Half-year market tops 10 billion; big firms reassess token‑sale business

Half-year market exceeds 10 billion, big firms re‑evaluate the Token‑selling business

Cover source | Public internet, “The Railway Bridge At Argenteuil”, Claude Monet, 1873

In 2011, Marc Andreessen, co‑founder of the Silicon Valley venture firm Andreessen Horowitz (a16z), coined the famous line “Software is eating the world,” marking the start of the SaaS decade. Six years later, Jensen Huang added the second half: software is eating the world, but AI will eat software.

However, even two years ago, most people didn’t believe that selling tokens could be a good business. By 2026, that has become reality.

Recently, IDC officially released its H1 2026 report. In the first half of 2026, China’s public‑cloud MaaS market generated revenue of ¥10.63 billion, a year‑over‑year increase of 889.4%.

When the public‑cloud market has been slowing its growth for years, only a handful of markets can reach a scale of tens of billions within two years.

MaaS has become a key indicator for tracking the development of the global AI market. It is often likened to “water, electricity, and coal”: model providers offer services via APIs, allowing enterprises or individuals to use models for text generation, code writing, and other tasks without building their own models or compute infrastructure. Providers charge based on token usage, compute time, or subscription plans—more model calls generate higher revenue.

In the new 2026 season, Volcano took the lead—according to IDC’s “China AI Software Market Semi‑Annual” report, in H1 2026 Volcano accounted for nearly half of the domestic public‑cloud MaaS revenue market.

From the discussions around MaaS metrics, we can also glimpse the shift in market consensus.

In 2025, the industry was still debating which cloud had higher token sales and better cost‑performance in the MaaS market; by 2026, facing a market that grew almost ninefold in a year, everyone’s focus metric gradually converged on: MaaS revenue.

Anthropic and OpenAI remain the “model students” for Chinese vendors. The recent buzz around Anthropic’s IPO prospectus stems from its commercial growth once again surpassing market expectations—by the first half of 2026, Anthropic’s total revenue reached $16.286 billion, with 80% coming from inference revenue billed by token usage.

The token consumption of agents is increasing exponentially, also driving up Anthropic’s average revenue per user.The prospectus shows that in Q3 2026, enterprise users who did not use Claude Code spent less than $75 per month on average; the top 5% high‑spending enterprise users who used Claude Code spent about 20 times more than the former.

This growth is far from peaking, prompting aggressive investment from domestic tech giants. The AI office competition that started mid‑year—featuring products like Doubao Work, Qianwen Office, and WorkBuddy—is essentially large companies vying for high‑value productivity‑focused task scenarios.

Today, industry consensus has become unprecedentedly clear: build top-tier models. This is the best way to quickly increase token consumption and get customers willing to pay for it.

The commercial potential of large models is also being unlocked. The ceiling of traditional cloud computing is largely constrained by enterprise IT budgets, while MaaS targets productivity— in theory, a country’s MaaS market size will ultimately be linked to its GDP and population. By this logic, the domestic market is only a part of the global market; an industry insider estimates that overseas market potential is at least three to four times larger than the domestic one.

We have also exclusively learned that starting from the second half of this year, Volcano will focus on expanding into overseas markets, and has already launched an aggressive commercialization process. By the end of the year, the overseas business revenue share of Volcano is expected to approach 40%.

A business more like “water, electricity, and gas” than cloud computing

Why will 2026 become a pivotal watershed for the MaaS market? One phenomenon is that the market is shifting from pursuing token consumption to pursuing revenue.

The focus on token consumption stems from the market still being in its early stage. Before 2025, most AI tasks were concentrated on simple dialogues like AI search, with limited interaction rounds and conversation depth.

At that time, most vendors’ primary task was to increase token consumption. After DeepSeek R1 exploded in popularity in 2025, domestic vendors quickly followed suit, whether by integrating as many third‑party models as possible, offering subsidies, or cutting prices—essentially all aimed at getting users to start using the service.

