The AI token takeover in China

[BEIJING] The token, the foundational unit that large language models use to process text, is becoming a key gauge of competitiveness in the artificial intelligence (AI) industry.

As coding tools and AI agents have grown more prevalent, demand for tokens is rising far faster than many company executives expected, setting off a contest across China’s AI industry over pricing, computing power, corporate clients and new layers of AI infrastructure.

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

As reasoning models became more common, a single exchange grew to about 10,000 tokens. This year, he said, the average has risen to more than 50,000 tokens.

For Chinese model developers, the surge is already translating into rapid revenue growth.

Zhipu AI said its application programming interface revenue in the first quarter rose sixty times from a year earlier.

Alibaba Cloud said its token revenue has grown more than 15 times since the start of 2026. ByteDance’s Doubao model surpassed 180 trillion token calls a day in June.

The boom is straining the AI supply chain. Since August 2025, a wave of data centre construction has tightened supply of memory chips, graphics processors, central processors, optical modules and liquid cooling systems.

Prices for some memory chips have more than doubled, with capacity for the coming year largely sold out.

SEE ALSO

Cheaper AI is better: Soaring bills are reshaping how businesses choose models

As ‘tokenmaxxing’ and AI costs take off, companies turn to private token factories: Dell

AI’s token economy revolution creates new China tech winners

The AI trade is losing one of its key signals

More

For now, profits in the token economy remain concentrated upstream. In 2025, Nvidia and the world’s three major memory chip makers generated more than US$160 billion in combined operating profit.

Model companies, by contrast, face heavy training and inference costs and remain in a high-investment, low-profit stage — with some still losing money.

“If tokens are destined to become the next generation of electricity, then data centres are power plants, and large-model and cloud companies are the power grid,” a China executive at an international cloud company told Caixin.

“Companies don’t want tokens that are simply easy to use or cheap. They want cost-performance.”

Pricing whiplash

For cloud and model companies, the token boom has created a pricing dilemma. Raising prices can encourage rationing of scarce computing power and protect margins, but cutting prices can stimulate usage, lock in customers and build market share.

Chinese cloud providers such as Tencent Cloud raised prices for core models by more than 430 per cent in March.

Alibaba Cloud said in March it would raise prices for some AI computing services by 5 per cent to 34 per cent and increase the price of its high-performance storage system by 30 per cent.

Model makers followed. Zhipu AI chief executive officer Zhang Peng said on an earnings call that Zhipu’s API prices rose 83 per cent quarter on quarter in the first quarter, while call volume still jumped more than four times.

On Jun 23, ByteDance released Doubao 2.1, doubling the price from the previous version. Its consumer-facing Doubao app also introduced a Pro subscription plan starting at 68 yuan (US$10.03) a month.

Cloud computing companies are adjusting prices because of supply shortages, Cui Tingting, a research manager at IDC China, told Caixin.

Rapid downstream demand has pushed cloud vendors to increase AI investment, while key hardware is in short supply and delivery times are lengthening, she said.

Alibaba Group CEO Eddie Wu told investors that as AI applications shift from chatbots to agents, customers are becoming more willing to accept token pricing.

At the same time, he said, server costs have more than doubled from January 2025. Over the next one to two years, prices for models and other cloud services will keep rising, he said, helping lift Alibaba Cloud’s gross margin.

Yet many industry executives told Caixin token prices must eventually fall for AI to be widely adopted. Token prices will inevitably drop as chip costs decline and technology advances, Chinese AI inference startup Beijing Silicon Flow Technology co-founder Hu Jian told Caixin.

The turning point, he said, will come when domestic chips achieve breakthroughs in performance and supply capacity.

DeepSeek, which led an earlier token price war in China, cut prices again in May.

Its DeepSeek V4 Pro was priced at one quarter of its original level. In late June, the company said it would launch the official version of DeepSeek V4 in mid-July and introduce peak-and-valley pricing. Daytime peak prices will be twice normal levels, though the model will still retain a clear price advantage.

One inference service provider told Caixin that DeepSeek can cut prices because it keeps computing power utilisation high and lowers inference costs through scale.

“Even if high-load operations cause some volatility in service stability, users are still highly tolerant of DeepSeek,” the person said.

Middle-layer moment

Anxiety over token supply and cost is giving new value to the companies sitting between applications and foundation models.

These “middle-layer” providers include model aggregation platforms, inference service companies, memory management systems, security and risk control providers, and AI infrastructure firms that optimise deployment across different chips.

The gap between enterprise demand and model company capacity creates room for middle-layer providers, said Jia Anya, head of SenseTime Group’s Raccoon agent application family.

Hardware supply will still take years to balance, she said, while model companies are focused mainly on improving performance and may not have the resources to focus on lowering inference costs.

