Volcano Engine President Tan Dai: Price hikes for large models are not industry behavior, and "Lobster" user numbers are far from reaching the ceiling

Ask AI · How does Token pricing reflect the true value of model capabilities?

“To judge the industry stage, you can work backward from the endgame. We can calculate the revenue scale that all Tokens (word tokens) will generate in the future, then compare it with the actual revenue of the global industry today. Mapped onto a 42-kilometer marathon, it clearly shows where we stand.”

On April 2, in an interview with media outlets including The Paper, Tan, President of Volcano Engine, said that the competition in the AI track is currently only at “the first kilometer of the marathon,” while his judgment last year was “500 meters,” and overall, “the industry is still in a very early stage.”

In this year’s “Hundred Shrimp Battle,” which kicked off at the beginning of the year, Volcano Engine launched the “Lobster” product ArkClaw. “Currently, the number of users for lobster-type products in China is only in the millions. If everyone has one or two lobsters in their hands, the market potential is enormous.” When talking about the market for the “Lobster” product, Tan said candidly.

That day, Volcano Engine disclosed that as of this March, the Doubao large model’s average daily Token usage had exceeded 120 trillion. In the past three months, it has doubled, which is 1,000 times compared with when it was released in May 2024. Currently, the number of enterprises with total Token usage exceeding one trillion on Volcano Engine has grown from 100 at the end of last year to 140.

“Raising shrimp” isn’t expensive because Tokens are expensive

Regarding the recent industry hotly discussed trend of large-model price increases, Tan said directly that price hikes are just a market behavior by some vendors. Within the industry, there are also vendors pushing for price reductions. The core difference lies in the logic used to judge Token value.

“The difference in Token prices, in essence, is the difference in capabilities they carry.” Tan said that next-generation model capabilities are stronger, so the cost per Token will increase somewhat, and the economic value they can create will also rise in step. “The price increases brought by improvements in the model’s intelligence are, in essence, because they can create greater value for customers.”

In Tan’s view, the industry’s widespread anxiety about “Tokens being too expensive” has a core misconception: focusing only on the per-Token price while ignoring the overall end-to-end cost to complete tasks, as well as the large volume of invalid Token consumption. “Many users report that Token consumption for the Lobster product is fast. The core problem isn’t that a single Token is expensive, but that when it completes a task, it makes a large number of invalid attempts—more than half of the Tokens are consumed in pointless exploration in order to find the final solution.”

“What matters isn’t how much a single Token costs, but whether spending this money can create value.” Tan believes that if the cost per Token is low but the model capability is insufficient—requiring the consumption of 10 times or even 20 times more Tokens to complete a task—then in the end it will only cause even greater waste.

It is worth noting that the explosive popularity of products like OpenClaw (lobster) has also come with extensive controversies over security and other risks: “Is raising shrimp truly something everyone needs?”

In Tan’s view, most ordinary people can indeed fully meet their needs by using the Doubao app. But even if the core audience for Lobster is enterprise employees, individual developers, and the geek community, the market size is still not something to be underestimated—corresponding to a population base that could reach hundreds of millions, with a high ceiling.

He revealed that Doubao 2.0 has made a large number of targeted optimizations for Agent scenarios, especially breakthroughs in VLM visual understanding capabilities, which provide key support for agents to handle complex multimodal tasks. On the security side, Volcano Engine has built a full-chain security detection mechanism covering every stage, including prompts, the Agent runtime environment, skills, and more—making it the first and only company in China that has obtained two relevant security certifications from the China Academy of Information and Communications Technology. On the ecosystem side, Volcano Engine has officially partnered with ClawHub to support and sponsor the building of its China mirror site, addressing the pain point of access latency for domestic developers, while also sticking to an open ecosystem instead of building a private system.

Aiming for 100-billion revenue targets may be achieved early

Two years ago, when Volcano Engine took the lead in pushing into the MaaS (Model as a Service) track, the mainstream view in the industry was not optimistic. People generally believed that Token businesses had poor stickiness and low gross margins—essentially “losing money to attract attention.” Today, almost all cloud providers and large model companies have moved into MaaS, and multiple companies have also adjusted their organizational structures and market strategies in a targeted way.

Tan said that the whole industry moving into MaaS is a positive for the industry. Sufficient market competition can not only drive technological iteration, but also expand the market pie so that more companies can benefit from the AI technology dividend.

Facing increasingly fierce market competition, Tan said bluntly that Volcano Engine’s core advantages have never been price discounts. “When enterprises choose MaaS services, the key is whether the model can get the work done and solve problems. No matter how low the discount is, if the model can’t do the work, then it is of no value to customers.” He said that feedback from the market for multimodal products such as Seedance shows that generational differences in model capabilities are the core yardstick for customer choice. He disclosed that currently, Volcano Engine’s MaaS market share is close to 50%.

Worth noting is that Volcano Engine once set a target of having MaaS business revenue exceed 10 billion yuan in 2026. Regarding this, Tan revealed that last year they revised their business plan and raised the target by quite a lot. “At the current development speed, the (100-billion-yuan annual revenue) target will very likely be achieved ahead of schedule.”

Copyright protection needs to adapt to the new AI paradigm

As a core product in Volcano Engine’s multimodal layout, Seedance 2.0 has attracted widespread industry attention since its self-testing.

An even more notable industry trend is that with the launch of Seedance 2.0, it is changing the industry norm of “mixing multiple models” in content production. Tan said that previously, the industry mixed multiple models mainly because the capabilities of different models were close, and there was no product that clearly led. But when a model’s capabilities achieve a generational breakthrough, users naturally choose the best solution, and their reliance on a single high-quality model will increase significantly.

“This is an inevitable stage in the development of the AI industry.” Tan believes that three years ago, the industry only talked about model parameters, and no one truly put them into use. After that, it entered a stage of comparing based on usage volume. With limited model capabilities, each had its own strengths, so mixing multiple models was the optimal solution. Now, more and more enterprises have already run through a closed loop for using general models to create business value, and demand for models is shifting toward value orientation.

This time, the Seedance 2.0 API was delayed and only began the public trial testing at the moment—it’s the core bottleneck of copyright protection. Tan believes that copyright protection is a dynamic process, and Volcano Engine has always placed it at the core. Only after building a complete and comprehensive copyright protection system can the model capabilities be officially opened up externally.

“Traditional copyright protection technology is already outdated.” Tan said that in the AI era, IP protection needs to cover all-scenario rights and interests of characters across different ages, different looks, and different art styles; traditional technology cannot meet this need. And based on Doubao’s strong VLM visual understanding capabilities, Volcano Engine has already built a whole new copyright IP protection mechanism that covers the entire chain, including prompts, intermediate outputs, and final generated outputs.