The turning point comes from Anthropic’s Claude Opus 4.6, which allows the model to finally operate in production‑grade coding and agent scenarios—planning tasks, sequentially invoking tools, and recording execution states. A single task may call dozens of tool interfaces in sequence, and token consumption will increase exponentially until the task is completed.

When a task shifts from a short conversation to a long‑running process, the model sells not just a single answer but becomes an actual service. Model providers, charging by token, can more easily align with high‑value work tasks.

The change in interaction modality also alters the growth and business logic of MaaS.

“From the perspective of language models, customers’ task characteristics are no longer divided by industry or modality; customers usually run any task, so Volcano’s pricing segmentation this year is also based on the input length of the task, not on customer type,” a Volcano representative said.

Therefore, the MaaS market is further maturing.

One sign is that after Coding opened up productivity scenarios, each model vendor chose an ecosystem and product positioning that better suits them, which has become a key market competition strategy.

For example, after the flagship model K2 was released, Kimi continued on an open‑source path and was the first to scale open‑source model parameters to 3 T, gaining huge market influence.

But Zhipu chose to focus on the Coding scenario, launching a model in the comfortable range—hundreds of billions of parameters—yet with faster inference speed, capable of handling most programming tasks and achieving rapid commercial deployment.

In the first half of 2026, Zhipu’s revenue grew nearly fourfold, climbing to 954 million yuan, with over 80% coming from API calls.

Although the Seedance series models have maintained the global video SOTA for nearly eight months, most AI industry professionals we talk to say that, regardless of modality, there is no absolute moat.

This is also the most fundamental difference between MaaS and traditional cloud computing. In the traditional cloud era, vendors could keep customers’ data on the platform through storage and database products, making migration costs high; but in the MaaS era, if the model invocation chain is encrypted, the platform can hardly lock customers by binding data. If another vendor releases a stronger SOTA model, customers may switch models at any time.

In other words, compared with cloud computing, MaaS is more like a pure “utility” business—water, electricity, gas. To retain customers, the only way is to truly embed the model into the enterprise’s production pipeline—first launch a sufficiently SOTA model and roll it out quickly; then users will apply it to more challenging tasks.

36Kr learned that before its launch, Seedance was positioned only as a creative scenario, but it quickly reached SOTA in the creative field. After achieving commercial coverage, some customer-initiated scenario innovations have also emerged—many embodied companies use Seedance for corner case (extreme situation) data processing.

When the call volume increases, real user feedback also grows, feeding back into inference engineering and product design, allowing us to release a better next version of the model before competitors catch up.

Not only pursuing revenue growth

But also pursuing intelligence density

MaaS is more like the utilities business than cloud computing; the market’s explosive growth rate now appears even more intense than the previous era: China’s public cloud market started in 2009 and took seven years to surpass the 10‑billion‑yuan annual scale threshold; today the MaaS market has halved that time.

However, another frequently discussed issue is: MaaS is more like utilities (water, electricity, gas), seemingly harder to create customer stickiness than traditional cloud. Why do big companies still pour money into it at any cost?

On one hand, the ceiling of traditional cloud computing is essentially limited by a company’s IT spending budget; when AI enters core production areas like coding, video creation, and daily office work, MaaS targets the productivity budget—companies are not spending money to improve IT efficiency, but essentially purchasing a batch of virtual digital employees.

Silicon Valley research firm SemiAnalysis surveyed more than 50 companies and found that most Fortune 500 firms still spend less than $100 per employee per year on AI, while top high‑intensity users spend over $90,000 per employee annually—a gap of nearly 660 times. This means the true explosion of token consumption on the enterprise side has not yet arrived.

MaaS is expected to dramatically raise the market ceiling. Most industry insiders believe that the MaaS boom will be even more intense than cloud computing. For a long time, the MaaS market will remain in a seller’s market phase.

In most cases, AI compute supply determines the overall growth rate of the MaaS market. “As long as the model can achieve SOTA, the amount of intelligence it produces can be consumed proportionally.” said an industry insider.

Regarding MaaS, while many cloud providers still treat it merely as a hook to drive traditional large cloud deals, Volcano has been positioning MaaS as its sole bet since 2023, believing it can generate a market far larger than traditional cloud.

