's first-mover dividend may gradually narrow.

At the same time, language models remain its weak spot. In the first half of 2026, IDC statistics show Alibaba Cloud's LLM MaaS share reached 46.6%, while Volcano Engine had only 10.2%. The core entry points for enterprise Agents are complex reasoning, Coding, and workflows, and ByteDance still needs to build a second growth pillar.

The second question is whether Agent Plan can form genuine customer stickiness. As model aggregation becomes an industry standard and pricing strategies across vendors change, the appeal of Volcano Engine's subscription price starting at 40 yuan may weaken. In the long run, what enterprises value more is API stability, service cost, security, and workflow migration cost. Only by entering core production systems can long-term dependence be established.

The third question is whether computing power investment can bring sustained returns. Multiple media outlets have reported that ByteDance's 2026 capital expenditure plan exceeds 200 billion yuan, and behind this huge investment are tests of computing power utilization, inference efficiency, and Token cost. Future MaaS competition will still come back to who can use the same computing power to continuously produce more Tokens with commercial value.

Conclusion: Video Generation Supports Half of MaaS, and ByteDance Needs More Pillars

Looking back at Volcano Engine's rise to the top this time, ByteDance built a revenue barrier with an ecosystem closed loop and extreme efficiency in a high-Token-density scenario where competitors had not yet fully committed, and then used Agent Plan and organizational integration to push entry points and switching costs, thereby gaining a considerable MaaS share.

Whether it can sit firmly on the top spot still depends on whether more pillars can grow beyond Seedance. In an interview in June this year, Tan Dai mentioned: "Although our current situation is that Seedance sells a bit more, I hope LLM can become the bulk later." From Seedance to Doubao, from model services to the Agent platform, ByteDance has clearly begun laying out for the next stage.