Volcano Engine, wants to create a cloud that belongs exclusively to the AI era
On December 18-19, the Volcano Winter Powertrain Conference was held in Shanghai. At the conference, Volcano Engine officially released the 1.8 version of the Bean Bun Large Model and updated the Bean Bun Image Creation Model Seedream to version 4.5, bringing users richer creation tools and lower threshold AI application landing solutions.
However, the most closely watched by the market is the token usage of the Volcano Engine.
According to Tan Dai, President of Volcano Engine, as of December this year, the daily average tuning volume of the model has exceeded 50 trillion times, an increase of more than ten times compared to the same period last year, and an explosive growth of 417 times compared to the initial release.
This is also considered by the market as the best example of Volcano Engine's good "realization" that it is cloud native AI.
However, behind everyone's astonishment at this report card, there is also a new consideration: how much tangible revenue can the exponentially increasing token consumption contribute to the volcano engine's report? Can AI, which claims to be a volcano engine for AI native cloud, overcome the trap of heavy manpower in domestic SaaS by utilizing AI?
How much real gold and silver can be exchanged for trillions of tokens?
To understand the ambition of a volcanic engine, one must first see its' origin '.
On the table of the domestic cloud market, the Volcano engine is a 'outlier'. Unlike Alibaba and Tencent, who are burdened with the heavy burden of "traditional cloud", the volcano engine has almost no historical burden - it was born in the dusk of mobile Internet, but just caught the dawn of the big model.
This' tardiness' has instead become a strategic dividend. Because there is no outdated data center architecture that needs to be compatible, Volcano Engine has been living like a special forces soldier in the AI era since day one: it does not want to be a traditional "rental company" that relies on selling storage and bandwidth to make a living, it wants to be the absolute dealer of MaaS (Model as a Service).
This is also the logic that President Tan Zai repeatedly preaches on various occasions: the Volcano Engine is a cloud born for AI, and its ultimate form is to sell model calls and intelligent services.
Nowadays, the market has indeed bought it. The daily average tuning volume of 50 trillion is not only an astronomical figure that catches the eye of peers, but also like a volcano engine running the first "highway" on the AI cloud track. But once the excitement subsides, rational investors and observers will eventually pick up their calculators and calculate the most realistic account.
How much revenue can the consumption of these 50 trillion tokens bring to Volcano Engine?
December 2024: The daily average call volume is only 4 trillion times.
April 2025: Rapidly climbing to 12.7 trillion times
By May 2025, it will reach 16.4 trillion times.
August 2025: Exceeding 25.9 trillion times.
October 2025: Steady growth to 30 trillion times
December 2025: Finally announced at the Powertrain Conference that the daily average call volume has officially exceeded 50 trillion times
If we connect these points and assume that the consumption is steadily increasing within the statistical interval, then by 2025, the total token consumption of the Big Bean Bun model will have exceeded 9000 trillion, approaching 10000 trillion.
However, this seemingly massive amount of data has not been converted into substantial revenue, and the core issue lies in the fact that the unit price of tokens is simply too low.
In order to break through a gap in the fiercely competitive cloud market, Volcano Engine has almost been sweeping the market with a "suicidal" price reduction strategy in the past year.
Last May, Volcano Engine was the first in the industry to engage in a price war, lowering the inference input price of the main model of tofu buns to 0.0008 yuan/thousand tokens, which was more than 99% lower than the industry price at that time. In the following year, it also continued the logic of "collapsing" price reductions.
This top-down pricing logic, although allowed the Volcano Engine to attract a massive number of users in a very short period of time, also means that it must face an extremely cruel reality: tokens have transformed from expensive "luxury goods" to cheap "commodities".
So, how much revenue will one trillion tokens bring to the Volcano Engine in 2025? Based on the current unit price on the official website of Volcano Engine, we have calculated that after integrating the prices of online inference, online inference context caching, and batch inference, the price per thousand tokens is approximately 0.0009 yuan.
According to the current unit price calculation, the annual total adjustment amount of "ten trillion" times can theoretically support a revenue space of nearly ten billion yuan for the Volcano Engine.
However, this 9 billion yuan is by no means the real gold and silver that ultimately fell into the bag.
In the early stages of market promotion, in order to gain market share, the vast majority of tokens were actually "given away for free". Under intense price wars and ecological subsidies, the proportion of paid traffic that truly generates deductions is extremely low. If we boldly assume that paid tokens only account for 10% of the total, then the actual MaaS revenue of Volcano Engine in 2025 will only be around 900 million yuan.
In sharp contrast to MaaS revenue, the revenue of Volcano Engine has exceeded 20 billion yuan (approximately 24 billion to 25 billion yuan) by 2025.