Because model iteration is so fast, to avoid spreading engineering adaptation effort, Volcano has not released many model sizes. Whether for language or vision models, Volcano now focuses on only two model sizes: a large flagship model, the top priority for exploring the intelligence ceiling; and the Flash model, which handles scaling and quickly expands commercialization.

In 2023, while ByteDance’s self-developed foundational models were still in their early stages, Volcano already introduced a large number of third‑party models to tap more application scenarios. Models of the scale of Flash, although not generating the highest revenue, have already been called several times more than the flagship model on Volcano, shouldering broader productivity penetration and the exploration of new scenarios.

In this process, Volcano also pre‑aligned the collaboration chain between the model engineering, application, and commercialization teams.

“Now MaaS revenue is Volcano’s top priority, and all resources are being reallocated for it,” an industry insider joked. “You can’t even sell GPU resources; the cards are all being used for inference and training.”

A Volcano insider recalled that starting in Q4 2025, after recognizing the potential of productivity scenarios, Volcano quickly shifted its strategy, making revenue—not token consumption—the team’s primary metric.

Around productivity scenarios, Volcano’s organization and sales teams are rapidly forming. After integrating Feishu, Doubao Work, and Doubao’s product team in July 2026, Volcano Engine recently completed another internal integration, establishing a “Creativity Service Platform,” merging Volcano and Feishu’s sales teams so that a single salesforce sells all products, including SaaS and token‑related offerings.

The series of adjustments signals that, with the opening of productivity scenarios, SaaS and token businesses are becoming increasingly convergent—users want delivered task value rather than just a tool product seat.

When it comes to commercialization, Volcano’s pricing and product principles are also quite stable: first, there must be healthy gross margins; second, they must be competitive in the market; and finally, differentiate based on different task requirements.

Seedance’s remarkable leap is one of the most noteworthy cases to reference in 2026.

Within just a few months, Seedance has already achieved extremely high coverage. In addition to reaching SOTA in the video domain in terms of parameters and generation quality, many of Volcano’s engineering efforts have also accelerated the establishment of a competitive advantage.

When Seedance 2.5 launched, Volcano highlighted the “Draft” (sample) feature, based on an observation from the Volcano team’s breakout success with 2.0: professional creators rarely get a satisfactory shot on the first try and often have to draw cards repeatedly. Trying each time in 1080p or even 4K is too costly, which hinders continued use. After Draft was introduced, users first generate low‑resolution previews with a cost‑effective model, then pay for the final high‑resolution output, saving the cost of most ineffective draws.

When task intelligence density is higher, the value of intelligence also rises. Because inference costs continue to be optimized, this does not erode revenue. Ultimately, for the same amount of finished output, the total expense using Seedance 2.5 is generally lower than using 2.0.

36Kr learned that Seedance 2.5 Draft has become the most deeply intensive operational mode for users, greatly boosting user stickiness to the model.

Now, even with their own flagship models, major firms have become unprecedentedly open at the application entry point—on October 29, Doubao Work opened third‑party large‑model services to enterprise customers, initially integrating models such as Kimi 3.1 Pro, GLM 5.3, and DeepSeek V4.1 Flash.

It’s not just Bytedance; Alibaba’s Qianwen Office also launched GLM-5.3 and DeepSeek V4 Pro in August, alongside its own Qianwen models; Tencent’s WorkBuddy has also supported integration with external models like DeepSeek and Kimi from the start.

Whether self-developed or third‑party, the underlying consideration is converging: the big players are trying to keep complex tasks within their own products—this is not only to increase current token consumption but also to use high‑frequency productivity entry points to identify which scenarios truly merit heavy investment and have higher intelligence density.

“We are now integrated with the model team,” said a Volcano Ark representative. As the competition for large models returns to the flagship model arena, to keep winning in this business, the big companies’ contest is no longer just about whether the next‑generation model can lead, but about comprehensive competition spanning models, inference engineering, and commercialization systems.