According to an exclusive report by LatePost on October 30, the model-as-a-service (MaaS) business of Volcano Engine, ByteDance's cloud arm, pulled in 4.8 billion yuan in revenue during the first half of 2026, with a gross margin above 55%. The company has also raised its original annual revenue target of 1.5 billion yuan for the unit. Note the phrasing: 1.5 billion yuan was the original target, and the new number has not been disclosed. The same report contained an even starker contrast: Doubao, ByteDance's consumer chatbot, is the largest AI application in China by user base, yet it earned less than 100 million yuan in the first half of the year. A national-scale app cannot scrape together pocket change, while a pay-per-token enterprise service books 4.8 billion. After two years of the "monetize the consumer app" storyline, the ledger tells the opposite story.
Let me draw the boundary first. This article covers commercialization and market structure only: revenue, market share, margins, pricing, and the statistical definitions behind them. Which model is smarter is a different question, handled by our flagship LLM comparison and our open-weight agent flagship comparison elsewhere on this site.
Where the 4.8 Billion Comes From
The MaaS business model fits in one sentence: open model capabilities to enterprises and charge by actual usage. Volcano Engine's shelf carries both its own models, including the Doubao large language model and the Seedance video generation model, and third-party models such as DeepSeek and Zhipu. According to LatePost, every inference service currently runs at a positive gross margin, including the hosted DeepSeek offering. The logic stacks two layers: DeepSeek itself drives down inference costs through its architecture and caching mechanisms, and Volcano Engine squeezes costs further by raising GPU utilization and optimizing compute scheduling, all without carrying DeepSeek's upstream training expenses.
The price list is public. Volcano Engine's published rates show Doubao Seed 2.1 Pro at 6 yuan per million input tokens and 30 yuan per million output tokens, with image and voice models billed per generated image or per minute of audio. What deserves attention is the shift in posture: early in the year the company was still chasing higher token consumption, but from midyear onward, commercial revenue replaced consumption as the single most important business metric. One person familiar with the matter put it to LatePost more bluntly: for Volcano Engine, MaaS revenue is the first priority, and the only goal that matters.
The shift has history behind it. LatePost's retrospective reads like this: Volcano Engine entered the cloud market around 2020 from the AI and GPU direction, when Alibaba Cloud, Huawei Cloud and Tencent Cloud were already far ahead. In 2022, startups like MiniMax and Zhipu trained their models on Volcano's compute, and after the ChatGPT wave, ByteDance elevated MaaS to its highest business priority. In the industry's early days, vendors had to expand token consumption to sharpen model capability and validate the business model. Now that the model is validated, commercial revenue is the more honest measure of how good a model really is.
The consumption numbers are indeed staggering. By Volcano Engine's own announcements across launch events, the Doubao model family went from processing 120 billion tokens per day at its debut in May 2024, to over 30 trillion by the end of September 2025, past 50 trillion in December 2025, 120 trillion in March 2026, and more than 180 trillion by June 2026, roughly a 1,500-fold increase in just over two years. The number of enterprise customers consuming over 1 trillion tokens per year grew from 100 in December 2025 to 200 in 2026. But that is a consumption narrative; 4.8 billion is the first monetization receipt it has delivered.
Three "Number Ones," Three Metrics: 45.1%, 49.5% and 10.2%
The 4.8 billion figure can be independently verified. IDC, the research firm, published its "China AI Software Market Semiannual Tracker, 26H1" on October 28: China's MaaS market reached 10.63 billion yuan in the first half of 2026, up 889.4% year over year, nearly a tenfold jump. On the all-modality revenue share metric, Volcano Engine ranked first with 45.1%, which works out to about 4.8 billion yuan and matches LatePost's number. Alibaba followed at 25.2%, roughly 2.68 billion yuan, with Zhipu third at about 530 million yuan, Baidu at 4.5%, and Kuaishou rounding out the top five.
