ByteDance’s Volcano Engine Tops 100 Trillion Daily Tokens, Turning China’s MaaS Fight Into a Scale Game
ByteDance’s Volcano Engine has pushed its Doubao large-model platform beyond 100 trillion tokens in average daily calls since the 2026 Lunar New Year, a scale milestone that tightens the link between AI adoption and cloud revenue—and raises the stakes in China’s intensifying Model-as-a-Service (MaaS) price-and-capacity war.
LatePost reported the daily token volume climbed more than 60% in less than two months, with token consumption from individual users rising sharply as agent-style products spread from developer tools to mass-market workflows. The acceleration underscores how token throughput is becoming an operating metric investors track alongside cloud bookings, as higher usage can translate into recurring inference spend—provided vendors can keep margins intact amid rising compute costs.
In global terms, LatePost said only three companies currently exceed 100 trillion tokens in daily consumption: OpenAI, Google and ByteDance. Unlike the first two, ByteDance’s volume is still primarily China-driven, while overseas growth remains in an earlier phase and constrained by product rollout—Volcano Engine’s Seedance 2.0 video-generation API has not formally launched overseas, according to a Volcano Engine source cited by LatePost.
Token surges amplify MaaS economics
Tokens—AI’s basic unit of text and multimodal computation—have shifted from a technical KPI to a commercial one because agents can burn far more tokens than chat-style interactions through multi-step reasoning and tool calls. Nvidia CEO Jensen Huang said at NVIDIA GTC on March 16 that tokens will become the most valuable “commodity” in the digital world, with throughput and generation speed turning into CEO-level dashboards—an argument cloud vendors are using to justify higher pricing and capacity investments.
China’s national-level usage data shows how quickly that commodity market is forming. The National Data Administration said China’s daily token consumption averaged 100 billion in early 2024 and surpassed 30 trillion by the end of June 2025—an increase of more than 300x in roughly 18 months. Volcano Engine’s move above 100 trillion daily tokens implies ByteDance alone now represents a material share of the country’s inferred demand, strengthening its leverage in upstream GPU supply, data-center utilization and enterprise procurement cycles.
Cloud rivals reposition as MaaS moves to the revenue core
The token boom is pushing MaaS from a peripheral add-on into a core battleground for Alibaba Group Holding Ltd.’s Alibaba Cloud, Tencent Cloud and Baidu AI Cloud—especially as traditional IaaS growth slows and vendors search for higher-frequency consumption revenue.
Amazon’s AWS told investors in an October earnings call that, over the long term, MaaS could contribute revenue comparable to EC2, its core compute product that accounts for more than 30% of AWS revenue. In China, Alibaba Cloud launched a “Bailian campaign” in late 2025 aiming to more than triple token volume on its Bailian MaaS platform, and on March 16 created an Alibaba Token Hub business group led directly by CEO Eddie Wu. Alibaba Cloud has publicly set a 2026 ambition to capture 80% of incremental growth in China’s AI cloud market, according to LatePost.
Volcano Engine, which started later than Alibaba, Huawei and Tencent in traditional cloud, has treated MaaS as a fast lane: it elevated MaaS in sales incentives in 2024 and, according to LatePost, previously set a target of more than RMB 10 billion (US$1.39 billion) in MaaS revenue in 2026—later raising that goal after launching models including Seed 2.0 and Seedance 2.0 and amid sustained momentum from agent products such as OpenClaw.
Agents shift demand from projects to subscriptions
Agents are changing buying behavior from one-off pilots to subscription-like usage curves. In 2025, Anthropic’s Claude Code helped normalize “coding plans” that encourage heavy, repeat usage, pushing token consumption higher per user. SemiAnalysis estimates about 4% of public GitHub commits are currently completed by Claude Code and projects the share could exceed 20% by the end of 2026, a trajectory that implies structural—not cyclical—growth in inference demand.
In China, the agent wave spread quickly in early 2026. Zhipu raised prices on its GLM Coding Plan by at least 30% in February, citing higher developer demand. MiniMax, which has emphasized subscription mechanics in its IPO materials and first earnings report, said on March 23 it upgraded its Coding Plan into a Token Plan—positioning it as a unified subscription across modalities. For Volcano Engine, the mass adoption angle is increasingly important: a company source told LatePost that token consumption from individual users grew about 16-fold over the past month.
Price hikes test whether scale can translate into profit
As token volumes surge, leading cloud vendors have begun signaling or implementing price increases for AI services. LatePost said AWS and Google Cloud pointed to higher pricing earlier, followed by moves in China from Tencent Cloud, Alibaba Cloud and Baidu AI Cloud. For investors, the near-term question is whether higher list prices can offset inference-driven GPU burn and networking costs—especially as multimodal generation (images and video) expands.
The operational challenge is also shifting. Volcano Engine executives have argued that agent-era cloud platforms must support long-term memory, self-evolution with business data, and reliable execution of complex multi-step tasks—capabilities that demand tighter integration across model endpoints, vector databases, workflow orchestration and security tooling. Those requirements could reward vendors that can bundle MaaS with broader PaaS and data services, echoing the Microsoft Azure playbook where model APIs become an entry point for larger cloud spend.
Related Coverage:
Doubao AI Assistant Sparks Struggle for System Control with Major Chinese Apps
ByteDance's Doubao Joins AI Red Envelope Battle with Spring Festival Campaign