AI token takeover in China

[BEIJING] Tokens, the basic units used by large language models to process text, are becoming an important measure of competitiveness in the artificial intelligence (AI) industry.

As coding tools and AI agents become more prevalent, demand for tokens is increasing far faster than many business executives expected, and competition for pricing, computing power, business customers, and new tiers of AI infrastructure has begun across China’s AI industry.

In the early days of AI chatbots, a typical user interaction included only a few hundred Chinese characters, said Li Boxun, chief technology officer at AI computing services company Infinigence AI.

As inference models became more popular, one exchange grew to around 10,000 tokens. This year, he said, the average has risen to more than 50,000 tokens.

For Chinese model developers, this surge is already leading to rapid revenue growth.

Zhipu AI announced that its application programming interface revenue increased 60 times year-on-year in the first quarter.

Alibaba Cloud announced that its token revenue has increased more than 15 times since the beginning of 2026. ByteDance’s Doubao model exceeded 180 trillion token calls per day in June.

This boom is putting a strain on the AI ​​supply chain. Since August 2025, a wave of data center construction has tightened supplies of memory chips, graphics processors, central processors, optical modules, and liquid cooling systems.

Prices for some memory chips have more than doubled, and next year’s capacity is nearly sold out.

For now, the benefits of the token economy are still concentrated in the upstream. In 2025, Nvidia and the world’s three major memory chip manufacturers generated combined operating profits of more than US$160 billion.

In contrast, model companies face high training and inference costs and are still in a stage of high investment and low profits, with some companies still in the red.

“If tokens are destined to become the next generation of power, data centers are power plants, and large models and cloud companies are power grids,” a Chinese executive of an international cloud company told Caixin.

“Companies are not just looking for tokens that are easy to use or cheap; they are looking for value for money.”

Caning pricing

For cloud and model companies, the token boom is creating a pricing dilemma. Raising prices may encourage the rationing of scarce computing power and protect profits, whereas lowering prices may stimulate usage, lock in customers, and expand market share.

Chinese cloud providers such as Tencent Cloud price increase The core model saw an increase of over 430% in March.

In March, Alibaba Cloud announced that it would raise prices for some AI computing services by 5% to 34% and increase prices for high-performance storage systems by 30%.

Model makers followed suit. Zhipu AI CEO Zhang Peng said in an earnings call that Zhipu’s API price rose 83% sequentially in the first quarter, but call volume still surged more than four times.

On June 23, ByteDance released Doubao 2.1, doubling the price of the previous version. Doubao app for consumers also introduces Pro subscription plan From 68 yuan (10.03 USD) per month.

Cui Tingting, research manager at IDC China, told Caixin that cloud computing companies are adjusting prices due to the lack of supply.

He said the surge in downstream demand has led cloud vendors to increase investment in AI, while leading to shortages of key hardware and longer delivery times.

Alibaba Group CEO Eddie Wu told investors that customers are increasingly willing to accept token prices as AI applications move from chatbots to agents.

At the same time, server costs have more than doubled since January 2025, he said, adding that in the next one to two years, the prices of models and other cloud services will continue to rise, which will lead to improved gross profit margins for Alibaba Cloud.

However, many industry executives said the price of Caixin tokens must eventually fall for AI to be widely adopted. Hu Jian, co-founder of Chinese AI inference startup Beijing Silicon Flow Technology, told Caixin that the token price will inevitably fall as chip costs decline and technology advances.

He said the tipping point will come when there is a breakthrough in the performance and supply capacity of domestically produced chips.

DeepSeek, which previously led a token price war in China, lowered its price again in May.

The price of DeepSeek V4 Pro was a quarter of the original level. The company announced in late June that it would launch the official version of DeepSeek V4 in mid-July and introduce peak and valley pricing. Even though the intraday peak price is double the normal level, this model still maintains a clear price advantage.

An inference service provider told Caixin that DeepSeek can reduce inference costs by keeping compute power utilization high and scaling up, thereby lowering prices.

“Users are still very tolerant of DeepSeek, even if heavy operations cause some instability in the stability of the service,” the official said.

middle moment

Concerns about token supply and cost are creating new value for companies that sit between application and underlying models.

These “middle tier” providers include model aggregation platforms, inference service companies, memory management systems, security and risk management providers, and AI infrastructure companies that optimize deployments across a variety of chips.

Jia Anya, head of the Raccoon agent application family at SenseTime Group, said the gap between enterprise demand and model enterprise capabilities creates room for middle-tier providers.

Balancing hardware supply is still years away, he said, but model companies are primarily focused on improving performance and may not have the resources to focus on reducing inference costs.

“Business users are very sensitive to price increases of 20% to 30%,” says Jia. “They need an AI infrastructure with a controllable budget, stability, and value for money.”

SiliconFlow, which recently filed to list in Hong Kong, said average daily token throughput on its platform increased from 47.8 billion in December 2024 to 578.5 billion in April.

Revenue in 2025 reached 55.3 million yuan, an increase of more than 650% from the previous year.

SiliconFlow’s flagship product, SiliconCloud, provides fine-tuning, hosting, and deployment services for over 170 large-scale models.