But the same report contains another "number one": looking only at large language model revenue share, Alibaba leads with 46.6%, Zhipu is second at 10.6%, and Volcano Engine sits third with just 10.2%. The two crowns do not contradict each other; they measure different things. IDC tracks LLM revenue separately from image, video and other modalities, and the revenue mix is the key to reading the table: in the first half of 2026, large language models accounted for 47.0% of total MaaS revenue, video generation models contributed 32.7%, and all other modalities made up the remaining 20.2%. Volcano Engine wins because Seedance has swallowed the video generation market; count only the text half and Alibaba's Qwen plus Bailian platform covers more ground. So the two circulating claims, "Alibaba leads by a mile while Volcano has barely 10%" and "Volcano is the runaway leader," each grab half of the same table: neither lying, neither complete.
There is a third number to handle with care: 49.5%. That is the token consumption share, an IDC figure Volcano Engine disclosed at its FORCE developer conference in June 2026. Leading on volume does not mean leading on revenue, because pricing, model mix and customer composition differ across vendors, and customers who burn more tokens do not necessarily contribute proportional revenue. To summarize: 45.1% is the all-modality revenue share, 49.5% is the token consumption share, and 10.2% is the LLM-only revenue share. Three numbers, three metrics. Cite any of them with its metric attached; mix them once and the conclusion bends.
Several background figures from the same report are worth keeping. Total market token consumption reached about 2,857.9 trillion in the first half of 2026, up 432.5% year over year. MaaS revenue as a share of China's public cloud PaaS market jumped from 5.2% in the second half of 2025 to 21.0% in the first half of 2026. Per reporting relayed by Beijing Daily, the statistics cover only public cloud LLM services sold to external customers, excluding vendors' internal usage. IDC's industry read is that enterprise demand is shifting from trial deployments to production-grade ones, and that model capability, inference efficiency, and compute scale with unit cost and supply-chain certainty form the hardest moat for competitors to copy.
From Chasing Tokens to Collecting Revenue
A gross margin above 55% deserves its own dissection. LatePost's earlier deep dive, "ByteDance's AI Ledger," published around midyear, offered finer numbers: Seedance's annualized revenue run rate had reached 2 billion US dollars at the time, roughly 14.3 billion yuan, with monthly revenue above 1 billion yuan, nearly offsetting Doubao's compute costs. Its gross margin hit 70%; for every 10 yuan of API sales, servers and inference consumed about 3 yuan. Pin down the metric here: the 2 billion dollar run rate is a point-in-time figure from when that piece was published, not first-half cumulative revenue, and the two must not be conflated. LatePost also noted that Seedance's revenue growth has since decelerated, a slowdown in growth, not a decline. Kuaishou's Kling saw its run rate stall after hitting 500 million US dollars in early May, so the video generation curve overall has entered a plateau phase.
The biggest cost sits in chips and data centers. LatePost has reported that ByteDance's main costs come from self-built data centers and self-purchased chips, and whether AI chips should be depreciated over a traditional server lifecycle or a shorter technology cycle remains unresolved, a debate that runs through Silicon Valley as well. The original report also did not specify whether the 55% margin includes depreciation. Treat it as a directional signal of improving economics, not a precise accounting figure.
The broader market is cooperating. IDC attributes the 2026 boom in China's MaaS market to three forces: model upgrades that lower capability and cost barriers simultaneously; stable revenue from high-concurrency scenarios like recommendation, search, chat and coding, with office and customer service scenarios next in line; and video generation crossing the commercialization threshold to become the second growth curve. ByteDance's product moves track the same logic: Seed 2.1 Pro strengthens coding and complex agent tasks, pushing the model into office agents, phone and computer automation, and embodied-AI data annotation, while Seedance 2.0 improves video generation usability for production pipelines in film, advertising and e-commerce. And there is now a global template. On October 26 local time, Anthropic filed its prospectus with the US SEC: first-half 2026 revenue of 16.286 billion US dollars, about fourteen times the year-ago period, with usage-based model access contributing roughly 13.1 billion dollars, an 80% share. Pay-per-use model services are being validated in the world's two largest markets at the same time.
The Scoreboard: Alibaba, Zhipu and Baidu, Each With Its Own Books
Reading Volcano Engine's position, the cardinal sin is comparing numbers from different metrics. Every competitor keeps its books on a different basis.
