When a consumer app serves hundreds of millions of users, the cheapest way to sell AI is to resell the same plumbing. That is the bet behind Volcano Engine (火山引擎), and in 2025 it quietly made ByteDance one of the largest model-serving operations on earth.
The number that matters: tokens moved, not headlines.
The platform, step by step
Volcano Engine is ByteDance's cloud and AI service platform. It launched in June 2020, initially exporting ByteDance's recommendation and content tooling to outside firms.
The pivot to AI came in two moves:
- June 2023: Volcano Engine released Volcano Ark (火山方舟), a one-stop large-model service platform — MaaS — covering fine-tuning, evaluation and inference, and aggregating third-party models (Baichuan, MiniMax, Zhipu and others) alongside ByteDance's own.
- May 2024: ByteDance opened the Doubao (豆包) large-model family to external API customers, and cut pricing hard — about 99.3% below then-current industry rates, to 0.0008 yuan (≈ US$0.0001) per 1,000 tokens, moving the market from "per-cent" to "per-milli" pricing.
The logic was stated plainly by Volcano Engine president Tan Dai (谭待): large usage volume is what grinds a model — and its inference cost — into shape.
Why MaaS, not just a model
Most Chinese labs released a model. ByteDance released a layer. Volcano Ark is the place where a developer picks a model, fine-tunes it, runs evaluation, and ships inference behind an enterprise SLA — OpenAI-compatible, so switching is a base-URL change.
That positioning paid off when open models surged: DeepSeek-R1 landed on Volcano Ark in February 2025, and Volcano Engine became the cloud that captured the most DeepSeek inference traffic in China.
The token numbers
External research firms now rank it clearly:
- Omdia (Dec 2025): by daily token calls, Volcano Engine was third in the world at over 30 trillion tokens/day — about 15% share — behind OpenAI (~70 trillion, ~31%) and Google Cloud (~43 trillion, ~19%). The three together held ~65% of the global MaaS market.
- IDC: in China's public-cloud large-model token calls, Volcano Engine held 49.2% in the first half of 2025 and 49.5% for the full year — the clear number one. China's external MaaS volume reached about 1,944 trillion tokens in 2025, roughly 16 times the prior year.
- Doubao's own daily usage passed 50 trillion tokens in December 2025 and was cited above 60 trillion later that month.
One nuance worth keeping: IDC's "share" counts tokens served to outside customers and excludes ByteDance's own apps (Doubao, Douyin, Jimeng). The headline "every other token in China runs on Volcano Engine" is real only on that external metric.
The ByteDance advantage
Volcano Engine's edge is not a secret algorithm. It is the spillover from ByteDance's consumer products, which hammer the same models at extreme scale. That gives the cloud team three things competitors buy rather than build:
- Battle-tested inference optimization (KV-cache, PD separation, xLLM-style frameworks).
- A feedback loop — real usage tells the next model what to fix.
- A C-end product (Doubao) that enterprises can imitate inside their own stack.
In December 2025, Volcano Engine was named the exclusive AI-cloud partner for the 2026 CCTV Spring Festival Gala, taking a role Alibaba Cloud held the year before.
What the pricing move did to the market
The May 2024 price cut was not just a discount; it reset the unit economics of Chinese AI. By moving to a per-milli pricing tier, Volcano Engine made small and medium-scale experimentation nearly free, which pulled a wave of developers onto its endpoint. Competitors matched or undercut, and the net effect was a steep drop in China's inference cost curve — the kind of shift that shows up later as more applications, not more headlines.
The risk in the model
The flywheel depends on consumer-scale demand continuing to grow. If Douyin and Doubao usage flattens, the inference volume that justifies the capex softens. And because "MaaS leader" is measured on external tokens, ByteDance's true dependence on its own apps is hidden inside the headline share. The moat is real, but it is a demand moat, not a technology one — and demand can turn faster than architecture.
Honest limitations
Market-share figures come from IDC and Omdia, whose methodologies differ (Omdia counts global daily tokens; IDC counts China external-customer calls). The "99.3% cheaper" claim is ByteDance's own launch pricing versus its stated industry comparison, not an independent benchmark. Volcano Engine remains second to Alibaba Cloud on full-chain cloud revenue (IaaS+PaaS+MaaS), so "MaaS leader" is not the same as "cloud leader."
What readers can do now
- If you run inference in China, benchmark Volcano Ark's OpenAI-compatible endpoint against Alibaba and Tencent — the price gap is real but check latency for your region.
- Watch the MaaS share number as the cleanest proxy for who actually operates China's model layer, separate from who trained the smartest model.
- Treat Doubao's token volume as evidence that consumer-scale usage, not model-release hype, is what compounds infrastructure advantage.