“Enterprise users are very sensitive to 20 per cent to 30 per cent price increases,” Jia said. “They need AI infrastructure with controllable budgets, stability and high cost-performance.”

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

Revenue in 2025 reached 55.3 million yuan, up more than 650 per cent from a year earlier.

SiliconFlow’s main product, SiliconCloud, provides fine tuning, hosting and deployment services for more than 170 large models.

On Jun 16, SiliconFlow said it had completed a Series B financing round of more than 2 billion yuan, the largest so far for a Chinese AI middle-layer service provider, valuing the company at US$7.7 billion.

SiliconFlow ranked fourth in China’s Model-as-a-Service market in 2025 by token call volume, IDC data show. ByteDance’s Volcano Engine ranked first with a share of more than 40 per cent, followed by Alibaba Cloud with about 30 per cent.

Other AI infrastructure providers are attacking different bottlenecks.

Qingcheng.AI has built AI Ping, a model calling platform that evaluates models from providers. It also offers price comparison and smart routing to help developers choose and call models more efficiently.

“Hardware determines the height of the ceiling, while software tries to reach as close as possible to that ceiling,” said Shi Tianhui, Qingcheng.AI’s co-founder.

On the same computing power and model, different infrastructure software can change hardware utilisation, model intelligence, token production efficiency and provider margins, he said.

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

Founder and CEO Xiong Feiyu said 50 per cent to 60 per cent of the company’s demand comes from industrial and financial agents, with the rest from emotional companionship, gaming and AI hardware.

AI goes to work

The largest immediate source of token consumption from agents is coding. Inside companies, however, the bigger prize is office workflow.

Analysts at China International Capital said enterprise office applications could break the old pattern of low-frequency, lightweight and experimental interactions.

Once AI agents connect to enterprises’ core data systems, token use will no longer be limited by human interaction frequency, but will scale with business flow.

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

In the co-pilot stage, AI saves time, but humans still modify formats, send emails and complete other steps. In the AI-native stage, agents will deliver entire projects, he said.

MuleRun launched overseas and, within two months, had paying users in 43 countries, Chen said on May 20. The 34-year-old later replaced DingTalk founder Chen Hang as CEO of DingTalk, becoming the youngest chief of any Alibaba business unit.

On Jul 2, Alibaba consolidated three office-agent products: QoderWork, a programmable agent desktop app; Wukong, an enterprise AI work platform incubated by DingTalk; and MuleRun.

The reshuffle came as DingTalk, with more than 800 million users and the largest share of China’s office software market, struggled to keep its lead in the AI-agent cycle.

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

Tencent said in its first-quarter earnings report that CodeBuddy and WorkBuddy had active user retention above 60 per cent and paying-user retention above 80 per cent.

WorkBuddy ranked first among China’s PC-based AI-native office agent platforms in March, with 8.85 million monthly visits and 831 per cent month-on-month user growth, according to research firm Analysys.

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

ByteDance’s Feishu connected its agent Aily to workplace scenarios in June, allowing users to make charts, build webpages, reply to document comments, answer emails and handle formatting through natural-language prompts.

Feishu said companies including Seres Group, Unitree Robotics and Luckin Coffee have migrated from other platforms to Feishu in the past six months, with more than 90 per cent also buying Feishu AI products.

Kingsoft Office said WPS AI had more than 80.1 million monthly active users in China by the end of 2025, up 307 per cent from a year earlier, and generated more than 200 billion daily token calls, up more than twelve times.

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

In 2025, AI proved itself in home appliance research and service, autonomous driving, financial customer consultation and unmanned mining, he said.

But industry adoption remains uneven. Jia of SenseTime said AI deployment depends on a sector’s level of informatisation and digitisation.

Internet, software and AI-native companies can transform fastest, while industries with weak digital foundations or heavy reliance on expert judgment face higher barriers.

That is changing talent demand. Demand for AI algorithm and model development talent has remained basically stable, but its share of total AI demand fell from about 50 per cent in 2022 to 20 per cent in the first quarter of 2026, according to a Jun 17 report by Tongdao Liepin Group and Tsinghua University’s Center for AI and Management Research.

Demand for AI-agent construction and generative-AI applications rose 40 per cent and 35 per cent, respectively, from the previous quarter.

DeepSeek’s Jun 25 recruitment notice reflected the same shift. Alongside research and algorithm jobs, it listed roles including AI product manager, AI product operations, general agent data product manager and agent harness team positions.

Sun Jiaxin, a senior AI headhunter at Robert Walters, said companies increasingly want people who can bring AI into business processes, organisational systems and production environments.

By mid-2026, frontier deployment engineers, agent engineering roles, AI application implementation and AI transformation delivery jobs had become a major source of AI hiring demand for his team. CAIXIN GLOBAL