Volcano Engine, the cloud arm of ByteDance, has seen its daily token consumption surge 50% to 180 trillion since end-2025, a 1,500-fold increase from two years ago, as its MaaS business enters core production scenarios. In an interview with 36Kr, Volcano Engine President Tan Dai said the new flagship model Doubao 2.1 Pro has reached usable standards, particularly in coding and agent capabilities, marking a milestone for the company.
Doubao 2.1 Pro: 'Finally on the table'
At the Volcano Engine Force conference on June 23, ByteDance unveiled its next-generation flagship model, Doubao 2.1 Pro. Tan Dai described the model as finally being "on the table" in terms of coding and agent abilities. On the Terminal Bench coding benchmark, Doubao 2.1 Pro is roughly on par with Claude Opus 4.7, performing well on long-horizon and complex tasks, meeting the threshold for production use.
Tan noted that coding and agent progress means models can enter core production workflows for enterprises and individuals, creating more commercial value. The model also powers the new task mode in the Doubao app. Alongside the flagship, Volcano Engine released Seedance 2.0 4K, Seedream 5.0, and a voice generation model 1.0, with Seedance 2.5 expected in July.
Seedance: From entertainment to production
Seedance 2.0 has become the first video generation model globally to unlock commercial production scenarios, according to Tan. Before its launch, most video models were used for UGC or PGC entertainment content, but usage patterns have shifted: weekday load now exceeds weekend load by more than double, indicating real work usage. Seedance is also being applied in embodied AI, industrial manufacturing, and autonomous driving for data synthesis and scenario simulation.
Tan emphasized that video generation is a viable path to world models, and Seedance's accurate understanding of the physical world enables high-quality visual data synthesis. He dismissed the idea that the video generation race is over, noting that AI penetration in video is still very low, and the technology's value extends beyond short-term revenue.
MaaS and cloud: Not opposing forces
Tan rejected the notion that MaaS and cloud are in conflict. "MaaS is part of the cloud," he said, envisioning a future where agents orchestrate IaaS, PaaS, and SaaS components. He argued that AI-driven cloud could be 10 to 20 times larger than traditional cloud, which he likened to investing in PC search in 2012 when ByteDance was founded.
On customer stickiness, Tan compared MaaS to early cloud computing when selling virtual machines had little stickiness. As models integrate deeper into core production systems, coupling increases. He noted that Volcano Engine's "trillion club" of customers consuming over a trillion tokens has doubled to 200 since December 2025, with nearly zero churn.
Pricing and profitability
Tan explained that pricing is not based on cost-plus but on value created. In 2024, price cuts were justified because models were only capable of chatbot functions. Now, with models entering production, higher prices are warranted. He dismissed external reports of 70% profit margins for Seedance, saying the company hasn't calculated that figure. Volcano Engine's priority remains profitable scale, not unprofitable growth.
The road ahead
Tan acknowledged that China's LLM progress in coding lags behind global leaders like Anthropic and OpenAI, but he expects catch-up by Q2 2025. He also stressed the need for a "middle layer" — harnesses tailored to industries — to fully unlock model capabilities. For video creation, Volcano Engine is hiring directors to build domain-specific tools, such as the new "3D white-box preview" feature for Seedance.
Looking forward, Tan said the company has raised its MaaS revenue target for the year, though he declined to specify figures. The ultimate goal is to build the best AI cloud company, with models as the core engine.