SiliconFlow announced on June 16 that it has completed a Series B funding round of over 2 billion yuan, the largest ever for a Chinese AI middle-tier service provider, valuing the company at USD 7.7 billion.

According to IDC data, SiliconFlow ranked 4th in China’s Model-as-a-Service market in 2025 by token call volume. ByteDance’s Volcano Engine ranked first with over 40% share, followed by Alibaba Cloud in second place with around 30%.

Other AI infrastructure providers are also attacking various bottlenecks.

Qingcheng.AI built AI Ping, a model calling platform that evaluates models from providers. It also provides price comparisons and smart routing to help developers select and invoke models more efficiently.

“The hardware determines the ceiling height, and the software tries to get as close to that ceiling as possible,” said Shi Tianhui, co-founder of Qingcheng.AI.

Even with the same computing power and model, different infrastructure software can lead to changes in hardware utilization, model intelligence, token production efficiency, and provider margins, he said.

Another mid-tier company, Memory Tensor, focuses on long-term memory systems for AI agents.

Founder and CEO Xiong Feiyu said that 50-60% of the company’s demand comes from industrial and financial agencies, and the rest comes from emotional companionship, gaming, and AI hardware.

AI starts work

Coding is the biggest source of direct token consumption by agents. However, within a company, office workflows offer greater benefits.

Analysts at China International Capital said enterprise office applications could break old patterns of infrequent, lightweight and experimental interactions.

When AI agents connect to a company’s core data systems, token usage is no longer limited by human interaction frequency and scales with business flows.

Alibaba, Tencent, and ByteDance are all trying to make office software an entry point for AI. Chen Yusen, vice president of Alibaba Cloud Intelligence and founder of AI workspace MuleRun, said that AI office products are transitioning in two stages.

During the co-pilot stage, AI saves time, but humans still perform formatting changes, sending emails, and other steps. At the AI-native stage, agents will deliver the entire project, he said.

Chen said on May 20 that MuleRun launched overseas and gained paying users in 43 countries within two months. The 34-year-old later became DingTalk’s CEO, replacing DingTalk founder Chen Han, making him the youngest head of a business unit at Alibaba.

On July 2, Alibaba integrated three office agent products. QoderWork, a programmable agent desktop app. Wukong is an enterprise AI work platform developed by DingTalk. And mule run.

The changes come as DingTalk, which has the largest share of China’s office software market with more than 800 million users, struggles to maintain its leadership in the AI ​​agent cycle.

Tencent’s WorkBuddy was generally available in mid-March and supports access to WeChat, WeCom, QQ, Feishu, and DingTalk, making it one of the fastest-growing AI office products this cycle.

In its first quarter earnings report, Tencent said that active user retention rates for CodeBuddy and WorkBuddy exceeded 60%, and paid user retention rates exceeded 80%.

According to research firm Analysis, WorkBuddy ranked first among PC-based AI-native office agent platforms in China in March, with 8.85 million monthly visitors and an 831 percent month-on-month increase in users.

This was followed by ByteDance’s Trae, Tencent’s personal QClaw app, Alibaba’s QoderWork, and Tencent’s CodeBuddy.

ByteDance’s Feishu connected Agent Aily to workplace scenarios in June, allowing users to create charts, build web pages, respond to document comments, respond to emails, and handle formatting with natural language prompts.

Feishu said companies such as Seres Group, Unitree Robotics, and Luckin Coffee have migrated to Feishu from other platforms in the past six months, and more than 90% have also purchased Feishu AI products.

According to Kingsoft Office, WPS AI will have more than 80.1 million monthly active users in China by the end of 2025, an increase of 307% year-on-year, and will generate more than 200 billion token calls every day, an increase of more than 12 times.

The next wave of token demand is expected to come from traditional industries. Guo Zhenxing, vice president of China government and enterprise business at Huawei, said 2026 will be a breakthrough year for the industry’s integration with AI.

According to him, by 2025, AI will have proven itself in areas such as home appliance research and services, autonomous driving, financial customer consulting, and unmanned mining.

However, industry adoption remains uneven. SenseTime’s Jia said that the adoption of AI will depend on the level of informatization and digitalization of the sector.

Internet, software, and AI-native companies can transform the fastest, but industries with weaker digital foundations or those that rely more on expert judgment face higher barriers.

This is due to changes in the demand for human resources. According to a June 17 report by Tongdao Liepin Group and Tsinghua University AI and Management Research Center, the demand for AI algorithm and model development talent is basically stable, but its share of total AI demand has declined from about 50 percent in 2022 to 20 percent in the first quarter of 2026.

Demand for AI agent building and generation AI applications grew 40% and 35%, respectively, from the previous quarter.

DeepSeek’s June 25 recruitment notice reflected the same change. In addition to research and algorithm work, roles such as AI product manager, AI product operations, general agent data product manager, and agent harness team positions are also listed.

Sun Jiaxin, senior AI headhunter at Robert Walters, said companies are increasingly seeking talent who can implement AI into business processes, organizational systems and production environments.

By mid-2026, frontier deployment engineer, agent engineering roles, AI application implementation, and AI transformation delivery jobs were the primary sources of AI hiring demand for his team. caixin global