This means that although Token call volume appears to be impressive in official terms, at this stage, the cornerstone of Volcano Engine's revenue is still IaaS, PaaS, and computing power leasing business, rather than pure AI calls.
This "top heavy, bottom light" data structure clearly indicates that the ideal of AI native cloud is very rich, but the reality of commercialization is still rigid.
So, in the upcoming year of 2026, the volcano engine must face the question of how to quickly drive token consumption to continue exponential growth when tokens are completely reduced to cheap infrastructure like electricity. Through economies of scale, MaaS can become the true revenue pillar and become the volcano engine that Tan Dai describes to the market?
When the agent encounters the traditional party A, does the volcano engine also need to send personnel to station on site?
If selling tokens is selling electricity, then the Agent promoted by Volcano Engine through platforms such as Coze is intended to directly sell customers a set of "automated factories".
In Tan Dai's blueprint, the ideal path is extremely light: the volcano engine is built with a base and tools, and customers only need to develop their own business adapted agents on the platform like building blocks. This is essentially reshaping the delivery logic of SaaS, from the past 'vendor feeding' to the current 'customer self-sufficiency'.
However, in the domestic business environment, this "lightness" often means some kind of huge challenge: how can you persuade a party that is accustomed to "paying by head" to pay for an invisible and intangible algorithm logic instead?
The deepest pitfall that the domestic SaaS industry has experienced in the past decade is the vicious cycle of "non heavy manpower investment without payment". Due to the low willingness to pay for software in China for a long time, the few corporate customers who are willing to pay often believe in an extremely simple logic: if I spend money, I have to see your people.
So, we see countless SaaS vendors who promote standardization, but in the end, they all become "decoration outsourcing teams".
In order to win a large order, the second party often needs to send a large delivery team to the first party's site, carrying computers and staying up late to accompany the other party to modify code and adjust interfaces. The sense of security brought by this "crowded crowd" was once the psychological cornerstone for SaaS transactions in China.
This model is not only heavy, but also extremely inefficient. Once entering the project-based system, software businesses that originally had diminishing marginal costs become physical labor with constant marginal costs. As a "latecomer", Volcano Engine is well aware that if it follows other cloud factories to engage in this "human flesh delivery", it will not only be difficult to surpass, but also make its advertised "AI native" background disappear completely.
So, the strategy of Volcano Engine is to "instrumentalize counterattack". It not only pushes out a large model, but also a complete set of Agent development kits. Its underlying message is: Don't ask us for hundreds of engineers anymore, give you a tool that is useful enough, and your own employees can handle it.
But this is precisely the focal point of the contradiction. The first party, who is accustomed to "opening their mouth when eating", finds it difficult to quickly adapt to this identity shift from "buying solutions" to "learning tools" in a short period of time. For most traditional enterprises, AI is still a black box. Although the threshold for Agent development has been lowered, there is still a barrier between "being able to run smoothly" and "being able to implement business".
Moreover, as a new phenomenon, the stability, logical loop, and understanding of complex business scenarios of agents are in the painful period of evolving from "semi-finished products" to "finished products". It is extremely difficult to achieve the SaaS software level that was manually polished by dozens of programmers and product managers in the past for a newly released agent.
When Party A found that the Agent they had worked on for a long time was still "artificially intellectually disabled", their most natural reaction was still: Volcano Engine, can you send someone who understands the industry to help me tune it?
This creates a huge paradox: the more Volcano Engine wants to promote standardized and low threshold agents, the more it realizes that in the actual "last mile" of implementation, it still needs a large number of professionals to enter the field to bridge the gap between technology and business.
If the Volcano Engine cannot achieve true "human like delivery" in technology, or cannot cultivate a large third-party partner ecosystem that can share physical labor for it, then its advertised AI native cloud is highly likely to involuntarily fall back into the vicious cycle of "stacking ten people to earn a hundred yuan" when moving towards deeper waters.
At that time, the ones supporting an average daily call volume of 500 to 500 trillion may no longer be just the servers that synchronize and jump in the background, but also the countless exhausted delivery engineers who run around in front of the platform. This situation of "new bottled old wine" is probably not the final answer that Tan Dai wants to deliver to the market.
Volcano Engine wants to be the cloud of the AI era, which is indeed a sexy vision. But the tough battle on this path from the frenzy of "trillions of tokens" to the deep cultivation of "real profits" has just begun.
When the myth of technology meets the common sense of business, the Volcano Engine needs to prove not only how fast its model is, but also how strong its endurance is for the domestic business ecosystem.