Zhipu is the easiest to get wrong. On LatePost's figures, Zhipu's total company revenue in the first half of 2026 was 954 million yuan, of which its open platform and API business contributed 825 million. On IDC's figures, Zhipu's public cloud MaaS revenue was only about 530 million yuan. The gap is nearly double, and the reason is not an error but a boundary: total company revenue includes private-delivery contracts, while IDC counts only public cloud MaaS services. Both numbers stand; neither may be mixed with the other, and writing "Zhipu's MaaS revenue was 954 million" would be wrong. The interesting wrinkle is that LatePost credits Zhipu's rapid gains in coding with boosting Volcano Engine's revenue too. Hosting third-party models generates real income; a platform's value does not end at its homegrown models.
Alibaba's ledger is bigger and broader. For the quarter ended June 30, 2026, Alibaba's AI cloud and compute services segment grew 45% year over year to 48.44 billion yuan. Note that this is a segment figure including compute services, far wider than IDC's MaaS revenue definition, and the two are not comparable. The more telling data point is the attitude curve. LatePost traces Alibaba CEO Eddie Wu's public statements: in February 2025 he described the Bailian platform's API revenue as small; by March this year he said MaaS would become Alibaba Cloud's largest product by revenue; and in May he added that fast-growing MaaS revenue should improve Alibaba Cloud's margins. In barely a year, the boss's framing moved from "small" to "largest." That curve is the industry's mood gauge.
Baidu runs a different route. Its government and enterprise large-model contract wins totaled 1.385 billion yuan in the first half of 2026, roughly 60% of the 2.044 billion yuan won across five cloud vendors. But project-based contract wins and pay-as-you-go public cloud MaaS revenue are entirely different metrics; the number shows Baidu's strength in private deployment and government business, nothing more. IDC separately measured China's private LLM deployment market at 2.88 billion yuan in the first half, up 108.4%, with Baidu and SenseTime among the leaders. Private deployment and public cloud MaaS are two markets; their numbers cannot be added, and their respective "leaders" cannot be translated into each other.
Next Stop: Office Scenarios and a Reorg
LatePost's judgment is that office scenarios could become Volcano Engine's next major revenue growth driver. The organizational groundwork came on July 30, when ByteDance restructured Doubao, Feishu and Volcano Engine: the Feishu and Doubao product teams merged, and Feishu's sales, marketing and customer service teams folded into Volcano Engine, connecting models, office products and enterprise sales channels in one line. Product moves followed: Doubao Work, ByteDance's office suite, opened third-party model access for its team and enterprise editions, starting with Kimi K3.1 Pro, GLM-5.3 and DeepSeek V4.1 Flash, billed through the existing AI credit system. LatePost describes Doubao Work as effectively another product form of MaaS, monetized through subscriptions plus pay-as-you-go credit packs.
The hierarchy of the business was spelled out in one line: even if subscription revenue from AI applications, Doubao Work, Jimeng and Doubao itself, exceeds that of traditional SaaS, it remains "the small change of the small change" within ByteDance's overall business. One more timeline link: around April this year, ByteDance executives visited Anthropic, then began reallocating AI resources from consumer-facing products toward enterprise-serving ones.
In the end, what the 4.8 billion changes is not the ranking itself but the industry's default answer to the question of how AI makes money. At ByteDance, at least, the answer is written plainly: not consumer subscriptions, but enterprise pay-per-use. Doubao handles the user scale; MaaS handles the revenue; the two legs walk their own paths. One earned less than 100 million yuan in six months, the other 4.8 billion. The more jarring the contrast, the clearer the answer.
Related reading
- Flagship LLM Comparison: How to Choose Opus 5.5, Astra or Argon — this article counts the money; the closed-source capability shootout lives here.
- KV Cache Gap of 39x: Four Open-Weight Agent Flagships Compared — why caching slashes inference costs, quantified across open-weight flagships.
- No.1 With 59, but You Can't Buy It: Gemini 4 Carbon's Gamble — another commercialization path: lock the throne first, open the cash register later.
- M3.1 Flash Weights Are Out, GitHub Comes Up Empty — a different combination of open weights and commercial strategy.