China builds computing network as AI token usage skyrockets

In May 2024, ByteDance's AI assistant Doubao was consuming around 100 billion tokens per day. By March 2026, that figure had surpassed 120 trillion – a roughly 1,000-fold surge in just two years, according to industry data.

Doubao is far from alone. Across China, daily average token consumption reached 140 trillion in March 2026, up from just 100 billion in early 2024, a more than 1,000-fold increase, according to the National Data Bureau. 

The explosive growth reflects the deepening adoption of AI agents and large-scale model applications across industries.

Globally, the trend is equally striking. From training complex models to running AI agents, tokens – the basic units of data processed by large language models – have become a fundamental metric of the AI era. And the world is burning through them at an unprecedented rate.

OpenAI's flagship GPT-5.5 charges $5 per million input tokens and $30 per million output tokens via API. Anthropic's Claude series follows a tiered API pricing structure: Opus 4.8 at $5/$25 per million tokens (input/output), Sonnet 4.6 at $3/$15, and Haiku 4.5 at $1/$5.

Meanwhile, in coding agent scenarios, token consumption can run 10 to 50 times higher than in standard chat due to context accumulation and search overhead.

In China, pricing remains highly competitive. Alibaba Cloud's Qwen3.7-Max charges $2.5/$7.5 per million tokens. DeepSeek's V3.1 comes in at 2 yuan ($0.28) per million input tokens and 8 yuan ($1.10) for output – significantly cheaper than most global counterparts. 

ByteDance's Doubao Lite is priced even lower, at just 0.3 yuan for input and 0.6 yuan for output per million tokens, while offering generous free quotas. Other domestic providers, including Baidu's ERNIE, Zhipu's GLM, and Moonshot, have adopted similarly accessible pricing.

But language matters. Chinese text processing consumes 11% to 64% more tokens than English on Claude and GPT models. 

On domestic models like Qwen and DeepSeek, however, Chinese is actually more token-efficient – a critical advantage for local developers.

Behind the price tags lies a vast and costly infrastructure. As agent-based applications explode, cloud providers are shifting from resource-based billing to token-based metering, directly linking costs to usage volumes. 

According to industry estimates, monthly token expenses for enterprise users can run into the hundreds of thousands – or even millions – of yuan.

To address the long-term sustainability of AI computing, China is building a national integrated computing power network – a unified infrastructure that connects data centers, supercomputing centers, and edge computing facilities across the country, enabling on-demand resource allocation much like the national power grid.

The network has been designated a key infrastructure project for the 15th Five-Year Plan period, alongside water, power and transport networks. 

The National Development and Reform Commission has projected that total investment in the network and related fields will exceed 7 trillion yuan ($968 billion) this year.

In telecommunications infrastructure, the foundation for this computing grid is already taking shape. As of April 2026, China had deployed over 5 million 5G base stations, covering 330 cities with 5G-A networks. 

During the 15th Five-Year Plan period, the network will evolve from "dual-gigabit" to "dual-ten-gigabit," with 500,000 new 5G-A base stations planned. 

6G is expected to be commercially launched around 2030, increasing network efficiency and connectivity capacity by more than tenfold. Thousands of low-orbit satellites will also be launched, enabling integrated air-space-ground-sea coverage across the entire territory.

The next-generation communication network is expected to drive 7 trillion yuan in total industrial output across upstream and downstream sectors, empowering the digital economy, low-altitude economy, and embodied intelligence.